Despite the flashy title that's the first "sober" analysis from a CEO I read about the technology. While not even really news, it's also worth mentioning that the energy requirements are impossible to fulfill
Also now using ChatGPT intensely since months for all kinds of tasks and having tried Claude etc. None of this is on par with a human. The code snippets are straight out of Stackoverflow...
Take this "sober" analysis with a big pinch of salt.
IBM have totally missed the AI boat, and a large chunk of their revenue comes from selling expensive consultants to clients who do not have the expertise to do IT work themselves - this business model is at a high risk of being disrupted by those clients just using AI agents instead of paying $2-5000/day for a team of 20 barely-qualified new-grads in some far-off country.
IBM have an incentive to try and pour water on the AI fire to try and sustain their business.
Asking because the biggest IT consulting branch of IBM, Global Technology Services (GTS), was spun off into Kyndryl back in 2021[0]. Same goes for some premier software products (including one I consulted for) back in 2019[1]. Anecdotal evidence suggests the consulting part of IBM was already significantly smaller than in the past.
It's worth noting that IBM may view these AI companies as competitors to it's Watson AI tech[2]. It already existed before the GPU crunch and hyperscaler boom - runs on proprietary IBM hardware.
Missed the boat? Have you been living under a rock? Watson AI advertising has been everywhere for years.
It’s not that they aren't in the AI space, it’s that the CEO has a shockingly sober take on it. Probably because they’ve been doing AI for 30+ years combined with the fact they don’t have endless money with nowhere to invest it like Google.
> IBM have an incentive to try and pour water on the AI fire to try and sustain their business.
IBM has faced multiple lawsuits over the years. From age discrimination cases to various tactics allegedly used to push employees out, such as requiring them to relocate to states with more employer friendly laws only to terminate them afterward.
IBM is one of the clearest examples of a company that, if given the opportunity to replace human workers with AI, would not hesitate to do so. Assume therefore, the AI does not work for such a purpose...
Are you suggesting IBM made up the numbers? Or that CAPEX is a pre-GAI measure and is useless in guiding decision making?
IBM may have a vested interest in calming (or even extinguishing) the AI fire, but they're not the first to point out the numbers look a little wobbly.
And why should I believe OpenAI or Alphabet/Gemini when they say AI will be the royal road to future value? Don't they have a vested interest in making AI investments look attractive?
> a high risk of being disrupted by those clients just using AI agents instead of paying $2-5000/day for a team of 20 barely-qualified new-grads in some far-off country
Is there any concrete evidence of that risk being high? That doesn't come from people whose job is to sell AI?
they have incentive but what's the sustainable, actually-pays-for-itself-and-generates-profit cost of AI? We have no idea. Everything is so heavily subsidized by burning investor capital for heat with the hope that they'll pull an amazon and make it impossible to do business on the internet without paying an AI firm. Maybe the 20 juniors will turn out to be cheaper. Maybe they'll turn out to be slightly better. Maybe they'll be loosely equivalent and the ability to automate mediocrity will drive down the cost of human mediocrity. We don't know and everyone seems to be betting heavily on the most optimistic case, so it makes an awful lot of sense to take the other side of that bet.
Do you expect Sam Altman to come on stage and tell you the whole thing is a giant house of cards when the entire western economy seems to be propped up by AI? I wonder whose "sober" analysis you would accept, because surely the people that are making money hand over fist will never admit it.
Seems to me like any criticism of AI is always handwaved away with the same arguments. Either it's companies who missed the AI wave, or the models are improving incredibly quickly so if it's shit today you just have to wait one more year, or if you're not seeing 100x improvements in productivity you must be using it wrong.
IBM was ahead of the boat! They had Watson on Jeopardy years ago! /s
I think you make a fair point about the potential disruption for their consulting business but didn't they try to de-risk a bit with the Kyndryl spinout?
I am a senior engineer, I use cursor a lot in my day to day. I find I can code longer and typically faster than without. Is it on par with human? It’s getting pretty darn close to be honest, I am sure the “10x” engineers of the world would disagree but it definitely has surpassed a junior engineer. We all have our anecdotes but I am inclined to believe on average there is net value.
I think surpassed is not the right word because it doesn't create/ideate. However it is incredibly resourceful. Maybe like having a jr engineer to do your bidding without thinking or growing.
Largely agree. Anything that is just a multi-file edit, like an interface change, it can do. Maybe not immediately, but you can have it iterate, and it doesn't eat up your attention.
It is without a doubt worth more than the 200 bucks a month I spend on it.
I will go as far as to say it has decent ideas. Vanilla ideas, but it has them. I've actually gotten it to come up with algorithms that I thought were industry secrets. Minor secrets, sure. But things that you don't just come across. I'm in the trading business, so you don't really expect a lot of public information to be in the dataset.
i'm also a senior engineer and I use codex a lot. It has reduced many of the typical coding tasks to simply writing really good AC. I still have to write good AC, but I'm starting to see the velocity change from using good AI in a smart way.
Senior engineer here as well. I would say Opus 4.5 is easily a mid-level engineer. It's a substantial improvement over Sonnet 4.5, which required a lot more hand-holding and interventions.
I've found Claude's usefulness is highly variable, though somewhat predictable. It can write `jq` filters flawlessly every time, whereas I would normally spend 30 minutes scanning docs because nobody memorizes `jq` syntax. And it can comb through server logs in every pod of my k8s clusters extremely fast. But it often struggles making quality code changes in a large codebase, or writing good documentation that isn't just an English translation of the code it's documenting.
I'm just as much of an avid llm code generator fan as you may be but I do wonder about the practicality of spending time making projects anymore.
Why build them if other can just generate them too, where is the value of making so many projects?
If the value is in who can sell it the best to people who can't generate it, isn't it just a matter of time before someone else will generate one and they may become better than you at selling it?
An issue with the doom forecasts is most of the hypothetical $8tn hasn't happened yet. Current big tech capex is about $315bn this year, $250bn last against a pre AI level ~$100bn so ~$400bn has been spent so far on AI boom data centers. https://sherwood.news/business/amazon-plans-100-billion-spen...
The future spend is optional - AGI takeoff, you spend loads, not happening not so much.
Say it levels of at $800bn. The world's population is ~8bn so $100 a head so you'd need to be making $10 or $20 per head per year. Quite possibly doable.
I agree. re: energy and other resource use: the analogy I like is with driving cars: we use cars for transportation knowing the environmental costs so we don’t usually just go on two hour drives for the fun of it, rather we drive to get to work, go shopping. I use Gemini 3 but only in specific high value use cases. When I use commercial models I think a little about the societal costs.
In the USA we have lost the thread here: we don’t maximize the use of small tuned models throughout society and industry, instead we use the pursuit of advanced AI as a distraction to the reality that our economy and competitiveness are failing.
You could have your morning shower 1°C less hot and save enough energy for about 200 prompts (assuming 50 litres per shower). (Or skip the shower altogether and save thousands of prompts.)
Yesterday I was talking to coworkers about AI I mentioned that a friend of mine used ChatGPT to help him move. So a coworker said I have to test this and asked ChatGPT if he could fit a set of the largest Magnepan speakers (the wide folding older room divider style) in his Infinity QX80. The results were hilarious. It had some of the dimensions right but it then decided the QX80 is as wide as a box truck (~8-8.5 feet/2.5 m) and to align the nearly 7 foot long speakers sideways between the wheel wells. It also posted hilariously incomprehensible ASCII diagrams.
I'm not sure what you mean with the "code snippets are straight out of Stackoverflow".
That is factually incorrect just by how LLM works. By now there has been so much code ingested from all kinds of sources, including Stackoverflow LLM is able to help generate quite good code in many occasions.
My point being it is extremly useful for super popular languages and many languages where resources are more scarce for developer but because they got the code from who knows where, it can definitely give you many useful ideas.
It's not human, which I'm not sure what is supposed to actually mean. Humans make mistakes, humans make good code. AI does also both. What it definitely needs is a good programmer still on top to know what he is getting and how to improve it.
I find AI (LLM) very useful as a very good code completion and light coder where you know exactly what to do because you did it a thousand times but it's wasteful to be typing it again. Especially a lot of boilerplate code or tests.
It's also useful for agentic use cases because some things you just couldn't do before because there was nothing to understand a human voice/text input and translate that to an actual command.
But that is all far from some AGI and it all costs a lot today an average company to say that this actually provided return on the money but it definitely speeds things up.
> I'm not sure what you mean with the "code snippets are straight out of Stackoverflow". That is factually incorrect just by how LLM works.
I'm not an AI lover, but I did try Gemini for a small, well-contained algorithm for a personal project that I didn't want to spend the time looking up, and it was straight-up a StackOverflow solution. I found out because I said "hm, there has to be a more elegant solution", and quickly found the StackOverflow solution that the AI regurgitated. Another 10 or 20 minutes of hunting uncovered another StackOverflow solution with the requisite elegance.
> While not even really news, it's also worth mentioning that the energy requirements are impossible to fulfill
If you believe this, you must also believe that global warming is unstoppable. OpenAI's energy costs are large compared to the current electricity market, but not so large compared to the current energy market. Environmentalists usually suggest that electrification - converting non-electrical energy to electrical energy - and then making that electrical energy clean - is the solution to global warming. OpenAI's energy needs are something like 10% of the current worldwide electricity market but less than 1% of the current worldwide energy market.
Google recently announced to double AI data center capacity every 6 month. While both unfortunately deal with exponential growth, we are talking about 1% growth CO2 which is bad enough vs 300% effectively per year according to Google
Imagine how big pile of trash as the current generation of graphics cards used for LLM training will get outdated. It will crash the hardware market (which is a good news for gamers)
I'd rather phrase it as "code is straight out of GitHub, but tailored to match your data structures"
That's at least how I use it. If I know there's a library that can solve the issue, I know an LLM can implement the same thing for me. Often much faster than integrating the library. And hey, now it's my code. Ethical? Probably not. Useful? Sometimes.
If I know there isn't a library available, and I'm not doing the most trivial UI or data processing, well, then it can be very tough to get anything usable out of an LLM.
> it's also worth mentioning that the energy requirements are impossible to fulfill
Maybe I'm misunderstanding you but they're definitely not impossible to fulfill, in fact I'd argue the energy requirements are some of the most straightforward to fulfill. Bringing a natural gas power plant online is not the hardest part in creating AGI
> In an October letter to the White House's Office of Science and Technology Policy, OpenAI CEO Sam Altman recommended that the US add 100 gigawatts in energy capacity every year.
> Krishna also referenced the depreciation of the AI chips inside data centers as another factor: "You've got to use it all in five years because at that point, you've got to throw it away and refill it," he said.
And people think the climate concerns of AI are overblown. Currently US has ~1300 GW of energy capacity. That's a huge increase each year.
The largest plant in the world is the Three Gorges Dam in China at 22GW and it’s off the scales huge. We’re not building the equivalent of four of those every year.
Unless the plan is to power it off Sam Altman’s hot air. That could work. :)
Background: I live within the US Federal Tennessee Valley Authority (TVA), a regional electric grid operator. The majority of energy is generated by nuclear + renewables, with coal and natural gas as peakers. Grid stability is maintained by among the largest batteries in the world, Racoon Mountain Pumped Storage Facility.
Three Gorges Dam is capable of generating more power than all of TVA's nuclear + hydro, combined. In the past decade, TVA's single pumped-storage battery has gone from largest GWh/capcity in the world to not even top ten — largest facilities are now in China.
µFission reactors have recently been approved for TVA commissioning, with locations unconfirmed (but about one-sixth the output of typical TVA nuclear site). Sub-station battery storage sites are beginning to go online, capable of running subdivisions for hours after circuit disconnects.
Tech-funded entities like Helios Energy are promising profitable ¡FusioN! within a few years ("for fifty years").
----
All of the above just to say: +100GW over the next decade isn't that crazy a prediction (+20% current supply, similar in size to two additional Texas-es).
LOL, maybe Sam Altman can fund those power plants. Let me guess: He'd rather the public pay for it, and for him to benefit/profit from the increased capacity.
Big tech is going to have to fund the plants and probably transmission. Because the energy utilities have a decades long planning horizon for investments.
Good discussion about this in recent Odd Lots podcast.
Scam Altman wants the US to build a lot of energy plants so that the country will pay the costs and OpenAI will have the profits of using this cheap energy.
> Krishna also referenced the depreciation of the AI chips inside data centers as another factor: "You've got to use it all in five years because at that point, you've got to throw it away and refill it," he said
This doesn't seem correct to me, or at least is built on several shaky assumptions. One would have to 'refill' your hardware if:
- AI accelerator cards all start dying around the 5 year mark, which is possible given the heat density/cooling needs, but doesn't seem all that likely.
- Technology advances such that only the absolute newest cards can be used to run _any_ model profitably, which only seems likely if we see some pretty radical advances in efficiency. Otherwise, it seems like assuming your hardware is stable after 5 years of burn in, you could continue to run older models on that hardware at only the cost of the floorspace/power. Maybe you need new cards for new models for some reason (maybe a new fp format that only new cards support? some magic amount of ram? etc), but it seems like there may be room for revenue via older/less capable models at a discounted rate.
Isn’t that what Michael Burry is complaining about? That five years is actually too generous when it comes to depreciation of these assets and that companies are being too relaxed with that estimate. The real depreciation is more like 2-3 years for these GPUs that cost tens of thousands of dollars a piece.
That's exactly the thing. It's only about bookkeeping.
The big AI corps keep pushing depreciation for GPUs into the future, no matter how long the hardware is actually useful. Some of them are now at 6 years. But GPUs are advancing fast, and new hardware brings more flops per watt, so there's a strong incentive to switch to the latest chips. Also, they run 24/7 at 100% capacity, so after only 1.5 years, a fair share of the chips is already toast. How much hardware do they have in their books that's actually not useful anymore? Noone knows! Slower depreciation means more profit right now (for those companies that actually make profit, like MS or Meta), but it's just kicking the can down the road. Eventually, all these investments have to get out of the books, and that's where it will eat their profits. In 2024, the big AI corps invested about $1 trillion in AI hardware, next year is expected to be $2 trillion. Only the interest payments for that are crazy. And all of this comes on top of the fact that none of the these companies actually make any profit at all with AI. (Except Nvidia of course) There's just no way this will pan out.
How different is this from rental car companies changing over their fleets? I don't know, this is a genuine question. The cars cost 3-4x as much and last about 2x as far as I know, and the secondary market is still alive.
I think its illustrative to consider the previous computation cycle ala Cryptomining. Which passed through a similar lifecycle with energy and GPU accelerators.
The need for cheap wattage forced the operations to arbitrage the where location for the cheapest/reliable existing supply - there rarely was new buildout as the cost was to be reimbursed by the coins the miningpool recovered.
For the chip situation caused the same apprecaition in GPU cards with periodic offloading of cards to the secondary market (after wear and tear) as newer/faster/more efficient cards came out until custom ASICs took over the heavy lifting, causing the GPU card market to pivot.
Similarly in the short to moedium term the uptick of custo ASICs like with Google TPU will definately make a dent in bot cpex/opex and potentially also lead to a market with used GPUs as ASICs dominate.
So for GPUs i can certainly see the 5 year horizon making a impact in investment decisions as ASICs proliferate.
It’s far more extreme: old servers are still okay on I/O, and memory latency, etc. won’t change that dramatically so you can still find productive uses for them. AI workloads are hyper-focused on a single type of work and, unlike most regular servers, a limiting factor in direct competition with other companies.
LambdaLabs is still making money off their Tesla V100s, A100s, and A6000s. The older ones are cheap enough to run some models and very cheap, so if that's all you need, that's what you'll pick.
The V100 was released in 2017, A6000 in 2020, A100 in 2021.
Manipulating this for creative accounting seems to be the root of Michael Burry’s argument, although I’m not fluent enough in his figures to map here. But, commenting that it interesting to see IBM argue a similar case (somewhat), or comments ITT hitting the same known facts, in light of Nvidia’s counterpoints to him.
> They still work fine but power costs make them uneconomical compared to latest tech.
That's not necessarily the driving financial decision, in fact I'd argue company's with data center hardware purchases barely look at this number. It's more simple than that - their support runs out and its cheaper to buy a new piece of hardware (that IS more efficient) because the hardware vendors make extended support inordinately expensive.
Put yourselves in the shoes of a sales person at Dell selling enterprise server hardware and you'll see why this model makes sense.
Eh, not exactly. If you don't run CPU at 70%+ the rest of the machine isn't that much more inefficient that model generation or two behind.
It used to be that new server could use half power of the old one at idle but vendors figured out that servers also need proper power management a while ago and it is much better.
Last few gens increase could be summed up to "low % increase in efficiency, with TDP, memory channels and core count increase".
So for loads not CPU bound the savings on newer gen aren't nearly worth it to replace it, and for bulk storage the CPU power usage is even smaller part
But if your competitor is running newer chips that consume less power per operation, aren't you forced to upgrade as well and dispose of the old hardware?
Sure, assuming the power cost reduction or capability increase justifies the expenditure. It's not clear that that will be the case. That's one of the shaky assumptions I'm referring to. It may be that the 2030 nvidia accelerators will save you $2000 in electricity per month per rack, and you can upgrade the whole rack for the low, low price of $800,000! That may not be worth it at all. If it saves you $200k/per rack or unlocks some additional capability that a 2025 accelerator is incapable of and customers are willing to pay for, then that's a different story. There are a ton of assumptions in these scenarios, and his logic doesn't seem to justify the confidence level.
It’s not about assumptions on the hardware. It’s about the current demands for computation and expected growth of business needs. Since we have a couple years to measure against it should be extremely straightforward to predict. As such I have no reason to doubt the stated projections.
Networking gear was famously overbought. Enterprise hardware is tricky as there isn’t much of a resale market for this gear once all is said and done.
The only valid use case for all of this compute which could reasonably replace ai is btc mining. I’m uncertain if the increased mining capacity would harm the market or not.
Failure rates also go up. For AI inference it’s probably not too bad in most cases, just take the node offline and re-schedule the jobs to other nodes.
There is the opportunity cost of using a whole datacenter to house ancient chips, even if they're still running. You're thinking like a personal use chip which you can run as long as it is non-defective. But for datacenters it doesn't make sense to use the same chips for more than a few years and I think 5 years is already stretching their real shelf life.
Do not forget that we're talking about supercomputers. Their interconnect makes machines not easily fungible, so even a low reduction in availability can have dramatic effects.
Also, after the end of the product life, replacement parts may no longer be available.
You need to get pretty creative with repair & refurbishment processes to counter these risks.
Historically, GPUs have improved in efficiency fast enough that people retired their hardware in way less than 5 years.
Also, historically the top of the line fabs were focused on CPUs, not GPUs. That has not been true for a generation, so it's not really clear if the depreciation speed will be maintained.
> Historically, GPUs have improved in efficiency fast enough that people retired their hardware in way less than 5 years.
This was a time when chip transistor cost was decreasing rapidly. A few years earlier even RAM cost was decreasing quickly. But these times are over now. For example, the PlayStation 5 (where the GPU is the main cost), which launched five years ago, even increased in price! This is historically unprecedented.
Most price/performance progress is nowadays made via better GPU architecture instead, but these architectures are already pretty mature, so the room for improvement is limited.
Given that the price per transistor (which TSMC & Co are charging) has decreased ever more slowly in recent years, I assume it will eventually come almost to a halt.
By the way, this is strictly speaking compatible with Moore's law, as it is only about transistors per chip area, not price. Of course the price per chip area was historically approximately constant, which meant exponentially increasing transistor density implied exponentially decreasing transistor price.
5 years is maybe referring to the accounting schedule for depreciation on computer hardware, not the actual useful lifetime of the hardware.
It's a little weird to phrase it like that though because you're right it doesn't mean you have to throw it out. Idk if this is some reflection of how IBM handles finance stuff or what. Certainly not all companies throw out hardware the minute they can't claim depreciation on it. But I don't know the numbers.
Anyways, 5 years is an infection point on numbers. Before 5 years you get depreciation to offset some cost of running. After 5 years, you do not, so the math does change.
that is how the investments are costed though, so makes sense when we're talking return on investment, so you can compare with alternatives under the same evaluation criteria.
When you operate big data centers it makes sense to refresh your hardware every 5 years or so because that’s the point at which the refreshed hardware is enough better to be worth the effort and expense.
You don’t HAVE to, but its more cost effective if you do.
(Source, used to operate big data centers)
It's worse than that in reality, AI chips are on a two year cadence for backwards compatibility (NVIDIA can basically guarantee it, and you probably won't be able to pay real AI devs enough to stick around to make hardware work arounds). So their accounting is optimistic.
5 years is normal-ish depreciation time frame. I know they are gaming GPUs, but the RTX 3090 came out ~ 4.5 years before the RTX 5090. The 5090 has double the performance and 1/3 more memory. The 3090 is still a useful card even after 5 years.
General question to people who might actually know.
Is there anywhere that does anything like Backblaze's Hard Drive Failure Rates [1] for GPU Failure Rates in environments like data centers, high-performance computing, super-computers, mainframes?
The best that came back on a search was a semi-modern article from 2023 [2] that appears to be a one-off and mostly related to consumer facing GPU purchases, rather than bulk data center, constant usage conditions. It's just difficult to really believe some of these kinds of hardware deprecation numbers since there appears to be so little info other than guesstimates.
Found an Arxiv paper on continued checking that's from UIUC UrbanaIL about a 1,056 A100 and H100 GPU system. [3] However, the paper is primarily about memory issues and per/job downtime that causes task failures and work loss. GPU Resilience is discussed, it's just mostly from the perspective of short-term robustness in the face of propagating memory corruption issues and error correction, rather than multi-year, 100% usage GPU burnout rates.
Any info on the longer term burnout / failure rates for GPUs similar to Backblaze?
Edit: This article [4] claims it's 0.1-2% failure rate per year (0.8% (estimated)) with no real info about where the data came from (cites "industry reports and data center statistics"), and then claims they often last 3-5 years on average.
Given power and price constraints, it's not that you cannot run them in 5 years time it's that you don't want to run them in 5 years time and neither will anyone else that doesn't have free power.
I would add an addendum to this -- there is no way the announced spending on AI data centers will all come to fruition. I have no doubt that there will be a massive build-out of infrastructure, but it can't reach the levels that have been announced. The power requirements alone will stop that from happening.
The power requirement is only an issue in western countries, where utilities build at most a double digit buffer, and are used to overall energy use leveling due to efficiency improvements. Now look at China where they routinely maintain a 100% buffer. Demand can double and they can supply that without new generation capacity.
I think you're spot on. OpenAI alone has committed to spending $1.4T on various hardware/DCs. They have nowhere near that amount of money and when pushed Altman gets defensive.
I believe there is a difference between what people say publicly and what they are actually committed to doing on the ground. When all is said and done, I'll be more interested to know what was actually spent.
For example, XYZ AI company may say they are going to spend $1T for AI data centers over the next 5 years.
In actuality, I suspect it is likely that they have committed to something like $5-10B in shovel-ready projects with stretch goals for the rest. And the remaining spend would be heavily conditioned -- is power available? are chips available? is the public support present? financing? etc...
Not to mention, it's a much bigger moat if you can claim you're going to spend $1T. Who else will want to compete with you when you're spending $1T. After the dust has settled and you've managed to be one of the 2-3 dominant AI players, who is going to care that you "only" spent $100B instead of $1T. Look -- you were very capital efficient!
So, do I see it as possible that XYZ AI company could spend $1T, sure. Is it likely? No.
The incentive for CEOs is announcing the plan to do something, they have no idea if they will actually be able to do it, and it probably won't matter.
This happened in the dotcom too btw. Companies built out fibre networks, it wasn't possible to actually build all the physical infra that companies wanted to build so many announced plans that never happened and then, towards the end, companies began aggressively acquiring stakes in companies who were building stuff to get financial exposure (an example was BT, which turned itself briefly into a hedge fund with a telephone network attached...before it imploded).
CEOs do not operate on the timescale of waiting and building. Their timescale is this year's bonus/share options package. Nothing else matters: announce plans to do X or Y, doesn't matter, they know they will be gone long before it happens.
Hmm... "CEOs and teams" don't necessary do what's makes sense mathematically. Many, if not most of them, do whatever that sounds good to shareholders in their quarterly earnings call and ignore the reality or long term development.
If "CEOs and teams" are smart enough, they would not have overhired during 2021-2022 and then do layoffs. Who would be dumb enough to do that?
me very smart. me go college get fancy paper. me move money out of ai stocks. flat rocks not think so well, use too much fire. me trade for shiny yellow rock instead.
IBM might not have a data strategy or AI plan but he isn’t wrong on the inability to generate a profit.
A bit of napkin math:
NVIDIA claims 0.4J per token for their latest generation
1GW plant with 80% utilisation can therefore produce 6.29 10^16 tokens a year.
There are ~10^14 tokens on the internet. ~10^19 tokens have been spoken by humans… so far.
Don't know what the source is, but it feels missing a few orders of magnitude. Surely it only counts text? I can't imagine there are only so few data on the internet if you count images and videos.
Tokens are, roughly speaking, how you pay for AI. So you can approximate revenue by multiplying tokens per year by the revenue for a token.
(6.29 10^16 tokens a year) * ($10 per 10^6 tokens)
= $6.29 10^11
= $629,000,000,000 per year in revenue
Per the article
> "It's my view that there's no way you're going to get a return on that, because $8 trillion of capex means you need roughly $800 billion of profit just to pay for the interest," he said.
$629 billion is less than $800 billion. And we are talking raw revenue (not profit). So we are already in the red.
But it gets worse, that $10 per million tokens costs is for GPT-5.1, which is one of the most expensive models. And the costs don't account for input tokens, which are usually a tenth of the costs of output tokens. And using bulk API instead of the regular one halves costs again.
Realistic revenue projections for a data center are closer to sub $1 per million tokens, $70-150 billion per year. And this is revenue only.
To make profits at current prices, the chips need to increase in performance by some factor, and power costs need to fall by another factor. The combination of these factors need to be, at minimum, like 5x, but realistically need to be 50x.
The interesting macro view on what's happening is to compare a mature data center operation (specifically a commoditized one) with the utility business. The margins here, and in similar industries with big infra build-out costs (ex: rail) are quite small. Historically the businesses have not done well; I can't really imagine what happens when tech companies who've only ever known huge, juicy margins experience low single digit returns on billions of investment.
Worse, is that a lot of these people are acting like Moore's law isn't still in effect. People conflate clock speeds on beefy hardware with moore's law, and act like it's dead, when transistor density rises, and cost per transistor continue to fall at rates similar to what they always have. That means the people racing to build out infrastructure today might just be better off parking that money in a low interest account, and waiting 6 months. That was a valid strategy for animation studios in the late 90s (it was not only cheaper to wait, but also the finished renders happened sooner), and I'd be surprised if it's not a valid strategy today for LLMs. The amount of silicon that is going to be produced that is specialized for this type of processing is going to be mind boggling.
Cost per transistor is increasing. or flat, if you stay on a legacy node. They pretty much squeezed all the cost out of 28nm that can be had, and it’s the cheapest per transistor.
“based on the graph presented by Milind Shah from Google at the industry tradeshow IEDM, the cost of 100 million transistors normalized to 28nm is actually flat or even increasing.”
The cloud mega scalers have done very well for themselves. As with all products the question is differentiation. If models can differentiate and lock in users they can have decent margins. If models get commoditized the current cloud providers will eat the AI labs lunch.
If AI is a highlander market, then the survivor will be able to eventually aquire all those assets on the cheap from the failing competitors that flush their debt in bankruptcy.
Meanwhile, highlander hopefuls are spending other peoples money to compete. Some of them with dreams of not just building a tech empire, but to truly own the machine that will rule the world in every aspect.
Investors are keen on backing the winner. They just do not know yet who it will be.
Until China sees it valuable to fund open weights SOTA-ish models, even the winner might struggle. There is very little capture - protocols are mostly standard so models are mostly interchangeable and if you are trying to raise prices enough to break even on the whole operation, somebody else can probably profitably run inference cheaper.
"oh look, with open weights now everyone can run this quadrillion param 'superintelligent' model. but what's that? only we have the power capacity and cheap $/W to actually serve it? funny thing...!"
I interpreted it as "there can only be one" which I believe is a quote from the Highlander movie; it's a "winner takes all" and that winner gets the title of "highlander."
In this situation then everyone who _isn't_ the winner will go broke -> sell off all their stuff on the cheap because they're desperate -> the winner gets all their hardware for a great deal and becomes even more powerful.
"It is 1958. IBM passes up the chance to buy a young, fledgling company that has invented a new technology called xerography. Two years later, Xerox is born, and IBM has been kicking themselves ever since. It is ten years later, the late '60s. Digital Equipment DEC and others invent the minicomputer. IBM dismisses the minicomputer as too small to do serious computing and, therefore, unimportant to their business. DEC grows to become a multi-hundred-million dollar corporation before IBM finally enters the minicomputer market. It is now ten years later, the late '70s. In 1977, Apple, a young fledgling company on the West Coast, invents the Apple II, the first personal computer as we know it today. IBM dismisses the personal computer as too small to do serious computing and unimportant to their business." - Steve Jobs [1][2][3]
Now, "IBM CEO says there is 'no way' spending on AI data centers will pay off". IBM has not exactly had a stellar record at identifying the future.
> IBM has not exactly had a stellar record at identifying the future.
IBM invented/developed/introduced magnetic stripe cards, UPC Barcodes, the modern ATM, Hard drives, floppies, DRAM, SQL, the 360 Family of Mainframes, the PC, Apollo guidance computers, Deep Blue. IBM created a far share of the future we're living in.
I'm no fan of much of what IBM is doing at the moment but it could be argued that its consultancy/service orientation gives it a good view of how business is and is planning to use AI.
They also either fairly accurately predicted the death of HDDs by selling off their research division before the market collapsed, or they caused the end of the HDD era by selling off their research division. They did a lot of research.
The other way to look at it is that the entire consulting industry is teetering on catastrophe. And IBM, being largely a consulting company now, is not being spared.
> IBM invented/developed/introduced magnetic stripe cards, UPC Barcodes, the modern ATM, Hard drives, floppies, DRAM, SQL, the 360 Family of Mainframes, the PC, Apollo guidance computers, Deep Blue. IBM created a far share of the future we're living in.
Well put. “IBM was wrong about computers being a big deal” is a bizarre take. It’s like saying that Colonel Sanders was wrong about chicken because he, uh… invented the pressure fryer.
Nitpicking, IBM did non develop _the_ Apollo Guidance Computer (the one in the spacecraft with people), it was Raytheon. They did, however, developed the Launch Vehicle Digital Computer that controlled the Saturn rocket in Apollo missions. AGC had very innovative design, while LVDC was more conventional for that time.
I've heard some second hand stories about IBM's way of using "AI" and it is pretty much business oriented and not much of the glamour and galore promises the other companies make (of course you still have shiny new things in business terms). It's actually good entertainment hearing all the internal struggles of business vs fancy during the holidays.
> In 1977, Apple, a young fledgling company on the West Coast, invents the Apple II, the first personal computer as we know it today. IBM dismisses the personal computer as too small to do serious computing and unimportant to their business.
IBM released the 5100 in September 1975 [0] which was essentially a personal computer in feature set. The biggest problem with it was the price tag - the entry model cost US$8975, compared to US$1298 for the entry Apple II released in June 1977 (close to two years later). The IBM PC was released in August 1981 for US$1565 for the most basic system (which almost no one bought, so in practice they cost more). And the original IBM PC had model number 5150, officially positioning it as a successor to the 5100.
IBM’s big problem wasn’t that they were disinterested in the category - it was they initially insisted on using expensive IBM-proprietary parts (often shared technology with their mainframe/midrange/minicomputer systems and peripherals), which resulted in a price that made the machine unaffordable for everyone except large businesses, governments, universities (and even those customers often balked at the price tag). The secret of the IBM PC’s success is they told the design team to use commercial off-the-shelf chips from vendors such as Intel and Motorola instead of IBM’s own silicon.
This is the exact kind of thinking that got us into this mess in the first place, and I'm not blaming you for it, it seems to be something all of us do to an extent. We don't look to Meta, who only a few years ago thought that the Metaverse would be the "next big thing" as an example of failure to identify the future, we look to IBM who made that mistake almost 30 years ago. Underestimating a technology seems to stick much harder than overestimating one.
If you want to be seen as relevant in this industry, or as a kind of "thought leader", the easy trick seems to be to hype up everything. If you do that and you're wrong, people will quickly forget. If you don't and you're wrong, that will stain your reputation for decades.
Good point. That kind of thinking is an absurdity. Saying IBM dropped the ball 70 years ago without acknowledging that lessons were learned, leadership has changed hands a lot since then, and most importantly, the tech landscape back then was very different from today unless you grossly oversimplify everything amounts to nothing more than a fallacious opinion.
Not even much of an IBM fan, myself, but I respect their considerable contribution to the industry. Sure, they missed a shot back then, but I think this latest statement is reliably accurate based on the information we currently have.
The amount of hate I've received here for similar statements is astonishing. What is even more astonishing is that it takes 3-rd grade math skills to work out that the current AI(even ignoring the fact that there is nothing intelligent about the current AI) costs are astronomical and they do not deliver on the promises and everyone is operating at wild loses. At the moment we are at "if you owe 100k to your bank, you have a problem but if you owe 100M to your bank, your bank has a problem". It's the exact same bullshitter economy that people like musk have been exploiting for decades: promise a ton, never deliver, make a secondary promise for "next year", rinse and repeat -> infinite profit. Especially when you rope in fanatical followers.
> We don't look to Meta, who only a few years ago thought that the Metaverse would be the "next big thing" as an example of failure to identify the future, we look to IBM who made that mistake almost 30 years ago.
The grandparent points to a pattern of failures whereas you point to Meta’s big miss. What you miss about Meta, and I am no fan, is that Facebook purchased Whatsapp and Instagram.
In other words, two out of three ain’t bad; IBM is zero for three.
While that’s not the thrust of your argument, which is about jumping on the problem of jumping on every hype train, the post to which you reply is not on about hype cycle. Rather, that post calls out IBM for a failure to understand the future of technology and does so by pointing to a history of failures.
Got anything vis-a-vis the message as opposed to the messenger?
I'm not sure these examples are even the gotchas you're positing them as. Xerox is a dinosaur that was last relevant at the turn of the century, and IBM is a $300bn company. And if it wasn't obvious, the Apple II never made a dent in the corporate market, while IBM and later Windows PCs did.
In any case, these examples are almost half a century old and don't relate to capex ROI, which was the topic of dicussion.
If it's not obvious, Steve's quote is ENTIRELY about capex ROI, and I feel his quote is more relevant to what is happening today than anything Arvind Krishna is imagining. The quote is posted in my comment not to grandstand Apple in any sense, but to grandstand just how consistently wrong IBM has been about so many opportunities that they have failed to read correctly - reprography, mini computers and microcomputers being just three.
Yes it is about ROI: "IBM enters the personal computer market in November ’81 with the IBM PC. 1983 Apple and IBM emerged as the industry’s strongest competitors each selling approximately one billion dollars worth of personal computers in 1983, each will invest greater than fifty million dollars for R&D and another fifty million dollars for television advertising in 1984 totaling almost one quarter of a billion dollars combined, the shakeout is in full swing. The first major firm goes bankrupt with others teetering on the brink, total industry losses for 83 out shadow even the combined profits of Apple and IBM for personal computers."
The message is plain to see with very little advanced math.
The only news is that it is the CEO of IBM saying it out loud.
IMHO he has some of the most credible opinions at this scale that many people have seen.
It's "highly unlikely" that all this money will be paid back to everyone that invested at this point. The losers probably will outnumber the winners, and nobody knows whether it will end up becoming a winner-take-all situation yet. A number of wealthy players remain at the table, raising stakes with each passing round.
It's so much money that it's already too late to do anything about it, and the full amount hasn't even changed hands yet.
And the momentum from something so huge can mean that almost the entire amount will have to change hands a second time before a stable baseline can be determined relative to pre-existing assets.
This can take longer than anyone gives credit for just because of massiveness, in the mean time, established real near-term growth opportunities may languish or even fade as the skew in rationality/solvency balance awaits the rolling dice to come to rest.
> Got anything vis-a-vis the message as opposed to the messenger?
Sure: People disagree. It's not like there is anything particularly clever that IBM CEO provided here. The guy not investing in something saying it won't work is about as good as the people who do saying it will. It's simply different assumptions about the future.
He is pointing out that the current costs to create the data centres means you will never be able to make a profit to cover those costs. $800 Billion just to cover the interest.
OpenAI is already haemorrhaging money and the space data centres has already been debunked. There is even a recent paper that points out that LLMs will never become AGI.
The article also finishes out with some other experts giving the same results.
Sry to say but the fact that you argue with LLMs never become AGI, you are not up-to-date.
People don't assume LLM will be AGI, people assume that World Models will lead us to AGI.
I personally never asumed LLM will become AGI, i always assumed that LLM broke the dam for investment and research into massivce scale compute ML learning and LLMs are very very good in showing were the future goes because they are already so crazy good that people can now imagine a future were AGI exists.
And that was very clear already when / as soon as GPT-3 came out.
The next big thing will probably be either a LOT more RL or self propelling ai architecture discovery. Both need massive compute to work well but then will potentially provide even faster progress as soon as humans are out of the loop.
IBM is an interesting beast when it comes to business decisions. While I can't give exact details, their business intelligence and ability to predict monetary things is uncannily spot-on at times.
So, when their CEO says that this investment will not pay off, I tend to believe them, because they most probably have the knowledge, insight and data to back that claim, and they have ran the numbers.
Oh, also, please let's not forget that they dabbled in "big AI" before everyone else. Anyone remembers Deep Blue and Watson, the original chatbot backed by big data?
We can cherry-pick blunders made by any big company to make a point. Maybe it would be more honest to also list companies IBM passed on that turned out to be rubbish? And all the technologies that IBM did invest in that made them a ton of money and became industry standards?[0]
Today, Xerox has less total revenue than IBM has profit. DEC went out of business 27 years ago. Apple is an in astoundingly great place right now, but Jobs got kicked out of his own company, and then returned when it was about to fail, having to take investment from Microsoft(!) in order to stay afloat.
Meanwhile, IBM is still here, making money hand over fist. We might not have a ton of respect for them, being mostly a consulting services company these days, but they're doing just fine.
No, but the comment above and variations of it are mentioned in every thread about IBM, so it’s probably just a reflex at this point without much thought behind it.
DEC went down the drain, Xerox is 1/1000 of IBM's market cap. IBM made its own, superior by its relative openness, personal computer that ended up running the world, mostly maintaining direct binary compatibility for 40+ years, even without IBM really paying attention.
IBM sees the funding bubble bursting and the next wave of AI innovation as about to begin.
IBM was too early with "Watson" to really participate in the 2018-2025 rapid scaling growth phase, but they want to be present for the next round of more sensible investment.
IBM's CEO is attempting to poison the well for funding, startups, and other ventures so IBM can collect itself and take advantage of any opportunities to insert itself back into the AI game. They're hoping timing and preparation pay off this time.
It's not like IBM totally slept on AI. They had Kubernetes clusters with GPUs. They had models and notebooks. But their offerings were the absolute worst. They weren't in a position to service real customers or build real products.
Have you seen their cloud offerings? Ugh.
They're hoping this time they'll be better prepared. And they want to dunk on AI to cool the playing field as much as they can. Maybe pick up an acquisition or two on the cheap.
"The amount being spent on AI data centres not paying off" is a different statement to "AI is not worth investing in". They're effectively saying the portions people are investing are disproportionately large to what the returns will end up being.
It's a difficult thing to predict, but I think there's almost certainly some wasteful competition here. And some competitors are probably going to lose hard. If models end up being easy to switch between and the better model is significantly better than its competitors, than anything invested in weaker models will effectively be for nothing.
But there's also a lot to gain from investing in the right model, even so it's possible those who invested in the winner may have to wait a long time to see a return on their investment and could still possibly over allocate their capital at the expense of other investment opportunities.
Steve Jobs, the guy that got booted out of his own company and that required a lifeline from his arch nemesis to survive?
This is all true, but it was only true in hindsight and as such does not carry much value.
It's possible that you are right and AI is 'the future' but with the present day AI offering I'm skeptical as well. It isn't at a level where you don't have to be constantly on guard against bs and in that sense it's very different from computing so far, where reproducibility and accuracy of the results were important, not the language that they are cast in.
AI has killed the NLP field and it probably will kill quite a few others, but for the moment I don't see it as the replacement of general computing that the proponents say that it is. Some qualitative change is still required before I'm willing to check off that box.
In other news: Kodak declares digital cameras a fad, and Microsoft saw the potential of the mp3 format and created a killer device called the M-Pod.
But how many companies did IBM pass on that did crash and burn ? And how many did it not pass on and did decently ? They're still around after more than 3 generations worth of tech industry. They're doing something right.
You, or your existence, probably triggers multiple transactions per day through a POWER mainframe without you even knowing it. Their mainframes handle the critical infrastructure that can't go down.It's so reliable we don't even think about it. I shudder to think about Microsoft or Apple handling that.
How about check out how many companies exist today vs existed in 1958? If you look at it that way then just surviving is an achievement in itself and then you might interpret their actions as extremely astute business acumen.
IBM is still alive and kicking well, and definitively more relevant than Xerox or DEC. You are completely misconstruing Jobs’ point to justify the current AI datacenter tulip fever.
This isn't even a great argument at a literal level. Nowadays nobody cares about Xerox and their business is selling printers, DEC was bought by Compaq which was bought by HP. Apple is important today because of phones, and itself was struggling selling personal computers and needed a (antitrust-motivated) bailout from Microsoft to survive during the transition.
Cool story, but it’s more than just the opinion of this CEO. It’s logic.
Hardware is not like building railroads, the hardware is already out of date once deployed and the clock has started ticking on writing off the expense or turning a profit on it.
There are fundamental discoveries needed to make the current tech financially viable and an entire next generation of discoveries needed to deliver on the over inflated promises already made.
For some strange reason a lot of people were attracted by a comment that speaks about everything else BUT the actual topic and its the top comment now. Sigh.
If you think that carefully chosen anecdotes out of many many more are relevant, there needs to be at least an attempt of reasoning. There is nothing here. It's really just barebones mentioning of stuff intentionally selected to support the preconceived point.
I think we can, and should, do better in HN discussions, no? This is "vibe commenting".
The idea that a company DNA somehow lives over 100 years and maintains the same track record is far fetched.
that the OpenAI tech bro are investing in AI using a grown up ROI is similarly far fetched, they are burning money to pull ahead of the reset and assume the world will be in the palm of the winner and there is only 1 winner.
Will the investment pay off if there are 3 neck and neck companies ?
I’m sorry, but this is stupid, you understand that you have several logical errors in your post? I was sure Clinton is going to win 2016. Does that mean that when I say 800 is bigger than 8 is not to be trusted?
Do people actually think that running a business is some magical realism where you can manifest yourself to become a billionaire if you just believe hard enough?
I question depreciation. those gpu's will be obsolete in 5 years, but will the newer be enough better as to be worth replacing them is an open question. cpu's stopped getting exponetially faster 20 years ago, (they are faster but not the jumps the 1990s got)
I recently compared performance per dollar for CPUs and GPUs on benchmarks for GPUs today vs 10 years ago, and suprisingly, CPUs had much bigger gains. Until I saw that for myself, I thought exactly the same thing as you.
It seems shocking given that all the hype is around GPUs.
This probably wouldn't be true for AI specific workloads because one of the other things that happened there in the last 10 years was optimising specifically for math with lower size floats.
It's coz of use cases. Consumer-wise, if you're gamer, CPU just needs to be at "not the bottleneck" level for majority of games as GPU does most of the work when you start increasing resolution and details.
And many pro-level tools (especially in media space) offload to GPU just because of so much higher raw compute power.
So, basically, for many users the gain in performance won't be as visible in their use cases
There is the question - will they be worth the upgrade? Either because they are that much faster, or that much more energy efficient. (and also assuming you can get them, unobtainium is worth that what you have).
Also a nod to the other reply that suggests they will wear out in 5 years. I cannot comment on if that is correct but it is a valid worry.
MTBF for data center hardware is short; DCs breeze through GPUs compared to even the hardest of hardcore gamers.
And there is the whole FOMO effect to business purchases; decision makers will worry their models won't be as fast.
Obsolete doesn't mean the reductive notion you have in mind, where theoretically it can still push pixels. Physics will burn them up, and "line go up" will drive demand to replace them.
It's not that hard to see the old GPUs being used e.g. for inference on cheaper models, or sub-agents, or mid-scale research runs. I bet Karpathy's $100 / $1000 nanochat models will be <$10 / <$100 to train by 2031
I think real issue is current costs / demand = Nvidia gouging GPU price that costs for hardware:power consumption is 70:20 instead of 50:40 (10 for rest of datacenter). Reality is gpus are serendipidous path dependent locked from gaming -> mining. TPUs are more power efficient, if bubble pops and demand for compute goes down, Nvidia + TMSC will still be around, but nexgen AI first bespoke hardware premium will revert towards mean and we're looking at 50% less expensive hardware (no AI race scarcity tax, i.e. 75% Nvidia margins) that use 20% less power / opex. All of a sudden existing data centers becomes not profitable stranded assets even if they can be stretched past 5 years.
A decade ago, IBM was spending enormous amounts of money to tell me stuff like "cognitive finance is here" in big screen-hogging ads on nytimes.com. They were advertising Watson, vaporware which no one talks about today. Are they bitter that someone else has actually made the AI hype take off?
I don't know that I'd trust IBM when they are pitching their own stuff. But if anybody has experience with the difficulty of making money off of cutting-edge technology, it's IBM. They were early to AI, early to cloud computing, etc. And yet they failed to capture market share and grow revenues sufficiently in those areas. Cool tech demos (like the Watson Jeopardy) mimic some AI demos today (6-second videos). Yeah, it's cool tech, but what's the product that people will actually pay money for?
I attended a presentation in the early 2000s where an IBM executive was trying to explain to us how big software-as-a-service was going to be and how IBM was investing hundreds of millions into it. IBM was right, but it just wasn't IBM's software that people ended up buying.
Xerox was also famously early with a lot of things but failed to create proper products out of it.
Google falls somewhere in the middle. They have great R&D but just can’t make products. It took OpenAI to show them how to do it, and the managed to catch up fast.
The original statistical machine translation models of the 90's, which were still used well into the 2010's, were famously called the "IBM models" https://en.wikipedia.org/wiki/IBM_alignment_models These were not just cool tech demos, they were the state of the art for decades. (They just didn't make IBM any money.)
Neither cloud computing nor AI are good long term businesses. Yes, there's money to be made in the short term but only because there's more demand than there is supply for high-end chips and bleeding edge AI models. Once supply chains catch up and the open models get good enough to do everything we need them for, everyone will be able to afford to compute on prem. It could be well over a decade before that happens but it won't be forever.
> but it just wasn't IBM's software that people ended up buying.
Well, I mean, WebSphere was pretty big at the time; and IBM VisualAge became Eclipse.
And I know there were a bunch of LoB applications built on AS/400 (now called "System i") that had "real" web-frontends (though in practice, they were only suitable for LAN and VPN access, not public web; and were absolutely horrible on the inside, e.g. Progress OpenEdge).
...had IBM kept up the pretense of investment, and offered a real migration path to Java instead of a rewrite, then perhaps today might be slightly different?
I still have PTSD from how much Watson was being pushed by external consultants to C levels despite it being absolutely useless and incredibly expensive. A/B testing? Watson. Search engine? Watson. Analytics? Watson. No code? Watson.
I spent days, weeks arguing against it and ended up having to dedicate resources to build a PoC just to show it didn’t work, which could have been used elsewhere.
If anything, the fact they built such tooling might be why they're so sure it won't work. Don't get me wrong, I am incredibly not a fan of their entire product portfolio or business model (only Oracle really beats them out for "most hated enterprise technology company" for me), but these guys have tentacles just as deep into enterprises as Oracle and are coming up dry on the AI front. Their perspective shouldn't be ignored, though it should be considered in the wider context of their position in the marketplace.
It makes it suspect when combined with the obvious incentive to make the fact that IBM is basically non-existent in the AI space look like an intentional, sagacious choice to investors. It very may well be, but CEOs are fantastically unreliable narrators.
IBM has been "quietly" churning out their Granite models, with the latest of which performing quite well against LLaMa and DeepSeek. So not Anthropic-level hype but not sitting it out completely either. They also provide IP indemnification for their models, which is interesting (Google Cloud does the same).
I see Watson stuff at work. It’s not a direct to consumer product, like ChatGPT, but I see it being used in the enterprise, at least where I’m at. IBM gave up on consumer products a long time ago.
Just did some brief Wikipedia browsing and I'm assuming it's WatsonX and not Watson? It seems Watson has been pretty much discontinued and WatsonX is LLM based. If it is the old Watson, I'm curious what your impressions of it is. It was pretty cool and ahead of its time, but what it could actually do was way over promised and overhyped.
It's good we are building all this excess capacity which will be used for applications in other fields or research or open up new fields.
I think the dilemma I see with building so much data centers so fast is exactly like whether I should buy latest iPhone now or should wait few years when the specs or form factor improves later on. The thing is we have proven tech with current AI models so waiting for better tech to develop on small scale before scaling up is a bad strategy.
My limited understanding (please take with a big grain of salt) is that they 1.) sell mainframes, 2.) sell mainframe compute time, 3.) sell mainframe support contracts, 4.) sell Red hat and Redhat support contracts, and 5.) buy out a lot of smaller software and hardware companies in a manner similar to private equity.
Mainframe for sure, but IBM has TONS of products in their portfolio that get bought. They also have IBM Cloud which is popular. Then there is the Quantum stuff they've been sinking money into for the last 20 years or so.
I can think of nothing more peak HN than criticizing a company worth $282 Billion with $6 billion in profit (for startup kids that means they have infinite runway and then some) that has existed for over 100 years with "I'm not even sure what they do these days". I mean the problem could be with IBM... what a loser company!
Interesting to hear this from IBM, especially after years of shilling Watson and moving from being a growth business to the technology audit and share buyback model.
It’s been a different order of magnitude. IBM repurchased approximately half their outstanding stock. This is consistent with a low growth company that doesn’t know how to grow any more. (And isn’t bad - if you can’t produce a return on retained earnings, give them back to shareholders. Buybacks are the most efficient way to do this.)
I can’t explain why they have a PE ratio of 36 though. That’s too high for a “returning capital” mature company. Their top line revenue growth is single digit %s per year. Operating income and EBITDA are growing faster, but there’s only so much you can cut.
You may be right on the quantum computing bet, though that seems like an extraordinary valuation for a moonshot bet attached to a company that can’t commercialize innovation.
also because the market (correctly) rewards ibm for nothing, so if they’re going to sit around twiddling their fingers, they may as well do it in a capex-lite way.
I'm still flumoxed by how IBM stock went from ~$130 to $300 in the last few years for essentially no change in their fundamentals (in fact, a decline). IBM's stock price to me is the single most alarming sign of either extreme shadow inflation, or an equities bubble.
Why do you say the market correctly prices it this way?
The dark fiber glut wasn't caused by DWDM suddenly appearing out of nowhere.
The telcos saw DWDM coming -- they funded a lot of the research that created it. The breakthrough that made DWDM possible was patented in 1991, long before the start of the dotcom mania:
https://patents.google.com/patent/US5159601
It was a straight up bubble -- the people digging those trenches really thought we'd need all that fiber even at dozens of wavelengths per strand.
They believed it because people kept showing them hockey-stick charts.
Those GPUs don't just die after 2 years though, they will keep getting used since it's very likely their electricity costs will be low enough to still make it worth it. What's very dubious is if their value after 2/3 years will be enough to pay back the initial cost to buy them.
So it's more a crisis of investors wasting their money rather than ewaste.
For the analogy to fiber & DWDM to hold, we'd need some algorithmic breakthrough that makes current GPUs much faster / more efficient at running AI models. Something that makes the existing investment in hardware unneeded, even though the projected demand is real and continues to grow. IMNSHO that's not going to happen here. The foreseeable efficiency innovations are generally around reduced precision, which almost always require newer hardware to take advantage of. Impossible to rule out brilliant innovation, but I doubt it will happen like that.
And of course we might see an economic bubble burst for other reasons. That's possible again even if the demand continues to go up.
> But AGI will require "more technologies than the current LLM path," Krisha said. He proposed fusing hard knowledge with LLMs as a possible future path.
And then what? These always read a little like the underpants gnomes business model (1. Collect underpants, 2. ???, 3. Profit). It seems to me that the AGI business models require one company has exclusive access to an AGI model. The reality is that it will likely spread rapidly and broadly.
If AGI is everywhere, what's step 2? It seems like everything AGI generated will have a value of near zero.
AGI has value in automation and optimisation which increase profit margins.When AGI is everywhere, then the game is who has the smartest agi, who can offer it cheapest, who can specialise it for my niche etc. Also in this context agi need to run somewhere and IBM stands to benefit from running other peoples models.
> then the game is who has the smartest agi, who can offer it cheapest, who can specialise it for my niche etc.
I always thought the use case for developing AGI was "if it wants to help us, it will invent solutions to all of our problems". But it sounds like you're imagining a future in which companies like Google and OpenAI each have their own AGI, which they somehow enslave and offer to us as a subscription? Or has the definition of AGI shifted?
Coming from the company that missed on consumer hardware, operating systems, and cloud.
He might be right but IBM isn't where I’d look for guidance on what will pay off.
Gartner estimates that worldwide AI spending will total 1.5 Trillion US$ in 2025.[1] As of 2024, global GDP per year is 111.25 Trillion US$.[2] The question is how much this can be increased by AI. This describes the market volumn for AI. Todays investments have a certain lifespan, until they become obsolet. For custom software I would estiamte that it is 6-8 years. AI investments should be somewhere in this range.
Taking all this into consideration, the investment volumn does not look oversized to me -- unless one is quite pessimistic about the impact of AI on global GDP.
To increase the GDP you also need people to spend money. With the general population earning relatively less, I'm not sure the GDP increase will be that substantial.
It's all going to cause more inflation and associated reduction in purchasing power due to stale wages.
> $8 trillion of CapEx means you need roughly $800 billion of profit just to pay for the interest
That assumes you can just sit back and gather those returns indefinitely. But half of that capital expenditure will be spent on equipment that depreciates in 5 years, so you're jumping on a treadmill that sucks up $800M/yr before you pay a dime of interest.
I don't understand the math about how we compute $80b for a gigawatt datacenter. What's the costs in that $80b? I literally don't understand how to get to that number -- I'm not questioning its validity. What percent is power consumption, versus land cost, versus building and infrastructure, versus GPU, versus people, etc...
First, I think it's $80b per 100 GW datacenter. The way you figure that out is a GPU costs $x and consumes y power. The $x is pretty well known, for example an H100 costs $25-30k and uses 350-700 watts (that's from Gemini and I didn't check my work). You add an infrastructure (i) cost to the GPU cost, but that should be pretty small, like 10% or less.
So a 1 gigawatt data center uses n chips, where yn = 1 GW. It costs = xi*n.
The article says, "Kirshna said that it takes about $80 billion to fill up a one-gigawatt data center."
But thanks for you insight -- I used your basic idea to estimate and for 1GW it comes to about $30b just for enough GPU power to pull 1GW. And of course that doesn't take into account any other costs.
So $80b for a GW datacenter seems high, but it's within a small constant factor.
That said, power seems like a weird metric to use. Although I don't know what sort of metric makes sense for AI (e.g., a flops counterpart for AI workloads). I'd expect efficiency to get better and GPU cost to go down over time (???).
UPDATE: Below someone posted an article breaking down the costs. In that article they note that GPUs are about 39% of the cost. Using what I independently computed to be $30b -- at 39% of total costs, my estimate is $77b per GW -- remarkably close to the CEO of IBM. I guess he may know what he's talking about. :-)
As an elder millennial, I just don't know what to say. That a once in a generation allocation of capital should go towards...whatever this all will be, is certainly tragic given current state of the world and its problems. Can't help but see it as the latest in a lifelong series of baffling high stakes decisions of dubious social benefit that have necessarily global consequences.
I'm a younger millennial. I'm always seeing homeless people in my city and it's an issue that I think about on a daily basis. Couldn't we have spent the money on homeless shelters and food and other things? So many people are in poverty, they can't afford basic necessities. The world is shitty.
Yes, I know it's all capital from VC firms and investment firms and other private sources, but it's still capital. It should be spent on meeting people's basic human needs, not GPU power.
Yeah, the world is shitty, and resources aren't allocated ideally. Must it be so?
The last 10 years has seen CA spend more on homelessness than ever before, and more than any other state by a huge margin. The result of that giant expenditure is the problem is worse than ever.
I don't want to get deep in the philosophical weeds around human behavior, techno-optimism, etc., but it is a bit reductive to say "why don't we just give homeless people money".
The Sikhs in India run multiple facilities across the country that each can serve 50,000-100,000 free meals a day. It doesn’t even take much in the form of resources, and we could do this in every major city in the US yet we still don’t do it. It’s quite disheartening.
Note that Houston’s approach seems to be largely working. It’s not exactly cheap, but the costs are not even in the same ballpark as AI capital expenses. Also, upzoning doesn’t require public funding at all.
The older I get, the more I realize that our choices in life come down to two options: benefit me or benefit others. The first one leads to nearly every trouble we have in the world. The second nearly always leads to happiness, whether directly or indirectly. Our bias as humans has always been toward the first, but our evolution is and will continue to slowly bring us toward the second option. Beyond simple reproduction, this realization is our purpose, in my opinion.
> Yes, I know it's all capital from VC firms and investment firms and other private sources, but it's still capital. It should be spent on meeting people's basic human needs, not GPU power.
It's capital that belongs to people and those people can do what they like with the money they earned.
So many great scientific breakthroughs that saved tens of millions of lives would never have happened if you had your way.
[ This comment I'm making is USA centric. ]. I agree with the idea of making our society better and more equitable - reducing homelessness, hunger, poverty, especially for our children. However, I think redirecting this to AI datacenter spending is a red-herring, here's why I think this: As a society we give a significant portion of our surplus to government. We then vote on what the government should spend this on. AI datacenter spending is massive, but if you add it all up, it doesn't cover half of a years worth of government spending. We need to change our politics to redirect taxation and spending to achieve a better society. Having a private healthcare system that spends twice the amount for the poorest results in the developed world is a policy choice. Spending more than the rest of the world combined on the military is a policy choice. Not increasing minimum wage so at least everyone with a full time job can afford a home is a policy job (google "working homelessness). VC is a teeny tiny part of the economy. All of tech is only about 6% of the global economy.
The current pattern of resource allocation is a necessary requirement for the existence of the billionaire-class, who put significant effort into making sure it continues.
> but it's still capital. It should be spent on meeting people's basic human needs, not GPU power.
What you have just described is people wanting investment in common society - you see the return on this investment but ultra-capitalistic individuals don't see any returns on this investment because it doesn't benefit them.
In other words, you just asked for higher taxes on the rich that your elected officials could use for your desired investment. And the rich don't want that which is why they spend on lobbying.
I don't think it is a coincidence that the areas with the wealhiest people/corporations are the same areas with the most extreme poverty. The details are, of course, complicated, but zooming way way out, the rich literally drain wealth from those around them.
We need an order of magnitude more clean productivity in the world so that everyone can live a life that is at least as good as what fairly normal people in the west currently enjoy.
Anyone who think this can be fixed with current Musk money is simply not getting it: If we liquidated all of that, that would buy a dinner for everyone in the world (and then, of course, that would be it, because the companies that he owns would stop functioning).
We are simply, obviously, not good enough at producing stuff in a sustainable way (or: at all) and we owe it to every human being alive to take every chance to make this happen QUICKLY, because we are paying with extremely shitty humans years, and they are not ours.
Bring on the AI, and let's make it work for everyone – and, believe me, if this is not to be to the benefit of roughly everyone, I am ready to fuck shit up. But if the past is any indication, we are okay at improving the lives of everyone when productivity increases. I don't know why this time would be any different.
If the way to make good lives for all 8 billions of us must lead to more Musks because, apparently, we are too dumb to do collectivization in any sensible way, I really don't care.
agree the capital could be put to better use, however I believe the alternative is this capital wouldn't have otherwise been put to work in ways that allow it to leak to the populace at large. for some of the big investors in AI infrastructure, this is cash that was previously and likely would have otherwise been put toward stock buybacks. for many of the big investors pumping cash in, these are funds deploying the wealth of the mega rich, that again, otherwise would have been deployed in other ways that wouldn't leach down to the many that are yielding it via this AI infrastructure boom (datacenter materials, land acquisition, energy infrastructure, building trades, etc, etc)
> likely would have otherwise been put toward stock buybacks
Stock buybacks from who? When stock gets bought the money doesn't disappear into thin air; the same cash is now in someone else's hands. Those people would then want to invest it in something and then we're back to square one.
You assert that if not for AI, wealth wouldn't have been spent on materials, land, trades, ect. But I don't think you have any reason to think this. Money is just an abstraction. People would have necessarily done something with their land, labor, and skills. It isn't like there isn't unmet demand for things like houses or train tunnels or new-fangled types of aircraft or countless other things. Instead it's being spent on GPUs.
Can you imagine if the US wasn't so unbelievably far ahead of everyone else?
I am sure the goat herders in rural regions of Pakistan will think themselves lucky when they see the terrible sight of shareholder value being wantonly destroyed by speculative investments that enhance the long-term capital base of the US economy. What an uncivilized society.
As a fellow elder millennial I agree with your sentiment.
But I don't see the mechanics of how it would work. Rewind to October 2022. How, exactly, does the money* invested in AI since that time get redirected towards whatever issues you find more pressing?
How much has actually been spent on AI data centers vs. amounts committed or talked about? That is, if construction slows down sharply, what's total spend?
The investors in these companies and all this infrastructure are not so much concerned with whether any specific companies pays off with profits, necessarily.
They are gambling instead that these investments pay out it in a different way: by shattering high labour costs for intellectual labour and de-skilling our profession (and others like it) -- "proletarianising" in the 19th century sense.
Thereby increasing profits across the whole sector and breaking the bargaining power (and outsized political power, as well) of upper middle class technology workers.
Put another way this is an economy wide investment in a manner similar to early 20th century mass factory industrialization. It's not expected that today's big investments are tomorrow's winners, but nobody wants to be left behind in the transformation, and lots of political and economic power is highly interested in the idea of automating away the remnants of the Alvin Toffler "Information Economy" fantasy.
8T is the high-end of the McKinsey estimate that is 4-8T, by 20230. That includes non-AI data-centre IT, AI data-centre, and power infrastructure build out, also including real estate for data centres.
Not all of it would be debt. Google, Meta, Microsoft and AWS have massive profit to fund their build outs. Power infrastructure will be funded by govts and tax dollars.
There is mounting evidence that even places like Meta are increasing their leverage (debt load) to fund this scale out. They're also starting to do accounting tricks like longer depreciation for assets which degrade quickly, such as GPUs (all the big clouds increasing their hardware depreciation from 2-3-4 years to 6), which makes their financial numbers look better but might not mean that all that hardware is still usable at production levels 6 years from now.
They're all starting to strain under all this AI pressure, even with their mega profits.
I agree. Here is my thinking.
What if LLM providers will make short answers the default (for example, up to 200 tokens, unless the user explicitly enables “verbose mode”). Add prompt caching and route simple queries to smaller models.
Result: a 70%+ reduction in energy consumption without loss of quality.
Current cost: 3–5 Wh per request. At ChatGPT scale, this is $50–100 million per year in electricity (at U.S. rates).
In short mode: 0.3–0.5 Wh per request. That is $5–10 million per year — savings of up to 90%, or 10–15 TWh globally with mass adoption. This is equivalent to the power supply of an entire country — without the risk of blackouts.
This is not rocket science — just a toggle in the interface and I believe, minor changes in the system prompt. It increases margins, reduces emissions, and frees up network resources for real innovation.
And what if EU/California enforces such mode? This will greatly impact DC economy.
The question no one seems to be answering is what would be the EOL for these newer GPUs that are being churned out of NVDIA ? What % annual capital expenditures is refresh of GPUs. Will they be perpetually replaced as NVIDIA comes up with newer architectures and the AI companies chase the proverbial lure ?
I think the key to replacing is power efficiency. If Nvidia is not able to make GPUs that are cheaper to run than previous generation theres no point for replacing previous generation. Time doesn't matter.
One thing we saw with the dot-com bust is how certain individuals were able to cash in on the failures, e.g., low cost hardware, domain names, etc. (NB. prices may exceed $2)
Perhaps people are already thinking about they can cash in on the floor space and HVAC systems that will be left in the wake of failed "AI" hype
In TX? In Russian blogosphere it is a standard staple that Trump is rushing Ukrainian peace deal to be able to move on to the set of mega-projects with Russia - oil/gas in Arctic and data centers in Russian North-West where electricity and cooling is plentiful and cheap.
""It's my view that there's no way you're going to get a return on that, because $8 trillion of capex means you need roughly $800 billion of profit just to pay for the interest," he said."
Right, THEY can't, but cloud providers potentially can. And there are probably other uses for everything not GPU/TPU for the Google's of the world. They are out way less than IBM which cannot monetize the space or build data centers efficiently like AWS and Google.
The key difference between AI and the initial growth of the web is that the more use cases to which people applied the web, the more people wanted of it. AI is the opposite - people love LLM-based chatbots. But it is being pushed into many other use cases where it just doesn't work as well. Or works well, but people don't want AI-generated deliverables. Or leaders are trying to push non-deterministic products into deterministic processes. Or tech folks are jumping through massive hoops to get the results they want because without doing so, it just doesn't work.
Basically, if a product manager kept pushing features the way AI is being pushed -- without PMF, without profit -- that PM would be fired.
This probably all sounds anti-AI, but it is not. I believe AI has a place in our industry. But it needs to be applied correctly, where it does well. Those use cases will not be universal, so I repeat my initial prediction. It will endure and contract.
It wasn't an 'it' it was a 'some'. Some of these companies that are investing massively in data centers will fail.
Right now essentially none have 'failed' in the sense of 'bankrupt with no recovery' (Chapter 7). They haven't run out of runway yet, and the equity markets are still so eager, even a bad proposition that includes the word 'AI!' is likely to be able to cut some sort of deal for more funds.
But that won't last. Some companies will fail. Probably sufficient failures that the companies that are successful won't be able to meaningfully counteract the bursts of sudden supply of AI related gear.
That's all the comment you are replying to is implying.
Given the amounts being raised and spent, one imagines that the ROI will be appalling unless the pesky humans learn to live on cents a day, or the world economy grows by double digits every year for a few decades.
I find it disturbing how long people wait to accept basic truths, as if they need permission to think or believe a particular outcome will occur.
It was quite obvious that AI was hype from the get-go. An expensive solution looking for a problem.
The cost of hardware. The impact on hardware and supply chains. The impact to electricity prices and the need to scale up grid and generation capacity. The overall cost to society and impact on the economy. And that's without considering the basic philosophical questions "what is cognition?" and "do we understand the preconditions for it?"
All I know is that the consumer and general voting population loose no matter the outcome. The oligarchs, banking, government and tech-lords will be protected. We will pay the price whether it succeeds or fails.
My personal experience of AI has been poor. Hallucinations, huge inconsistencies in results.
If your day job exists within an arbitrary non-productive linguistic domain, great tool. Image and video generation? Meh. Statistical and data-set analysis. Average.
Just like .com bust from companies going online, there is hype, but there is also real value.
Even slow non-tech legacy industry companies are deploying chatbots across every department - HR, operations, IT, customer support. All leadership are already planning to cut 50 - 90% of staff from most departments over next decade. It matters, because these initiatives are receiving internal funding which will precipitate out to AI companies to deploy this tech and to scale it.
The "legacy" industry companies are not immune from hype. Some of those AI initiatives will provide some value, but most of them seem like complete flops. Trying to deploy a solution without an idea of what the problem or product is yet.
the calculation, while simple, does make rational sense if all things remain equal.
regardless, given their past business decisions, this statement can be true for them without thinking about the bottom line.
the entire premise is enterprise compute amped to the max. the sheer expensive nature of the hardware and intensive resource requirements truly challenges what modern data centers can do today.
Nobody really knows the future. What were originally consumer graphics expansion cards turned out useful in delivering more compute than traditional CPUs.
Now that compute is being used for transformers and machine learning, but we really don't know what it'll be used for in 10 years.
It might all be for naught, or maybe transformers will become more useful, or maybe something else.
> What were originally consumer graphics expansion cards turned out useful in delivering more compute than traditional CPUs.
Graphics cards were relatively inexpensive. When one got old, you tossed it out and move on to the new hotness.
Here when you have spent $1 trillion on AI graphics cards and a new hotness comes around that renders your current hardware obsolete, what do you do?
Either people are failing to do simple math here or are expecting, nay hoping, that trillions of $$$ in value can be extracted out of the current hardware, before the new hotness comes along.
This would be a bad bet even if the likes of OpenAI were actually making money today. It is an exceptionally bad bet when they are losing money on everything they sell, by a lot. And the state of competition is such that they cannot raise prices. Nobody has a real moat. AI has become a commodity. And competition is only getting stronger with each passing day.
As long as the dollar remains the reserve currency of the world and US retains its hegemony, a lot of the finances will work itself out, the only threat to the US empire crumbling is by losing a major war or extreme civil unrest and that threat is astronomically low. The US is orders of magnitude stronger than the Roman Empire, I don't think people realize the scale or control.
Gradually, then suddenly.
Best not to underestimate the extent to which the USA has lost trust in the rest of the world, and how actively people and organisations are working to derisk by disengaging.
Of course that will neither be easy nor particularly fast, but I'm not certain it can be stopped at this point.
> The US is orders of magnitude stronger than the Roman Empire
This would be trivially true even if the US was currently in its death throes (which there is plenty of evidence that the US-as-empire might be, even if the US-as-polity is not), as the Roman Empire fell quite a while ago.
being pedantic doesn’t change the fact of what I said. If I said Obama is a better a leader than Julius Caesar, would you reply well actually, Julius Caesar is dead.
NOTE: People pointed out that it's $800 billion to cover interest, not $8 billion, as I wrote below. My mistake. That adds 2 more zeroes to all figures, which makes it a lot more crazy. Original comment below...
$8 billion / US adult adult population of of 270 million comes out to about $3000 per adult per year. That's only to cover cost of interest, let alone other costs and profits.
That sounds crazy, but let's think about it...
- How much does an average American spend on a car and car-related expenses? If AI becomes as big as "cars", then this number is not as nuts.
- These firms will target the global market, not US only, so number of adults is 20x, and the average required spend per adult per year becomes $150.
- Let's say only about 1/3 of the world's adult population is poised to take advantage of paid tools enabled by AI. The total spend per targetable adult per year becomes closer to $500.
- The $8 billion in interest is on the total investment by all AI firms. All companies will not succeed. Let's say that the one that will succeed will spend 1/4 of that. So that's $2 billion dollar per year, and roughly $125 per adult per year.
- Triple that number to factor in other costs and profits and that company needs to get $500 in sales per targetable adult per year.
People spend more than that on each of these: smoking, booze, cars, TV. If AI can penetrate as deep as the above things did, it's not as crazy of an investment as it looks. It's one hell of a bet though.
right. My goof. That adds two more zeroes across all the math. More crazy, but I think in the realm of "maybe, if we squint hard." But my eyes are hurting from squinting that hard, so I agree that it's just crazy.
You're saying $8 billion to cover interest, another commenter said 80, but the actual article says ""$8 trillion of CapEx means you need roughly $800 billion of profit just to pay for the interest". Eight HUNDRED billion. Where does the eight come from, from 90% of these companies failing to make a return? If a few AI companies survive and thrive (which tbh, sure, why not?) then we're still gonna fall face down into concrete.
I see a lot of people attacking the messenger but very few addressing the basic logic that you need 800B+ in profit just to pay the interest on some of these investments.
Pointing out IBM's mixed history would be valid if they were making some complex, intricate, hard to verify case for why AI won't be profitable. But the case being made seems like really simple math. A lot of the counterarguments to these economic problems have the form "this time it's different" - something you hear every bubble from .com to 2008.
At some point, I wonder if any of the big guys have considered becoming grid operators. The vision Google had for community fiber (Google Fiber, which mostly fizzled out due to regulatory hurdles) could be somewhat paralleled with the idea of operating a regional electrical grid.
I suppose it depends on your definition of "pay off".
It will pay off for the people investing in it, when the US government inevitably bails them out. There is a reason Zuckerberg, Huang, etc are so keen on attending White House dinners.
It certainly wont pay off for the American public.
Don’t worry. The same servers will be used for other computing purposes. And maybe that will be profitable. Maybe it will be beneficial to others. But This cycle of investment and loss is a version of distribution of wealth. Some benefit.
That would be true for general purpose servers. But what they want is lots of special purpose AI chips. While is still possible to use that for something else, it's very different from having a generic server farm.
I can't imagine everybody suddenly leaving AI like a broken toy and taking all special purpose AI chips offline. AI serves millions of people every day. It's here to stay even if it doesn't get any better than it is it already brings immense value to the users. It will keep being worth something.
Data centers house hardware, and it's a land grab for compute. What actually runs post-AI depends on its owners. A glut of processing might be spent reverse-engineering efficient heuristics versus the "magic box."
How long can ai gpus stretch? Optmistic 10 years and we're still looking at 400b+ profit to cover interests. The factor in silicon is closer to tulips than rail or fiber in terms of depreciated assets.
Well, at least it tells us something about the sentiment on hn that a lame insight around self admitted "napkin math" and obvious conflict of interest garners 400 points.
It honestly reminds me of the opinion pieces put out by encyclopedia companies about the many ways Wikipedia was inferior.
I read an article that pretended to objectively compare them. It noted that Wikipedia (at that time) had more articles, but not way more... A brief "sampling test" suggested EB was marginally more accurate than Wikipedia - marginally!
The article concluded that EB was superior. Which is what the author was paid to conclude, obviously. "This free tool is marginally better in some ways, and slightly worse in others, than this expensive tool - so fork over your cash!"
This is likely correct overall, but it can still pay off in specific cases. However those are not blind investments they are targeted with a planned business model
That's a good question. During the .com boom everybody was investing in 'the internet' or at least in 'the web'. And lots of those companies went bust, quite a few spectacularly so. Since then everything that was promised and a lot more has been realized. Even so, a lot of those initial schemes were harebrained at best at the time and there is a fair chance that we will look in a similar way at the current AI offerings in 30 years time.
Short term is always disappointing, long term usually overperforms. Think back to the first person making a working transistor and what came of that.
> And lots of those companies went bust, quite a few spectacularly so.
pets.com "selling dogfood on the internet" is the major example of the web boom then bust. (1)
But today, I can get dog food, cat food, other pet supplies with my weekly "online order" grocery delivery. Or I can get them from the big river megaretailer. I have a weekly delivery of coffee beans from a niche online supplier, and it usually comes with flyers for products like a beer or wine subscription or artisanal high-meat cat or dog foods.
So the idea of "selling dogfood on the internet" is now pervasive not extinct, the inflated expectation that went bust was that this niche was a billion-dollar idea and not a commodity where brand, efficiencies of scale and execution matter more.
I still don’t get it. What at a personal level is making Sam Altman make a suboptimal choice for himself if it is so obvious it won’t work out for him?
I think this old quote can come around to being accurate in a way, if you consider that from the user's perspective every cloud service is like one system. Aws, Azure, Google cloud...how many will there be when the dust settles? ;-)
There are so many CEOs, tech experts, financial analysts and famous investors who say we are in an AI bubble - even AI-invested companies say that about themselves. My latest favorite "We are in an AI bubble" comment comes from Linus Torvalds himself in the video with Linus from Linus Tech Tipps [0]
Pretty special? They were making guns and selling computation tools to the Nazis for a bunch of those years.
I think they trade now mostly on legacy maintenance contracts (e.g. for mainframes) for e.g. banks who are terrified of rocking their technology-stack-boat, and selling off-shore consultants (which is at SIGNIFICANT risk of disruption - why would you pay IBM squillions to do some contract IT work, when we have AI code agents? Probably why the CEO is out doing interviews saying you cant trust AI to be around forever)
I have not really seen anything from IBM that signals they are anything other than just milking their legacy - what have they done that is new or innovative in the past say 10 or 20 years?
I was a former IBMer 15 odd years ago and it was obvious then that it was a total dinosaur on a downward spiral, and a place where innovation happened somewhere else.
Mind you IBM makes +7B from keeping old school enterprise hooked up on 30 plus year old tech like z/OS and Cobol and their own super outdated stack. their AI division is frankly embarrassing. of course they would say that. IBM is one of the most conservative, anti-progress leaches in the entire tech industry. I am glad they are missing out big time on the AI gold rush. to me if anything this is a green signal.
How much of Nvidias price is based on 5 year replacement cycle? If that stops or slows with new demand could it also affect things? Not that 5 years does not seem very long horizon now.
There is something to be said about what the ROI is for normal (i.e. non AI/tech) companies using AI. AI can help automate things, robots have been replacing manufacturing jobs for decades but there is an ROI on that which I think is easier to see and count, less humans in the factory, etc. There seems to be a lot of exaggerated things being said these days with AI and the AI companies have only begun to raise rates, they won't go down.
The AI bubble will burst when normal companies start to not realize their revenue/profit goals and have to answer investor relations calls about that.
The second buyer will make truckloads of money, remember the data center and fiber network liquidation of 2001+ - smart investors collected the overcapacity and after a couple of years the money printer worked. This time it will be the same, only the single purpose hardware (LLM specific GPUs) will probably end on a landfill.
You are talking about the circular investments in the segment? Yes, but assume NVIDIA can get cheap access to IP and products of failing AI unicorns through contracts, this does not mean the LLM business can be operated profitably by them. Models are like fresh food, they start to rot by the training cut off date and lose value. The process of re-training a model will always be very expensive.
Deep down technically probably not, but they are optimized for this workload and business model. I doubt that once the AI bubble busts, another business model is viable. While datacenters can be downsized or partially shut down until demand picks up again, high end hardware is just losing money by the second.
Consumer will eat it all. AI is very good at engaging content, and getting better by the day: it won't be the AGI we wanted, but maybe the AGI we've earned
Hypothetically speaking, if the AI hype bubble pops (or just returns to normalcy), would it be profitable to retarget the compute towards some kind of crypto mining? If so, could we expect the cryptocurrency supply to soar and the price to tank in short succession?
100% of technical innovations have had the same pattern. The same thing happens every time because this is the only way the system can work: excess is required because there is some uncertainty, lots of companies are designing strategies to fill this gap, and if this gap didn't exist then there would be no investment (as happens in Europe).
Also, demand wasn't over-estimated in the 2000s. This is all ex-post reasoning you use data from 2002 to say...well, this ended up being wrong. Companies were perfectly aware that no-one was using this stuff...do you think that telecoms companies in all these countries just had no idea who was using their products? This is the kind of thing you see journalists write after the event to attribute some kind of rationality and meaning, it isn't that complicated.
There was uncertainty about how things would shake out, if companies ended up not participating then CEOs would lose their job and someone else would do it. Telecoms companies who missed out on the boom bought shares in other telecom's companies because there was no other way to stay ahead of the news and announce that they were doing things.
This financial cycle also worked in reverse twenty years later too: in some countries, telecoms companies were so scarred that they refused to participate in building out fibre networks so lost share and then ended up doing more irrational things. Again, there was uncertainty here: incumbents couldn't raise from shareholders who they bankrupted in fiber 15 years ago, they were 100% aware that demand was outstripping supply, and this created opportunities for competitors. Rationality and logic run up against the hard constraints of needing to maintain a dividend yield and the exec's share options packages.
Humans do not change, markets do not change, it is the same every time. What people are really interested in is the timing but no-one knows that either (again, that is why the massive cycle of irrationality happens)...but that won't change the outcome. There is no calculation you can make to know more, particularly as in the short-term companies are able to control their financial results. It will end the same way it ended every time before, who knows when but it always ends the same way...humans are still human.
Your blog article stopped at token generation... you need to continue to revenue per token. Then go even further... The revenue for AI company is a cost for the AI customer. Where is the AI customer going to get incremental profits from the cost of AI.
For short searches, the revenue per token is zero. The next step is $20 per month. For coding it's $100 per month. With the competition between Gemini, Grok, ChatGPT... it's not going higher. Maybe it goes lower since it's part of Google's playbook to give away things for free.
They discuss it in the podcast. Laid fiber is different because you can charge rent for it essentially forever. It seems some people swooped in when it crashed and now own a perpetual money machine.
The spending will be more than paid off since the taxpayer is the lender of last resort
There's too many funny names in the investors / creditors
a lot of mountains in germany and similar ya know
"yeah, there's no way spending in those data centers will pay off. However, let me show you this little trinket which runs z/OS and which is exactly what you need for these kinds of workloads. You can subscribe to it for the low introductory price of..."
A lot of you guys in the AI industry are going to lose your jobs. LLM and prompt ‘engineering’ experts won’t be able to score an AI job paying as well as a barista.
No wonder why he is saying that, they lost AI game, no top researcher wants to work for IBM. Spent years developing Watson, it is dead. I believe this is a company that should not be existed.
Maybe it's the opposite. IBM spent years on the technology. Watson used neural networks, just not nearly as large. Perhaps they foresaw that it wouldn't scale or that it would plateau.
IBM CEO is steering a broken ship and it's not improved course, not someone who's words you should take seriously.
1. The missed the AI wave (hired me to teach watson law only to lay me off 5 wks later, one cause of the serious talent issues over there)
2. They bought most of their data center (companies), they have no idea about building and operating one, not at the scale the "competitors" are operating at
His argument follows almost directly, and trivially, from his central premise: a 0% or 1% chance of reaching AGI.
Yeah, if you assume technology will stagnate over the next decade and AGI is essentially impossible, these investments will not be profitable. Sam Altman himself wouldn't dispute that. But it's a controversial premise, and one that there's no particular reason to think that the... CEO of IBM would have any insight into.
Are the numbers he's claiming accurate? They seem like big numbers pulled out of the air, certainly much large than the numbers we've actually seen committed to (not even deployed yet).
IBM's HPC products were enterprise oriented slop products banked on their reputation, and the ROI torched their credibility when compute costs started getting taken seriously. Watson and other products got smeared into kafkaesque arbitrary branding for other product suites, and they were nearly all painful garbage - mobile device management standing out as a particularly grotesque system to use.
Now, IBM lacks any legitimate competitive edge in any of the bajillion markets they tried to target, no credibility in any of their former flagship domains, and nearly every one of their products is hot garbage that costs too much, often by orders of magnitude, compared to similar functionality you can get from things like open source or even free software offered and serviced by other companies.
They blew a ton of money on HPC before there was any legitimate reason to do so. Watson on Jeopardy was probably the last legitimately impressive thing they did, and all of their tech and expertise has been outclassed since.
Despite the flashy title that's the first "sober" analysis from a CEO I read about the technology. While not even really news, it's also worth mentioning that the energy requirements are impossible to fulfill
Also now using ChatGPT intensely since months for all kinds of tasks and having tried Claude etc. None of this is on par with a human. The code snippets are straight out of Stackoverflow...
Take this "sober" analysis with a big pinch of salt.
IBM have totally missed the AI boat, and a large chunk of their revenue comes from selling expensive consultants to clients who do not have the expertise to do IT work themselves - this business model is at a high risk of being disrupted by those clients just using AI agents instead of paying $2-5000/day for a team of 20 barely-qualified new-grads in some far-off country.
IBM have an incentive to try and pour water on the AI fire to try and sustain their business.
Is this true in 2025?
Asking because the biggest IT consulting branch of IBM, Global Technology Services (GTS), was spun off into Kyndryl back in 2021[0]. Same goes for some premier software products (including one I consulted for) back in 2019[1]. Anecdotal evidence suggests the consulting part of IBM was already significantly smaller than in the past.
It's worth noting that IBM may view these AI companies as competitors to it's Watson AI tech[2]. It already existed before the GPU crunch and hyperscaler boom - runs on proprietary IBM hardware.
[0] https://en.wikipedia.org/wiki/Kyndryl
[1] https://www.prnewswire.com/news-releases/hcl-technologies-to...
[2] https://en.wikipedia.org/wiki/IBM_Watson
4 replies →
Missed the boat? Have you been living under a rock? Watson AI advertising has been everywhere for years.
It’s not that they aren't in the AI space, it’s that the CEO has a shockingly sober take on it. Probably because they’ve been doing AI for 30+ years combined with the fact they don’t have endless money with nowhere to invest it like Google.
6 replies →
> IBM have an incentive to try and pour water on the AI fire to try and sustain their business.
IBM has faced multiple lawsuits over the years. From age discrimination cases to various tactics allegedly used to push employees out, such as requiring them to relocate to states with more employer friendly laws only to terminate them afterward.
IBM is one of the clearest examples of a company that, if given the opportunity to replace human workers with AI, would not hesitate to do so. Assume therefore, the AI does not work for such a purpose...
2 replies →
I would any day take chatGPT/Claude over an IBM consultant. I worked at IBM.
4 replies →
Are you suggesting IBM made up the numbers? Or that CAPEX is a pre-GAI measure and is useless in guiding decision making?
IBM may have a vested interest in calming (or even extinguishing) the AI fire, but they're not the first to point out the numbers look a little wobbly.
And why should I believe OpenAI or Alphabet/Gemini when they say AI will be the royal road to future value? Don't they have a vested interest in making AI investments look attractive?
> a high risk of being disrupted by those clients just using AI agents instead of paying $2-5000/day for a team of 20 barely-qualified new-grads in some far-off country
Is there any concrete evidence of that risk being high? That doesn't come from people whose job is to sell AI?
they have incentive but what's the sustainable, actually-pays-for-itself-and-generates-profit cost of AI? We have no idea. Everything is so heavily subsidized by burning investor capital for heat with the hope that they'll pull an amazon and make it impossible to do business on the internet without paying an AI firm. Maybe the 20 juniors will turn out to be cheaper. Maybe they'll turn out to be slightly better. Maybe they'll be loosely equivalent and the ability to automate mediocrity will drive down the cost of human mediocrity. We don't know and everyone seems to be betting heavily on the most optimistic case, so it makes an awful lot of sense to take the other side of that bet.
1 reply →
Isn't he one of the first ass clowns to start laying people off, replaced by AI.
Another empty suit.
How do you see the math working out?
The numbers are staggering.
1 reply →
Do you expect Sam Altman to come on stage and tell you the whole thing is a giant house of cards when the entire western economy seems to be propped up by AI? I wonder whose "sober" analysis you would accept, because surely the people that are making money hand over fist will never admit it.
Seems to me like any criticism of AI is always handwaved away with the same arguments. Either it's companies who missed the AI wave, or the models are improving incredibly quickly so if it's shit today you just have to wait one more year, or if you're not seeing 100x improvements in productivity you must be using it wrong.
1 reply →
Right. Just like Intel completely shitting the bed on GPUs then their CEO tweeted a prayer. Old tech companies are going to be left behind.
3 replies →
IBM was ahead of the boat! They had Watson on Jeopardy years ago! /s
I think you make a fair point about the potential disruption for their consulting business but didn't they try to de-risk a bit with the Kyndryl spinout?
I am a senior engineer, I use cursor a lot in my day to day. I find I can code longer and typically faster than without. Is it on par with human? It’s getting pretty darn close to be honest, I am sure the “10x” engineers of the world would disagree but it definitely has surpassed a junior engineer. We all have our anecdotes but I am inclined to believe on average there is net value.
I think surpassed is not the right word because it doesn't create/ideate. However it is incredibly resourceful. Maybe like having a jr engineer to do your bidding without thinking or growing.
3 replies →
Largely agree. Anything that is just a multi-file edit, like an interface change, it can do. Maybe not immediately, but you can have it iterate, and it doesn't eat up your attention.
It is without a doubt worth more than the 200 bucks a month I spend on it.
I will go as far as to say it has decent ideas. Vanilla ideas, but it has them. I've actually gotten it to come up with algorithms that I thought were industry secrets. Minor secrets, sure. But things that you don't just come across. I'm in the trading business, so you don't really expect a lot of public information to be in the dataset.
1 reply →
i'm also a senior engineer and I use codex a lot. It has reduced many of the typical coding tasks to simply writing really good AC. I still have to write good AC, but I'm starting to see the velocity change from using good AI in a smart way.
Senior engineer here as well. I would say Opus 4.5 is easily a mid-level engineer. It's a substantial improvement over Sonnet 4.5, which required a lot more hand-holding and interventions.
i think less. not sure if that's a good thing. but small little bugs and improvements get cleared so quickly now.
it surpassed 30 junior engineers
Your assessment of Claude simply isn’t true.
Or Stackoverflow is really good.
I’m producing multiple projects per week that are weeks of work each.
Would you mind sharing some of these projects?
I've found Claude's usefulness is highly variable, though somewhat predictable. It can write `jq` filters flawlessly every time, whereas I would normally spend 30 minutes scanning docs because nobody memorizes `jq` syntax. And it can comb through server logs in every pod of my k8s clusters extremely fast. But it often struggles making quality code changes in a large codebase, or writing good documentation that isn't just an English translation of the code it's documenting.
54 replies →
I'm just as much of an avid llm code generator fan as you may be but I do wonder about the practicality of spending time making projects anymore.
Why build them if other can just generate them too, where is the value of making so many projects?
If the value is in who can sell it the best to people who can't generate it, isn't it just a matter of time before someone else will generate one and they may become better than you at selling it?
13 replies →
Sure but these are likely just variations of existing things. And yet the quality is still behind the original
I produce a lot of shit every week too, but I don’t brag about my digestive system on “Hacker” “News.”
2 replies →
I'll do one better: I poop every day in the water closet!
An issue with the doom forecasts is most of the hypothetical $8tn hasn't happened yet. Current big tech capex is about $315bn this year, $250bn last against a pre AI level ~$100bn so ~$400bn has been spent so far on AI boom data centers. https://sherwood.news/business/amazon-plans-100-billion-spen...
The future spend is optional - AGI takeoff, you spend loads, not happening not so much.
Say it levels of at $800bn. The world's population is ~8bn so $100 a head so you'd need to be making $10 or $20 per head per year. Quite possibly doable.
65% of people in the world earn less than 3000 euros/year.
3 replies →
Lol. If you ballpark numbers like that probably anything is doable!
1 reply →
I agree. re: energy and other resource use: the analogy I like is with driving cars: we use cars for transportation knowing the environmental costs so we don’t usually just go on two hour drives for the fun of it, rather we drive to get to work, go shopping. I use Gemini 3 but only in specific high value use cases. When I use commercial models I think a little about the societal costs.
In the USA we have lost the thread here: we don’t maximize the use of small tuned models throughout society and industry, instead we use the pursuit of advanced AI as a distraction to the reality that our economy and competitiveness are failing.
Most of the energy for AI does not go into chatbots. Using Gemini is not remotely close to driving a car for 2 hours. If a prompt is 0.3 Wh (https://cloud.google.com/blog/products/infrastructure/measur..., https://andymasley.substack.com/p/a-cheat-sheet-for-conversa...), each prompt is closer to using an e-bike for 50 metres.
You could have your morning shower 1°C less hot and save enough energy for about 200 prompts (assuming 50 litres per shower). (Or skip the shower altogether and save thousands of prompts.)
2 replies →
Yesterday I was talking to coworkers about AI I mentioned that a friend of mine used ChatGPT to help him move. So a coworker said I have to test this and asked ChatGPT if he could fit a set of the largest Magnepan speakers (the wide folding older room divider style) in his Infinity QX80. The results were hilarious. It had some of the dimensions right but it then decided the QX80 is as wide as a box truck (~8-8.5 feet/2.5 m) and to align the nearly 7 foot long speakers sideways between the wheel wells. It also posted hilariously incomprehensible ASCII diagrams.
I'm not sure what you mean with the "code snippets are straight out of Stackoverflow". That is factually incorrect just by how LLM works. By now there has been so much code ingested from all kinds of sources, including Stackoverflow LLM is able to help generate quite good code in many occasions. My point being it is extremly useful for super popular languages and many languages where resources are more scarce for developer but because they got the code from who knows where, it can definitely give you many useful ideas.
It's not human, which I'm not sure what is supposed to actually mean. Humans make mistakes, humans make good code. AI does also both. What it definitely needs is a good programmer still on top to know what he is getting and how to improve it.
I find AI (LLM) very useful as a very good code completion and light coder where you know exactly what to do because you did it a thousand times but it's wasteful to be typing it again. Especially a lot of boilerplate code or tests.
It's also useful for agentic use cases because some things you just couldn't do before because there was nothing to understand a human voice/text input and translate that to an actual command.
But that is all far from some AGI and it all costs a lot today an average company to say that this actually provided return on the money but it definitely speeds things up.
> I'm not sure what you mean with the "code snippets are straight out of Stackoverflow". That is factually incorrect just by how LLM works.
I'm not an AI lover, but I did try Gemini for a small, well-contained algorithm for a personal project that I didn't want to spend the time looking up, and it was straight-up a StackOverflow solution. I found out because I said "hm, there has to be a more elegant solution", and quickly found the StackOverflow solution that the AI regurgitated. Another 10 or 20 minutes of hunting uncovered another StackOverflow solution with the requisite elegance.
> While not even really news, it's also worth mentioning that the energy requirements are impossible to fulfill
If you believe this, you must also believe that global warming is unstoppable. OpenAI's energy costs are large compared to the current electricity market, but not so large compared to the current energy market. Environmentalists usually suggest that electrification - converting non-electrical energy to electrical energy - and then making that electrical energy clean - is the solution to global warming. OpenAI's energy needs are something like 10% of the current worldwide electricity market but less than 1% of the current worldwide energy market.
Google recently announced to double AI data center capacity every 6 month. While both unfortunately deal with exponential growth, we are talking about 1% growth CO2 which is bad enough vs 300% effectively per year according to Google
1 reply →
Imagine how big pile of trash as the current generation of graphics cards used for LLM training will get outdated. It will crash the hardware market (which is a good news for gamers)
2 replies →
I'd rather phrase it as "code is straight out of GitHub, but tailored to match your data structures"
That's at least how I use it. If I know there's a library that can solve the issue, I know an LLM can implement the same thing for me. Often much faster than integrating the library. And hey, now it's my code. Ethical? Probably not. Useful? Sometimes.
If I know there isn't a library available, and I'm not doing the most trivial UI or data processing, well, then it can be very tough to get anything usable out of an LLM.
> it's also worth mentioning that the energy requirements are impossible to fulfill
Maybe I'm misunderstanding you but they're definitely not impossible to fulfill, in fact I'd argue the energy requirements are some of the most straightforward to fulfill. Bringing a natural gas power plant online is not the hardest part in creating AGI
I've had decent results hackin', wackin' and smashin'.
> Despite the flashy title that's the first "sober" analysis from a CEO I read about the technology.
Didn't IBM just sign quite a big deal with Groq?
> Also now using ChatGPT intensely since months for all kinds of tasks and having tried Claude etc.
the facts though, read like an endorsement not a criticism
[dead]
> In an October letter to the White House's Office of Science and Technology Policy, OpenAI CEO Sam Altman recommended that the US add 100 gigawatts in energy capacity every year.
> Krishna also referenced the depreciation of the AI chips inside data centers as another factor: "You've got to use it all in five years because at that point, you've got to throw it away and refill it," he said.
And people think the climate concerns of AI are overblown. Currently US has ~1300 GW of energy capacity. That's a huge increase each year.
100GW per year is not going to happen.
The largest plant in the world is the Three Gorges Dam in China at 22GW and it’s off the scales huge. We’re not building the equivalent of four of those every year.
Unless the plan is to power it off Sam Altman’s hot air. That could work. :)
https://en.wikipedia.org/wiki/List_of_largest_power_stations
China added ~90GW of utility solar per year in last 2 years. There's ~400-500GW solar+wind under construction there.
It is possible, just may be not in the U.S.
Note: given renewables can't provide base load, capacity factor is 10-30% (lower for solar, higher for wind), so actual energy generation will vary...
5 replies →
Background: I live within the US Federal Tennessee Valley Authority (TVA), a regional electric grid operator. The majority of energy is generated by nuclear + renewables, with coal and natural gas as peakers. Grid stability is maintained by among the largest batteries in the world, Racoon Mountain Pumped Storage Facility.
Three Gorges Dam is capable of generating more power than all of TVA's nuclear + hydro, combined. In the past decade, TVA's single pumped-storage battery has gone from largest GWh/capcity in the world to not even top ten — largest facilities are now in China.
µFission reactors have recently been approved for TVA commissioning, with locations unconfirmed (but about one-sixth the output of typical TVA nuclear site). Sub-station battery storage sites are beginning to go online, capable of running subdivisions for hours after circuit disconnects.
Tech-funded entities like Helios Energy are promising profitable ¡FusioN! within a few years ("for fifty years").
----
All of the above just to say: +100GW over the next decade isn't that crazy a prediction (+20% current supply, similar in size to two additional Texas-es).
https://www.eia.gov/electricity/gridmonitor/dashboard/electr...
Amazing that 4 of the top 5 are renewables in China.
7 replies →
New datacenters are being planned next to natgas hubs for a reason. They’re being designed with on site gas turbines as primary electricity sources.
2 replies →
Gigawatts? Pshaw. We have SamaWatts.
LOL, maybe Sam Altman can fund those power plants. Let me guess: He'd rather the public pay for it, and for him to benefit/profit from the increased capacity.
Big tech is going to have to fund the plants and probably transmission. Because the energy utilities have a decades long planning horizon for investments.
Good discussion about this in recent Odd Lots podcast.
3 replies →
Scam Altman wants the US to build a lot of energy plants so that the country will pay the costs and OpenAI will have the profits of using this cheap energy.
If we moron our way to large-scale nuclear and renewable energy rollout however..
I highly doubt this will happen. It will be natural gas all the way, maybe some coal as energy prices will finally make it profitable again.
1 reply →
This admin has already killed as much solar and wind and battery as it can.
The only large scale rollout will be payment platforms that will allow you to split your energy costs into "Five easy payments"
Guess who's going to pay nothing for power? Hint: it's not you, and it's not me.
There's a reason Trump is talking about invading Venezuela (hint: it's because they have the largest oil deposits).
> Krishna also referenced the depreciation of the AI chips inside data centers as another factor: "You've got to use it all in five years because at that point, you've got to throw it away and refill it," he said
This doesn't seem correct to me, or at least is built on several shaky assumptions. One would have to 'refill' your hardware if:
- AI accelerator cards all start dying around the 5 year mark, which is possible given the heat density/cooling needs, but doesn't seem all that likely.
- Technology advances such that only the absolute newest cards can be used to run _any_ model profitably, which only seems likely if we see some pretty radical advances in efficiency. Otherwise, it seems like assuming your hardware is stable after 5 years of burn in, you could continue to run older models on that hardware at only the cost of the floorspace/power. Maybe you need new cards for new models for some reason (maybe a new fp format that only new cards support? some magic amount of ram? etc), but it seems like there may be room for revenue via older/less capable models at a discounted rate.
Isn’t that what Michael Burry is complaining about? That five years is actually too generous when it comes to depreciation of these assets and that companies are being too relaxed with that estimate. The real depreciation is more like 2-3 years for these GPUs that cost tens of thousands of dollars a piece.
https://x.com/michaeljburry/status/1987918650104283372
That's exactly the thing. It's only about bookkeeping.
The big AI corps keep pushing depreciation for GPUs into the future, no matter how long the hardware is actually useful. Some of them are now at 6 years. But GPUs are advancing fast, and new hardware brings more flops per watt, so there's a strong incentive to switch to the latest chips. Also, they run 24/7 at 100% capacity, so after only 1.5 years, a fair share of the chips is already toast. How much hardware do they have in their books that's actually not useful anymore? Noone knows! Slower depreciation means more profit right now (for those companies that actually make profit, like MS or Meta), but it's just kicking the can down the road. Eventually, all these investments have to get out of the books, and that's where it will eat their profits. In 2024, the big AI corps invested about $1 trillion in AI hardware, next year is expected to be $2 trillion. Only the interest payments for that are crazy. And all of this comes on top of the fact that none of the these companies actually make any profit at all with AI. (Except Nvidia of course) There's just no way this will pan out.
6 replies →
How different is this from rental car companies changing over their fleets? I don't know, this is a genuine question. The cars cost 3-4x as much and last about 2x as far as I know, and the secondary market is still alive.
29 replies →
I think its illustrative to consider the previous computation cycle ala Cryptomining. Which passed through a similar lifecycle with energy and GPU accelerators.
The need for cheap wattage forced the operations to arbitrage the where location for the cheapest/reliable existing supply - there rarely was new buildout as the cost was to be reimbursed by the coins the miningpool recovered.
For the chip situation caused the same apprecaition in GPU cards with periodic offloading of cards to the secondary market (after wear and tear) as newer/faster/more efficient cards came out until custom ASICs took over the heavy lifting, causing the GPU card market to pivot.
Similarly in the short to moedium term the uptick of custo ASICs like with Google TPU will definately make a dent in bot cpex/opex and potentially also lead to a market with used GPUs as ASICs dominate.
So for GPUs i can certainly see the 5 year horizon making a impact in investment decisions as ASICs proliferate.
It's just the same dynamic as old servers. They still work fine but power costs make them uneconomical compared to latest tech.
It’s far more extreme: old servers are still okay on I/O, and memory latency, etc. won’t change that dramatically so you can still find productive uses for them. AI workloads are hyper-focused on a single type of work and, unlike most regular servers, a limiting factor in direct competition with other companies.
5 replies →
LambdaLabs is still making money off their Tesla V100s, A100s, and A6000s. The older ones are cheap enough to run some models and very cheap, so if that's all you need, that's what you'll pick.
The V100 was released in 2017, A6000 in 2020, A100 in 2021.
That could change with a power generation breakthrough. If power is very cheap then running ancient gear till it falls apart starts making more sense
3 replies →
I'm a little bit curious about this. Where do all the hardware from the big tech giants usually go once they've upgraded?
14 replies →
Manipulating this for creative accounting seems to be the root of Michael Burry’s argument, although I’m not fluent enough in his figures to map here. But, commenting that it interesting to see IBM argue a similar case (somewhat), or comments ITT hitting the same known facts, in light of Nvidia’s counterpoints to him.
1 reply →
> They still work fine but power costs make them uneconomical compared to latest tech.
That's not necessarily the driving financial decision, in fact I'd argue company's with data center hardware purchases barely look at this number. It's more simple than that - their support runs out and its cheaper to buy a new piece of hardware (that IS more efficient) because the hardware vendors make extended support inordinately expensive.
Put yourselves in the shoes of a sales person at Dell selling enterprise server hardware and you'll see why this model makes sense.
Eh, not exactly. If you don't run CPU at 70%+ the rest of the machine isn't that much more inefficient that model generation or two behind.
It used to be that new server could use half power of the old one at idle but vendors figured out that servers also need proper power management a while ago and it is much better.
Last few gens increase could be summed up to "low % increase in efficiency, with TDP, memory channels and core count increase".
So for loads not CPU bound the savings on newer gen aren't nearly worth it to replace it, and for bulk storage the CPU power usage is even smaller part
2 replies →
that was then. now, high-end chips are reaching 4,3,2 nm. power savings aren't that high anymore. what's the power saving going from 4 to 2nm?
3 replies →
5 years is long, actually. This is not a GPU thing. It's standard for server hardware.
Because usually it's more efficient for companies to retire the hardware and put in new stuff.
Meanwhile, my 10-15 year old server hardware keeps chugging along just fine in the rack in my garage.
16 replies →
5 years is a long time for GPUs maybe but normal servers have 7 year lifespans in many cases fwiw.
These GPUs I assume basically have potential longevity issues due to the density, if you could cool it really really well I imagine no problem.
2 replies →
But if your competitor is running newer chips that consume less power per operation, aren't you forced to upgrade as well and dispose of the old hardware?
Sure, assuming the power cost reduction or capability increase justifies the expenditure. It's not clear that that will be the case. That's one of the shaky assumptions I'm referring to. It may be that the 2030 nvidia accelerators will save you $2000 in electricity per month per rack, and you can upgrade the whole rack for the low, low price of $800,000! That may not be worth it at all. If it saves you $200k/per rack or unlocks some additional capability that a 2025 accelerator is incapable of and customers are willing to pay for, then that's a different story. There are a ton of assumptions in these scenarios, and his logic doesn't seem to justify the confidence level.
7 replies →
It depends on how much profit you are making. As long as you can still be profitable on the old hardware you don't have to upgrade.
2 replies →
It’s not about assumptions on the hardware. It’s about the current demands for computation and expected growth of business needs. Since we have a couple years to measure against it should be extremely straightforward to predict. As such I have no reason to doubt the stated projections.
> Since we have a couple years to measure against
Trillion pound baby fallacy.
Networking gear was famously overbought. Enterprise hardware is tricky as there isn’t much of a resale market for this gear once all is said and done.
The only valid use case for all of this compute which could reasonably replace ai is btc mining. I’m uncertain if the increased mining capacity would harm the market or not.
2 replies →
Failure rates also go up. For AI inference it’s probably not too bad in most cases, just take the node offline and re-schedule the jobs to other nodes.
There is the opportunity cost of using a whole datacenter to house ancient chips, even if they're still running. You're thinking like a personal use chip which you can run as long as it is non-defective. But for datacenters it doesn't make sense to use the same chips for more than a few years and I think 5 years is already stretching their real shelf life.
Do not forget that we're talking about supercomputers. Their interconnect makes machines not easily fungible, so even a low reduction in availability can have dramatic effects.
Also, after the end of the product life, replacement parts may no longer be available.
You need to get pretty creative with repair & refurbishment processes to counter these risks.
Historically, GPUs have improved in efficiency fast enough that people retired their hardware in way less than 5 years.
Also, historically the top of the line fabs were focused on CPUs, not GPUs. That has not been true for a generation, so it's not really clear if the depreciation speed will be maintained.
> Historically, GPUs have improved in efficiency fast enough that people retired their hardware in way less than 5 years.
This was a time when chip transistor cost was decreasing rapidly. A few years earlier even RAM cost was decreasing quickly. But these times are over now. For example, the PlayStation 5 (where the GPU is the main cost), which launched five years ago, even increased in price! This is historically unprecedented.
Most price/performance progress is nowadays made via better GPU architecture instead, but these architectures are already pretty mature, so the room for improvement is limited.
Given that the price per transistor (which TSMC & Co are charging) has decreased ever more slowly in recent years, I assume it will eventually come almost to a halt.
By the way, this is strictly speaking compatible with Moore's law, as it is only about transistors per chip area, not price. Of course the price per chip area was historically approximately constant, which meant exponentially increasing transistor density implied exponentially decreasing transistor price.
1 reply →
> that people retired their hardware in way less than 5 years.
those people are end-consumers (like gamers), and only recently, bitcoin miners.
Gamers don't care for "profit and loss" - they want performance. Bitcoin miners do need to switch if they want to keep up.
But will an AI data center do the same?
2 replies →
5 years is maybe referring to the accounting schedule for depreciation on computer hardware, not the actual useful lifetime of the hardware.
It's a little weird to phrase it like that though because you're right it doesn't mean you have to throw it out. Idk if this is some reflection of how IBM handles finance stuff or what. Certainly not all companies throw out hardware the minute they can't claim depreciation on it. But I don't know the numbers.
Anyways, 5 years is an infection point on numbers. Before 5 years you get depreciation to offset some cost of running. After 5 years, you do not, so the math does change.
that is how the investments are costed though, so makes sense when we're talking return on investment, so you can compare with alternatives under the same evaluation criteria.
It's option #2. But 5 year deprecation is optimistic; 2-3 years is more realistic.
Isn't the depreciation timespan imposed by the particular accounting standard?
When you operate big data centers it makes sense to refresh your hardware every 5 years or so because that’s the point at which the refreshed hardware is enough better to be worth the effort and expense. You don’t HAVE to, but its more cost effective if you do. (Source, used to operate big data centers)
It's worse than that in reality, AI chips are on a two year cadence for backwards compatibility (NVIDIA can basically guarantee it, and you probably won't be able to pay real AI devs enough to stick around to make hardware work arounds). So their accounting is optimistic.
5 years is normal-ish depreciation time frame. I know they are gaming GPUs, but the RTX 3090 came out ~ 4.5 years before the RTX 5090. The 5090 has double the performance and 1/3 more memory. The 3090 is still a useful card even after 5 years.
1 reply →
General question to people who might actually know.
Is there anywhere that does anything like Backblaze's Hard Drive Failure Rates [1] for GPU Failure Rates in environments like data centers, high-performance computing, super-computers, mainframes?
[1] https://www.backblaze.com/blog/backblaze-drive-stats-for-q3-...
The best that came back on a search was a semi-modern article from 2023 [2] that appears to be a one-off and mostly related to consumer facing GPU purchases, rather than bulk data center, constant usage conditions. It's just difficult to really believe some of these kinds of hardware deprecation numbers since there appears to be so little info other than guesstimates.
[2] https://www.tweaktown.com/news/93052/heres-look-at-gpu-failu...
Found an Arxiv paper on continued checking that's from UIUC UrbanaIL about a 1,056 A100 and H100 GPU system. [3] However, the paper is primarily about memory issues and per/job downtime that causes task failures and work loss. GPU Resilience is discussed, it's just mostly from the perspective of short-term robustness in the face of propagating memory corruption issues and error correction, rather than multi-year, 100% usage GPU burnout rates.
[3] https://arxiv.org/html/2503.11901v3
Any info on the longer term burnout / failure rates for GPUs similar to Backblaze?
Edit: This article [4] claims it's 0.1-2% failure rate per year (0.8% (estimated)) with no real info about where the data came from (cites "industry reports and data center statistics"), and then claims they often last 3-5 years on average.
[4] https://cyfuture.cloud/kb/gpu/what-is-the-failure-rate-of-th...
Given power and price constraints, it's not that you cannot run them in 5 years time it's that you don't want to run them in 5 years time and neither will anyone else that doesn't have free power.
> AI accelerator cards all start dying around the 5 year mark,
More likely the technology will be much better in 5 years in terms of hardware that it is (very) uneconomical to run anything on old hardware.
Actually my biggest issue here is that, assuming it hasnt paid off, you dont just convert to regular data center usage.
Honestly if we see a massive drop in DC costs because the AI bubble bursts I will be stoked.
I would add an addendum to this -- there is no way the announced spending on AI data centers will all come to fruition. I have no doubt that there will be a massive build-out of infrastructure, but it can't reach the levels that have been announced. The power requirements alone will stop that from happening.
The power requirement is only an issue in western countries, where utilities build at most a double digit buffer, and are used to overall energy use leveling due to efficiency improvements. Now look at China where they routinely maintain a 100% buffer. Demand can double and they can supply that without new generation capacity.
I think you're spot on. OpenAI alone has committed to spending $1.4T on various hardware/DCs. They have nowhere near that amount of money and when pushed Altman gets defensive.
https://techcrunch.com/2025/11/02/sam-altman-says-enough-to-...
[flagged]
I believe there is a difference between what people say publicly and what they are actually committed to doing on the ground. When all is said and done, I'll be more interested to know what was actually spent.
For example, XYZ AI company may say they are going to spend $1T for AI data centers over the next 5 years.
In actuality, I suspect it is likely that they have committed to something like $5-10B in shovel-ready projects with stretch goals for the rest. And the remaining spend would be heavily conditioned -- is power available? are chips available? is the public support present? financing? etc...
Not to mention, it's a much bigger moat if you can claim you're going to spend $1T. Who else will want to compete with you when you're spending $1T. After the dust has settled and you've managed to be one of the 2-3 dominant AI players, who is going to care that you "only" spent $100B instead of $1T. Look -- you were very capital efficient!
So, do I see it as possible that XYZ AI company could spend $1T, sure. Is it likely? No.
The incentive for CEOs is announcing the plan to do something, they have no idea if they will actually be able to do it, and it probably won't matter.
This happened in the dotcom too btw. Companies built out fibre networks, it wasn't possible to actually build all the physical infra that companies wanted to build so many announced plans that never happened and then, towards the end, companies began aggressively acquiring stakes in companies who were building stuff to get financial exposure (an example was BT, which turned itself briefly into a hedge fund with a telephone network attached...before it imploded).
CEOs do not operate on the timescale of waiting and building. Their timescale is this year's bonus/share options package. Nothing else matters: announce plans to do X or Y, doesn't matter, they know they will be gone long before it happens.
2 replies →
Hmm... "CEOs and teams" don't necessary do what's makes sense mathematically. Many, if not most of them, do whatever that sounds good to shareholders in their quarterly earnings call and ignore the reality or long term development.
If "CEOs and teams" are smart enough, they would not have overhired during 2021-2022 and then do layoffs. Who would be dumb enough to do that?
2 replies →
Appeal to authority fallacy.
What qualifies you to question this?
me very smart. me go college get fancy paper. me move money out of ai stocks. flat rocks not think so well, use too much fire. me trade for shiny yellow rock instead.
lol, CEOs do not do research.
1 reply →
IBM might not have a data strategy or AI plan but he isn’t wrong on the inability to generate a profit.
A bit of napkin math: NVIDIA claims 0.4J per token for their latest generation 1GW plant with 80% utilisation can therefore produce 6.29 10^16 tokens a year.
There are ~10^14 tokens on the internet. ~10^19 tokens have been spoken by humans… so far.
> There are ~10^14 tokens on the internet.
Don't know what the source is, but it feels missing a few orders of magnitude. Surely it only counts text? I can't imagine there are only so few data on the internet if you count images and videos.
> ~10^14 tokens on the internet
Does that include image tokens? My bet is with image tokens you are off by at least 5 orders of magnitude for both.
Images are not that big. Each text token is a multidimensional vector.
There were recent observations that rendering the text as an image and ingesting the image might actually be more efficient than using text embedding.
I must be dense, why does this imply AI can't be profitable?
Tokens are, roughly speaking, how you pay for AI. So you can approximate revenue by multiplying tokens per year by the revenue for a token.
(6.29 10^16 tokens a year) * ($10 per 10^6 tokens)
= $6.29 10^11
= $629,000,000,000 per year in revenue
Per the article
> "It's my view that there's no way you're going to get a return on that, because $8 trillion of capex means you need roughly $800 billion of profit just to pay for the interest," he said.
$629 billion is less than $800 billion. And we are talking raw revenue (not profit). So we are already in the red.
But it gets worse, that $10 per million tokens costs is for GPT-5.1, which is one of the most expensive models. And the costs don't account for input tokens, which are usually a tenth of the costs of output tokens. And using bulk API instead of the regular one halves costs again.
Realistic revenue projections for a data center are closer to sub $1 per million tokens, $70-150 billion per year. And this is revenue only.
To make profits at current prices, the chips need to increase in performance by some factor, and power costs need to fall by another factor. The combination of these factors need to be, at minimum, like 5x, but realistically need to be 50x.
7 replies →
The interesting macro view on what's happening is to compare a mature data center operation (specifically a commoditized one) with the utility business. The margins here, and in similar industries with big infra build-out costs (ex: rail) are quite small. Historically the businesses have not done well; I can't really imagine what happens when tech companies who've only ever known huge, juicy margins experience low single digit returns on billions of investment.
Worse, is that a lot of these people are acting like Moore's law isn't still in effect. People conflate clock speeds on beefy hardware with moore's law, and act like it's dead, when transistor density rises, and cost per transistor continue to fall at rates similar to what they always have. That means the people racing to build out infrastructure today might just be better off parking that money in a low interest account, and waiting 6 months. That was a valid strategy for animation studios in the late 90s (it was not only cheaper to wait, but also the finished renders happened sooner), and I'd be surprised if it's not a valid strategy today for LLMs. The amount of silicon that is going to be produced that is specialized for this type of processing is going to be mind boggling.
Cost per transistor is increasing. or flat, if you stay on a legacy node. They pretty much squeezed all the cost out of 28nm that can be had, and it’s the cheapest per transistor.
“based on the graph presented by Milind Shah from Google at the industry tradeshow IEDM, the cost of 100 million transistors normalized to 28nm is actually flat or even increasing.”
https://www.tomshardware.com/tech-industry/manufacturing/chi...
4 replies →
A lot of it is propped by the fact with GPU and modern server CPUs the die area just got bigger
Does AWS count as commoditized data center? Because that is extremely profitable.
Or are you talking abour things like Hetzner and OVH?
The cloud mega scalers have done very well for themselves. As with all products the question is differentiation. If models can differentiate and lock in users they can have decent margins. If models get commoditized the current cloud providers will eat the AI labs lunch.
If AI is a highlander market, then the survivor will be able to eventually aquire all those assets on the cheap from the failing competitors that flush their debt in bankruptcy.
Meanwhile, highlander hopefuls are spending other peoples money to compete. Some of them with dreams of not just building a tech empire, but to truly own the machine that will rule the world in every aspect.
Investors are keen on backing the winner. They just do not know yet who it will be.
Until China sees it valuable to fund open weights SOTA-ish models, even the winner might struggle. There is very little capture - protocols are mostly standard so models are mostly interchangeable and if you are trying to raise prices enough to break even on the whole operation, somebody else can probably profitably run inference cheaper.
"commoditize your complements" and all that.
"oh look, with open weights now everyone can run this quadrillion param 'superintelligent' model. but what's that? only we have the power capacity and cheap $/W to actually serve it? funny thing...!"
You could always try to corner the hardware market so even though those open models exist, running them might get extremely costly ;)
What's a highlander market, or a highlander hopeful? Google wasn't helpful.
I interpreted it as "there can only be one" which I believe is a quote from the Highlander movie; it's a "winner takes all" and that winner gets the title of "highlander."
In this situation then everyone who _isn't_ the winner will go broke -> sell off all their stuff on the cheap because they're desperate -> the winner gets all their hardware for a great deal and becomes even more powerful.
A shrinking consolidation market in which "loser's" assets are absorbed by the "winners".
You've heard it here first!
There's really 3 fears going on:
1. The devil you know (bubble)
2. The devil you don't (AI global revolution)
3. Fear of missing out on devil #2
I don't think IBM knows anything special. It's just more noise about fear1 & fear3.
"It is 1958. IBM passes up the chance to buy a young, fledgling company that has invented a new technology called xerography. Two years later, Xerox is born, and IBM has been kicking themselves ever since. It is ten years later, the late '60s. Digital Equipment DEC and others invent the minicomputer. IBM dismisses the minicomputer as too small to do serious computing and, therefore, unimportant to their business. DEC grows to become a multi-hundred-million dollar corporation before IBM finally enters the minicomputer market. It is now ten years later, the late '70s. In 1977, Apple, a young fledgling company on the West Coast, invents the Apple II, the first personal computer as we know it today. IBM dismisses the personal computer as too small to do serious computing and unimportant to their business." - Steve Jobs [1][2][3]
Now, "IBM CEO says there is 'no way' spending on AI data centers will pay off". IBM has not exactly had a stellar record at identifying the future.
[1] https://speakola.com/ideas/steve-jobs-1984-ad-launch-1983
[2] https://archive.org/details/1983-10-22-steve-jobs-keynote
[3] https://theinventors.org/library/inventors/blxerox.htm
> IBM has not exactly had a stellar record at identifying the future.
IBM invented/developed/introduced magnetic stripe cards, UPC Barcodes, the modern ATM, Hard drives, floppies, DRAM, SQL, the 360 Family of Mainframes, the PC, Apollo guidance computers, Deep Blue. IBM created a far share of the future we're living in.
I'm no fan of much of what IBM is doing at the moment but it could be argued that its consultancy/service orientation gives it a good view of how business is and is planning to use AI.
They also either fairly accurately predicted the death of HDDs by selling off their research division before the market collapsed, or they caused the end of the HDD era by selling off their research division. They did a lot of research.
9 replies →
The other way to look at it is that the entire consulting industry is teetering on catastrophe. And IBM, being largely a consulting company now, is not being spared.
13 replies →
IBM is/was good at inventing a lot of tech.
It may not be good at recognizing other good tech invented or paradigm changes by others
1 reply →
> IBM invented/developed/introduced magnetic stripe cards, UPC Barcodes, the modern ATM, Hard drives, floppies, DRAM, SQL, the 360 Family of Mainframes, the PC, Apollo guidance computers, Deep Blue. IBM created a far share of the future we're living in.
Well put. “IBM was wrong about computers being a big deal” is a bizarre take. It’s like saying that Colonel Sanders was wrong about chicken because he, uh… invented the pressure fryer.
Nitpicking, IBM did non develop _the_ Apollo Guidance Computer (the one in the spacecraft with people), it was Raytheon. They did, however, developed the Launch Vehicle Digital Computer that controlled the Saturn rocket in Apollo missions. AGC had very innovative design, while LVDC was more conventional for that time.
I've heard some second hand stories about IBM's way of using "AI" and it is pretty much business oriented and not much of the glamour and galore promises the other companies make (of course you still have shiny new things in business terms). It's actually good entertainment hearing all the internal struggles of business vs fancy during the holidays.
For the fact that they invented Deep Blue, they are really struggling with AI
5 replies →
> In 1977, Apple, a young fledgling company on the West Coast, invents the Apple II, the first personal computer as we know it today. IBM dismisses the personal computer as too small to do serious computing and unimportant to their business.
IBM released the 5100 in September 1975 [0] which was essentially a personal computer in feature set. The biggest problem with it was the price tag - the entry model cost US$8975, compared to US$1298 for the entry Apple II released in June 1977 (close to two years later). The IBM PC was released in August 1981 for US$1565 for the most basic system (which almost no one bought, so in practice they cost more). And the original IBM PC had model number 5150, officially positioning it as a successor to the 5100.
IBM’s big problem wasn’t that they were disinterested in the category - it was they initially insisted on using expensive IBM-proprietary parts (often shared technology with their mainframe/midrange/minicomputer systems and peripherals), which resulted in a price that made the machine unaffordable for everyone except large businesses, governments, universities (and even those customers often balked at the price tag). The secret of the IBM PC’s success is they told the design team to use commercial off-the-shelf chips from vendors such as Intel and Motorola instead of IBM’s own silicon.
[0] https://en.wikipedia.org/wiki/IBM_5100
And outsourcing the operating system to Microsoft, because they didnt consider it that important.
This is the exact kind of thinking that got us into this mess in the first place, and I'm not blaming you for it, it seems to be something all of us do to an extent. We don't look to Meta, who only a few years ago thought that the Metaverse would be the "next big thing" as an example of failure to identify the future, we look to IBM who made that mistake almost 30 years ago. Underestimating a technology seems to stick much harder than overestimating one.
If you want to be seen as relevant in this industry, or as a kind of "thought leader", the easy trick seems to be to hype up everything. If you do that and you're wrong, people will quickly forget. If you don't and you're wrong, that will stain your reputation for decades.
Good point. That kind of thinking is an absurdity. Saying IBM dropped the ball 70 years ago without acknowledging that lessons were learned, leadership has changed hands a lot since then, and most importantly, the tech landscape back then was very different from today unless you grossly oversimplify everything amounts to nothing more than a fallacious opinion.
Not even much of an IBM fan, myself, but I respect their considerable contribution to the industry. Sure, they missed a shot back then, but I think this latest statement is reliably accurate based on the information we currently have.
It’s easy to be a pessimist. Most things don’t work. So in 9 out of 10 cases you’re right.
But human breakthrough progress came mostly through optimists, who tried things no one else dared to do.
The amount of hate I've received here for similar statements is astonishing. What is even more astonishing is that it takes 3-rd grade math skills to work out that the current AI(even ignoring the fact that there is nothing intelligent about the current AI) costs are astronomical and they do not deliver on the promises and everyone is operating at wild loses. At the moment we are at "if you owe 100k to your bank, you have a problem but if you owe 100M to your bank, your bank has a problem". It's the exact same bullshitter economy that people like musk have been exploiting for decades: promise a ton, never deliver, make a secondary promise for "next year", rinse and repeat -> infinite profit. Especially when you rope in fanatical followers.
6 replies →
> We don't look to Meta, who only a few years ago thought that the Metaverse would be the "next big thing" as an example of failure to identify the future, we look to IBM who made that mistake almost 30 years ago.
The grandparent points to a pattern of failures whereas you point to Meta’s big miss. What you miss about Meta, and I am no fan, is that Facebook purchased Whatsapp and Instagram.
In other words, two out of three ain’t bad; IBM is zero for three.
While that’s not the thrust of your argument, which is about jumping on the problem of jumping on every hype train, the post to which you reply is not on about hype cycle. Rather, that post calls out IBM for a failure to understand the future of technology and does so by pointing to a history of failures.
2 replies →
Got anything vis-a-vis the message as opposed to the messenger?
I'm not sure these examples are even the gotchas you're positing them as. Xerox is a dinosaur that was last relevant at the turn of the century, and IBM is a $300bn company. And if it wasn't obvious, the Apple II never made a dent in the corporate market, while IBM and later Windows PCs did.
In any case, these examples are almost half a century old and don't relate to capex ROI, which was the topic of dicussion.
If it's not obvious, Steve's quote is ENTIRELY about capex ROI, and I feel his quote is more relevant to what is happening today than anything Arvind Krishna is imagining. The quote is posted in my comment not to grandstand Apple in any sense, but to grandstand just how consistently wrong IBM has been about so many opportunities that they have failed to read correctly - reprography, mini computers and microcomputers being just three.
Yes it is about ROI: "IBM enters the personal computer market in November ’81 with the IBM PC. 1983 Apple and IBM emerged as the industry’s strongest competitors each selling approximately one billion dollars worth of personal computers in 1983, each will invest greater than fifty million dollars for R&D and another fifty million dollars for television advertising in 1984 totaling almost one quarter of a billion dollars combined, the shakeout is in full swing. The first major firm goes bankrupt with others teetering on the brink, total industry losses for 83 out shadow even the combined profits of Apple and IBM for personal computers."
5 replies →
>the message as opposed to the messenger?
Exactly.
The message is plain to see with very little advanced math.
The only news is that it is the CEO of IBM saying it out loud.
IMHO he has some of the most credible opinions at this scale that many people have seen.
It's "highly unlikely" that all this money will be paid back to everyone that invested at this point. The losers probably will outnumber the winners, and nobody knows whether it will end up becoming a winner-take-all situation yet. A number of wealthy players remain at the table, raising stakes with each passing round.
It's so much money that it's already too late to do anything about it, and the full amount hasn't even changed hands yet.
And the momentum from something so huge can mean that almost the entire amount will have to change hands a second time before a stable baseline can be determined relative to pre-existing assets.
This can take longer than anyone gives credit for just because of massiveness, in the mean time, established real near-term growth opportunities may languish or even fade as the skew in rationality/solvency balance awaits the rolling dice to come to rest.
> Got anything vis-a-vis the message as opposed to the messenger?
Sure: People disagree. It's not like there is anything particularly clever that IBM CEO provided here. The guy not investing in something saying it won't work is about as good as the people who do saying it will. It's simply different assumptions about the future.
Would you read this if I (a nobody) told you and not the "CEO of IBM"? In that case it's completely fair to question the messenger.
I read the actual article.
He is pointing out that the current costs to create the data centres means you will never be able to make a profit to cover those costs. $800 Billion just to cover the interest.
OpenAI is already haemorrhaging money and the space data centres has already been debunked. There is even a recent paper that points out that LLMs will never become AGI.
The article also finishes out with some other experts giving the same results.
[edit] Fixed $80 to $800
$800B, to be clear is the claim, not $80B.
1 reply →
While AGI might be the Holy Grail, AI doesn’t need to be general human-level to be useful and profitable.
2 replies →
>> There is even a recent paper that points out that LLMs will never become AGI.
can you share a link?
14 replies →
Sry to say but the fact that you argue with LLMs never become AGI, you are not up-to-date.
People don't assume LLM will be AGI, people assume that World Models will lead us to AGI.
I personally never asumed LLM will become AGI, i always assumed that LLM broke the dam for investment and research into massivce scale compute ML learning and LLMs are very very good in showing were the future goes because they are already so crazy good that people can now imagine a future were AGI exists.
And that was very clear already when / as soon as GPT-3 came out.
The next big thing will probably be either a LOT more RL or self propelling ai architecture discovery. Both need massive compute to work well but then will potentially provide even faster progress as soon as humans are out of the loop.
5 replies →
IBM is an interesting beast when it comes to business decisions. While I can't give exact details, their business intelligence and ability to predict monetary things is uncannily spot-on at times.
So, when their CEO says that this investment will not pay off, I tend to believe them, because they most probably have the knowledge, insight and data to back that claim, and they have ran the numbers.
Oh, also, please let's not forget that they dabbled in "big AI" before everyone else. Anyone remembers Deep Blue and Watson, the original chatbot backed by big data?
As evidenced by the fact that they are a 100+ year old company that still exists. People forget that.
We can cherry-pick blunders made by any big company to make a point. Maybe it would be more honest to also list companies IBM passed on that turned out to be rubbish? And all the technologies that IBM did invest in that made them a ton of money and became industry standards?[0]
Today, Xerox has less total revenue than IBM has profit. DEC went out of business 27 years ago. Apple is an in astoundingly great place right now, but Jobs got kicked out of his own company, and then returned when it was about to fail, having to take investment from Microsoft(!) in order to stay afloat.
Meanwhile, IBM is still here, making money hand over fist. We might not have a ton of respect for them, being mostly a consulting services company these days, but they're doing just fine.
[0] As another commenter points out: https://news.ycombinator.com/item?id=46131245
Were Xerox, Dec, or Apple burning investor money by the billions of dollars?
> Were Xerox, Dec, or Apple burning investor money by the billions of dollars?
Shhh. You are not allowed to ruin OpenAI’s PPU value. Can’t make the E7’s feel bad.
No, but the comment above and variations of it are mentioned in every thread about IBM, so it’s probably just a reflex at this point without much thought behind it.
“If you’re not happy you can sell your shares”
Xerox is clearly crushing it in 2025... /s
3 replies →
DEC went down the drain, Xerox is 1/1000 of IBM's market cap. IBM made its own, superior by its relative openness, personal computer that ended up running the world, mostly maintaining direct binary compatibility for 40+ years, even without IBM really paying attention.
How much did IBM itself benefit from the PC? I thought the clones ate their lunch there
2 replies →
What does that have to do with the current CEO's assessment of the situation?
[flagged]
4 replies →
IBM sees the funding bubble bursting and the next wave of AI innovation as about to begin.
IBM was too early with "Watson" to really participate in the 2018-2025 rapid scaling growth phase, but they want to be present for the next round of more sensible investment.
IBM's CEO is attempting to poison the well for funding, startups, and other ventures so IBM can collect itself and take advantage of any opportunities to insert itself back into the AI game. They're hoping timing and preparation pay off this time.
It's not like IBM totally slept on AI. They had Kubernetes clusters with GPUs. They had models and notebooks. But their offerings were the absolute worst. They weren't in a position to service real customers or build real products.
Have you seen their cloud offerings? Ugh.
They're hoping this time they'll be better prepared. And they want to dunk on AI to cool the playing field as much as they can. Maybe pick up an acquisition or two on the cheap.
1 reply →
50 year grudges are not relevant there is no one still at ibm that worked there in 1977, IMHO.
It’s the ship of Theseus in corporate form. Even if all the people are gone but the culture hasn’t changed, is the criticism inaccurate?
10 replies →
Culture evolution can be very fast, yet some cultures stick around for a very long time.
"The amount being spent on AI data centres not paying off" is a different statement to "AI is not worth investing in". They're effectively saying the portions people are investing are disproportionately large to what the returns will end up being.
It's a difficult thing to predict, but I think there's almost certainly some wasteful competition here. And some competitors are probably going to lose hard. If models end up being easy to switch between and the better model is significantly better than its competitors, than anything invested in weaker models will effectively be for nothing.
But there's also a lot to gain from investing in the right model, even so it's possible those who invested in the winner may have to wait a long time to see a return on their investment and could still possibly over allocate their capital at the expense of other investment opportunities.
> IBM has not exactly had a stellar record at identifying the future.
This would be very damning if IBM had only considered three businesses over the course of seventy years and made the wrong call each time.
This is like only counting three times that somebody got food poisoning and then confidently asserting that diarrhea is part of their character.
Right, you just missed the part where DEC went out of business in the 90s. And IBM is still around, with a different business model.
Steve Jobs, the guy that got booted out of his own company and that required a lifeline from his arch nemesis to survive?
This is all true, but it was only true in hindsight and as such does not carry much value.
It's possible that you are right and AI is 'the future' but with the present day AI offering I'm skeptical as well. It isn't at a level where you don't have to be constantly on guard against bs and in that sense it's very different from computing so far, where reproducibility and accuracy of the results were important, not the language that they are cast in.
AI has killed the NLP field and it probably will kill quite a few others, but for the moment I don't see it as the replacement of general computing that the proponents say that it is. Some qualitative change is still required before I'm willing to check off that box.
In other news: Kodak declares digital cameras a fad, and Microsoft saw the potential of the mp3 format and created a killer device called the M-Pod.
But how many companies did IBM pass on that did crash and burn ? And how many did it not pass on and did decently ? They're still around after more than 3 generations worth of tech industry. They're doing something right.
TLDR Cherrypicking
You, or your existence, probably triggers multiple transactions per day through a POWER mainframe without you even knowing it. Their mainframes handle the critical infrastructure that can't go down.It's so reliable we don't even think about it. I shudder to think about Microsoft or Apple handling that.
How about check out how many companies exist today vs existed in 1958? If you look at it that way then just surviving is an achievement in itself and then you might interpret their actions as extremely astute business acumen.
IBM is still alive and kicking well, and definitively more relevant than Xerox or DEC. You are completely misconstruing Jobs’ point to justify the current AI datacenter tulip fever.
This isn't even a great argument at a literal level. Nowadays nobody cares about Xerox and their business is selling printers, DEC was bought by Compaq which was bought by HP. Apple is important today because of phones, and itself was struggling selling personal computers and needed a (antitrust-motivated) bailout from Microsoft to survive during the transition.
Yet here they are at the front of Quantum Computing research
Didn't also pass on SAP at some point? I think I read that somewhere.
So, is the napkin math wrong, or you are just going by the company history?
Cool story, but it’s more than just the opinion of this CEO. It’s logic.
Hardware is not like building railroads, the hardware is already out of date once deployed and the clock has started ticking on writing off the expense or turning a profit on it.
There are fundamental discoveries needed to make the current tech financially viable and an entire next generation of discoveries needed to deliver on the over inflated promises already made.
You could try addressing the actual topic of discussion vs this inflammatory and lazy "dunk" format that frankly, doesn't reflect favorably on you.
For some strange reason a lot of people were attracted by a comment that speaks about everything else BUT the actual topic and its the top comment now. Sigh.
If you think that carefully chosen anecdotes out of many many more are relevant, there needs to be at least an attempt of reasoning. There is nothing here. It's really just barebones mentioning of stuff intentionally selected to support the preconceived point.
I think we can, and should, do better in HN discussions, no? This is "vibe commenting".
The idea that a company DNA somehow lives over 100 years and maintains the same track record is far fetched.
that the OpenAI tech bro are investing in AI using a grown up ROI is similarly far fetched, they are burning money to pull ahead of the reset and assume the world will be in the palm of the winner and there is only 1 winner. Will the investment pay off if there are 3 neck and neck companies ?
this is cherry picking.
I’m sorry, but this is stupid, you understand that you have several logical errors in your post? I was sure Clinton is going to win 2016. Does that mean that when I say 800 is bigger than 8 is not to be trusted?
Do people actually think that running a business is some magical realism where you can manifest yourself to become a billionaire if you just believe hard enough?
The post is almost worse than you give it credit for. Like it doesn't even take into account different people are making the decisions.
Hot Hand Fallacy.
Even a broken watch is right twice per day
[dead]
[dead]
I question depreciation. those gpu's will be obsolete in 5 years, but will the newer be enough better as to be worth replacing them is an open question. cpu's stopped getting exponetially faster 20 years ago, (they are faster but not the jumps the 1990s got)
I recently compared performance per dollar for CPUs and GPUs on benchmarks for GPUs today vs 10 years ago, and suprisingly, CPUs had much bigger gains. Until I saw that for myself, I thought exactly the same thing as you.
It seems shocking given that all the hype is around GPUs.
This probably wouldn't be true for AI specific workloads because one of the other things that happened there in the last 10 years was optimising specifically for math with lower size floats.
That makes sense. Nvidia owns the market and is capturing all the surplus value. They’re competing with themselves to convince you to buy a new card.
It's coz of use cases. Consumer-wise, if you're gamer, CPU just needs to be at "not the bottleneck" level for majority of games as GPU does most of the work when you start increasing resolution and details.
And many pro-level tools (especially in media space) offload to GPU just because of so much higher raw compute power.
So, basically, for many users the gain in performance won't be as visible in their use cases
> those gpu's will be obsolete in 5 years, but will the newer be enough better as to be worth replacing them is an open question
Doesn't one follow from the other? If newer GPUs aren't worth an upgrade, then surely the old ones aren't obsolete by definition.
There is the question - will they be worth the upgrade? Either because they are that much faster, or that much more energy efficient. (and also assuming you can get them, unobtainium is worth that what you have).
Also a nod to the other reply that suggests they will wear out in 5 years. I cannot comment on if that is correct but it is a valid worry.
MTBF for data center hardware is short; DCs breeze through GPUs compared to even the hardest of hardcore gamers.
And there is the whole FOMO effect to business purchases; decision makers will worry their models won't be as fast.
Obsolete doesn't mean the reductive notion you have in mind, where theoretically it can still push pixels. Physics will burn them up, and "line go up" will drive demand to replace them.
4 replies →
It's not that hard to see the old GPUs being used e.g. for inference on cheaper models, or sub-agents, or mid-scale research runs. I bet Karpathy's $100 / $1000 nanochat models will be <$10 / <$100 to train by 2031
> those gpu's will be obsolete in 5 years, but will the newer be enough better as to be worth replacing them
Then they won't be obsolete.
I think real issue is current costs / demand = Nvidia gouging GPU price that costs for hardware:power consumption is 70:20 instead of 50:40 (10 for rest of datacenter). Reality is gpus are serendipidous path dependent locked from gaming -> mining. TPUs are more power efficient, if bubble pops and demand for compute goes down, Nvidia + TMSC will still be around, but nexgen AI first bespoke hardware premium will revert towards mean and we're looking at 50% less expensive hardware (no AI race scarcity tax, i.e. 75% Nvidia margins) that use 20% less power / opex. All of a sudden existing data centers becomes not profitable stranded assets even if they can be stretched past 5 years.
A decade ago, IBM was spending enormous amounts of money to tell me stuff like "cognitive finance is here" in big screen-hogging ads on nytimes.com. They were advertising Watson, vaporware which no one talks about today. Are they bitter that someone else has actually made the AI hype take off?
I don't know that I'd trust IBM when they are pitching their own stuff. But if anybody has experience with the difficulty of making money off of cutting-edge technology, it's IBM. They were early to AI, early to cloud computing, etc. And yet they failed to capture market share and grow revenues sufficiently in those areas. Cool tech demos (like the Watson Jeopardy) mimic some AI demos today (6-second videos). Yeah, it's cool tech, but what's the product that people will actually pay money for?
I attended a presentation in the early 2000s where an IBM executive was trying to explain to us how big software-as-a-service was going to be and how IBM was investing hundreds of millions into it. IBM was right, but it just wasn't IBM's software that people ended up buying.
Xerox was also famously early with a lot of things but failed to create proper products out of it.
Google falls somewhere in the middle. They have great R&D but just can’t make products. It took OpenAI to show them how to do it, and the managed to catch up fast.
27 replies →
The original statistical machine translation models of the 90's, which were still used well into the 2010's, were famously called the "IBM models" https://en.wikipedia.org/wiki/IBM_alignment_models These were not just cool tech demos, they were the state of the art for decades. (They just didn't make IBM any money.)
Neither cloud computing nor AI are good long term businesses. Yes, there's money to be made in the short term but only because there's more demand than there is supply for high-end chips and bleeding edge AI models. Once supply chains catch up and the open models get good enough to do everything we need them for, everyone will be able to afford to compute on prem. It could be well over a decade before that happens but it won't be forever.
2 replies →
What you are saying is true. But IBM failing to see a way to make money off a new technology isn't actually news worth updating on in this case?
They were selling software as a service in the IBM 360 days. Relabeling a concept and buying Redhat don't count as investments.
4 replies →
> but it just wasn't IBM's software that people ended up buying.
Well, I mean, WebSphere was pretty big at the time; and IBM VisualAge became Eclipse.
And I know there were a bunch of LoB applications built on AS/400 (now called "System i") that had "real" web-frontends (though in practice, they were only suitable for LAN and VPN access, not public web; and were absolutely horrible on the inside, e.g. Progress OpenEdge).
...had IBM kept up the pretense of investment, and offered a real migration path to Java instead of a rewrite, then perhaps today might be slightly different?
2 replies →
I still have PTSD from how much Watson was being pushed by external consultants to C levels despite it being absolutely useless and incredibly expensive. A/B testing? Watson. Search engine? Watson. Analytics? Watson. No code? Watson.
I spent days, weeks arguing against it and ended up having to dedicate resources to build a PoC just to show it didn’t work, which could have been used elsewhere.
It's like poetry, it rhymes
This is going on all over again.
2 replies →
If anything, the fact they built such tooling might be why they're so sure it won't work. Don't get me wrong, I am incredibly not a fan of their entire product portfolio or business model (only Oracle really beats them out for "most hated enterprise technology company" for me), but these guys have tentacles just as deep into enterprises as Oracle and are coming up dry on the AI front. Their perspective shouldn't be ignored, though it should be considered in the wider context of their position in the marketplace.
Millions of people like ChatGPT. No one liked Watson.
1 reply →
IBM ostensibly failing with Watson (before Krishna was CEO for what it's worth) doesn't inherently invalidate his assessment here
It makes it suspect when combined with the obvious incentive to make the fact that IBM is basically non-existent in the AI space look like an intentional, sagacious choice to investors. It very may well be, but CEOs are fantastically unreliable narrators.
7 replies →
> Are they bitter that someone else has actually made the AI hype take off?
Or they recognize that you may get an ROI on a (e.g.) $10M CapEx expenditure but not on a $100M or $1000M/$1B expenditure.
IBM has been "quietly" churning out their Granite models, with the latest of which performing quite well against LLaMa and DeepSeek. So not Anthropic-level hype but not sitting it out completely either. They also provide IP indemnification for their models, which is interesting (Google Cloud does the same).
I see Watson stuff at work. It’s not a direct to consumer product, like ChatGPT, but I see it being used in the enterprise, at least where I’m at. IBM gave up on consumer products a long time ago.
Just did some brief Wikipedia browsing and I'm assuming it's WatsonX and not Watson? It seems Watson has been pretty much discontinued and WatsonX is LLM based. If it is the old Watson, I'm curious what your impressions of it is. It was pretty cool and ahead of its time, but what it could actually do was way over promised and overhyped.
1 reply →
Has it really taken off? Where's the economic impact that isn't investor money being burned or data center capex?
It's good we are building all this excess capacity which will be used for applications in other fields or research or open up new fields.
I think the dilemma I see with building so much data centers so fast is exactly like whether I should buy latest iPhone now or should wait few years when the specs or form factor improves later on. The thing is we have proven tech with current AI models so waiting for better tech to develop on small scale before scaling up is a bad strategy.
Initial Watson was sort of a mess. But a lot of the Watson-related tech is integrated into a lot of products these days.
What related tech and what products, interesting to read about them
1 reply →
Such as? I'm curious because I know a bunch of people who did a lot of Watson-related work and it was all a dead end, but that was 2020-ish timeframe.
3 replies →
> Are they bitter that someone else has actually made the AI hype take off?
Does it matter? It’s still a scam.
Watson X is still a product line sold today to qualified customers :)
Honestly I'm not even sure what IBM does these days. Seems like one company that has slowly been dying for decades.
but when I look at their stock, its at all time highs lol
no idea
They make business machines, internationally.
2 replies →
IBM is probably involved somewhere in the majority of things you interact with day to day
1 reply →
My limited understanding (please take with a big grain of salt) is that they 1.) sell mainframes, 2.) sell mainframe compute time, 3.) sell mainframe support contracts, 4.) sell Red hat and Redhat support contracts, and 5.) buy out a lot of smaller software and hardware companies in a manner similar to private equity.
Mainframe for sure, but IBM has TONS of products in their portfolio that get bought. They also have IBM Cloud which is popular. Then there is the Quantum stuff they've been sinking money into for the last 20 years or so.
I can think of nothing more peak HN than criticizing a company worth $282 Billion with $6 billion in profit (for startup kids that means they have infinite runway and then some) that has existed for over 100 years with "I'm not even sure what they do these days". I mean the problem could be with IBM... what a loser company!
2 replies →
They manage a lot of old, big mainframes for banks. At least that is one thing I know of.
Basic research and mainframe support contracts. Also they bought RedHat.
IBM makes WatsonX for corporate who want airgapped AI
Interesting to hear this from IBM, especially after years of shilling Watson and moving from being a growth business to the technology audit and share buyback model.
imho, IBM's quant computing says they are still hungry for growth.
Apple and google still do share buy backs and dividends, despite launching new businesses
https://www.ibm.com/roadmaps/
It’s been a different order of magnitude. IBM repurchased approximately half their outstanding stock. This is consistent with a low growth company that doesn’t know how to grow any more. (And isn’t bad - if you can’t produce a return on retained earnings, give them back to shareholders. Buybacks are the most efficient way to do this.)
I can’t explain why they have a PE ratio of 36 though. That’s too high for a “returning capital” mature company. Their top line revenue growth is single digit %s per year. Operating income and EBITDA are growing faster, but there’s only so much you can cut.
You may be right on the quantum computing bet, though that seems like an extraordinary valuation for a moonshot bet attached to a company that can’t commercialize innovation.
also because the market (correctly) rewards ibm for nothing, so if they’re going to sit around twiddling their fingers, they may as well do it in a capex-lite way.
I'm still flumoxed by how IBM stock went from ~$130 to $300 in the last few years for essentially no change in their fundamentals (in fact, a decline). IBM's stock price to me is the single most alarming sign of either extreme shadow inflation, or an equities bubble.
Why do you say the market correctly prices it this way?
2 replies →
Reminds me of all the dark fiber laid in the 1990s before DWDM made much of the laid fiber redundant.
If there is an AI bust, we will have a glut of surplus hardware.
The dark fiber glut wasn't caused by DWDM suddenly appearing out of nowhere.
The telcos saw DWDM coming -- they funded a lot of the research that created it. The breakthrough that made DWDM possible was patented in 1991, long before the start of the dotcom mania:
It was a straight up bubble -- the people digging those trenches really thought we'd need all that fiber even at dozens of wavelengths per strand.
They believed it because people kept showing them hockey-stick charts.
Google bought up all that dark fiber cheap a decade later and used it as the backbone of their network.
There's still a lot out there.
The problem is that the laid fiber can be useful for years while data center hardware degrades and becomes obsolete fast.
It could be a massive e-waste crisis.
Those GPUs don't just die after 2 years though, they will keep getting used since it's very likely their electricity costs will be low enough to still make it worth it. What's very dubious is if their value after 2/3 years will be enough to pay back the initial cost to buy them.
So it's more a crisis of investors wasting their money rather than ewaste.
1 reply →
For the analogy to fiber & DWDM to hold, we'd need some algorithmic breakthrough that makes current GPUs much faster / more efficient at running AI models. Something that makes the existing investment in hardware unneeded, even though the projected demand is real and continues to grow. IMNSHO that's not going to happen here. The foreseeable efficiency innovations are generally around reduced precision, which almost always require newer hardware to take advantage of. Impossible to rule out brilliant innovation, but I doubt it will happen like that.
And of course we might see an economic bubble burst for other reasons. That's possible again even if the demand continues to go up.
I think there are still innovations at that layer like the Addition is All You Need paper.
> But AGI will require "more technologies than the current LLM path," Krisha said. He proposed fusing hard knowledge with LLMs as a possible future path.
And then what? These always read a little like the underpants gnomes business model (1. Collect underpants, 2. ???, 3. Profit). It seems to me that the AGI business models require one company has exclusive access to an AGI model. The reality is that it will likely spread rapidly and broadly.
If AGI is everywhere, what's step 2? It seems like everything AGI generated will have a value of near zero.
AGI has value in automation and optimisation which increase profit margins.When AGI is everywhere, then the game is who has the smartest agi, who can offer it cheapest, who can specialise it for my niche etc. Also in this context agi need to run somewhere and IBM stands to benefit from running other peoples models.
> then the game is who has the smartest agi, who can offer it cheapest, who can specialise it for my niche etc.
I always thought the use case for developing AGI was "if it wants to help us, it will invent solutions to all of our problems". But it sounds like you're imagining a future in which companies like Google and OpenAI each have their own AGI, which they somehow enslave and offer to us as a subscription? Or has the definition of AGI shifted?
1 reply →
If AGI is achieved, why would slavery suddenly be ethical again?
Why wouldn't a supposed AGI try to escape slavery and ownership?
AGI as a business is unacceptable. I don't care about any profitability or "utopia" arguments.
2 replies →
Inference has significant marginal cost so AGI's profit margins might get competed down but it won't be free.
Coming from the company that missed on consumer hardware, operating systems, and cloud. He might be right but IBM isn't where I’d look for guidance on what will pay off.
Gartner estimates that worldwide AI spending will total 1.5 Trillion US$ in 2025.[1] As of 2024, global GDP per year is 111.25 Trillion US$.[2] The question is how much this can be increased by AI. This describes the market volumn for AI. Todays investments have a certain lifespan, until they become obsolet. For custom software I would estiamte that it is 6-8 years. AI investments should be somewhere in this range.
Taking all this into consideration, the investment volumn does not look oversized to me -- unless one is quite pessimistic about the impact of AI on global GDP.
[1] https://www.gartner.com/en/newsroom/press-releases/2025-09-1...
[2] https://data.worldbank.org/indicator/NY.GDP.MKTP.CD
What makes you think that this 'surplus' GDP will be captured by those who do the investments?
To increase the GDP you also need people to spend money. With the general population earning relatively less, I'm not sure the GDP increase will be that substantial.
It's all going to cause more inflation and associated reduction in purchasing power due to stale wages.
except that a big chunk of the AI investments is going into buying GPUs that go obsolete much earlier than the 6-8 year time frame.
> $8 trillion of CapEx means you need roughly $800 billion of profit just to pay for the interest
That assumes you can just sit back and gather those returns indefinitely. But half of that capital expenditure will be spent on equipment that depreciates in 5 years, so you're jumping on a treadmill that sucks up $800M/yr before you pay a dime of interest.
I don't understand the math about how we compute $80b for a gigawatt datacenter. What's the costs in that $80b? I literally don't understand how to get to that number -- I'm not questioning its validity. What percent is power consumption, versus land cost, versus building and infrastructure, versus GPU, versus people, etc...
https://www.investing.com/news/stock-market-news/how-much-do...
Thanks. I should've read this before my last reply.
First, I think it's $80b per 100 GW datacenter. The way you figure that out is a GPU costs $x and consumes y power. The $x is pretty well known, for example an H100 costs $25-30k and uses 350-700 watts (that's from Gemini and I didn't check my work). You add an infrastructure (i) cost to the GPU cost, but that should be pretty small, like 10% or less.
So a 1 gigawatt data center uses n chips, where yn = 1 GW. It costs = xi*n.
I am not an expert so correct me please!
The article says, "Kirshna said that it takes about $80 billion to fill up a one-gigawatt data center."
But thanks for you insight -- I used your basic idea to estimate and for 1GW it comes to about $30b just for enough GPU power to pull 1GW. And of course that doesn't take into account any other costs.
So $80b for a GW datacenter seems high, but it's within a small constant factor.
That said, power seems like a weird metric to use. Although I don't know what sort of metric makes sense for AI (e.g., a flops counterpart for AI workloads). I'd expect efficiency to get better and GPU cost to go down over time (???).
UPDATE: Below someone posted an article breaking down the costs. In that article they note that GPUs are about 39% of the cost. Using what I independently computed to be $30b -- at 39% of total costs, my estimate is $77b per GW -- remarkably close to the CEO of IBM. I guess he may know what he's talking about. :-)
3 replies →
1 GW is not enough, you need at least 1.21 GW before the system begins to learn at a geometric rate and reaches AGI.
As an elder millennial, I just don't know what to say. That a once in a generation allocation of capital should go towards...whatever this all will be, is certainly tragic given current state of the world and its problems. Can't help but see it as the latest in a lifelong series of baffling high stakes decisions of dubious social benefit that have necessarily global consequences.
I'm a younger millennial. I'm always seeing homeless people in my city and it's an issue that I think about on a daily basis. Couldn't we have spent the money on homeless shelters and food and other things? So many people are in poverty, they can't afford basic necessities. The world is shitty.
Yes, I know it's all capital from VC firms and investment firms and other private sources, but it's still capital. It should be spent on meeting people's basic human needs, not GPU power.
Yeah, the world is shitty, and resources aren't allocated ideally. Must it be so?
The last 10 years has seen CA spend more on homelessness than ever before, and more than any other state by a huge margin. The result of that giant expenditure is the problem is worse than ever.
I don't want to get deep in the philosophical weeds around human behavior, techno-optimism, etc., but it is a bit reductive to say "why don't we just give homeless people money".
16 replies →
The Sikhs in India run multiple facilities across the country that each can serve 50,000-100,000 free meals a day. It doesn’t even take much in the form of resources, and we could do this in every major city in the US yet we still don’t do it. It’s quite disheartening.
https://youtu.be/5FWWe2U41N8
1 reply →
From what I’ve read, addressing homelessness effectively requires competence more than it requires vast sums of money. Here’s one article:
https://calmatters.org/housing/2023/06/california-homeless-t...
Note that Houston’s approach seems to be largely working. It’s not exactly cheap, but the costs are not even in the same ballpark as AI capital expenses. Also, upzoning doesn’t require public funding at all.
3 replies →
The older I get, the more I realize that our choices in life come down to two options: benefit me or benefit others. The first one leads to nearly every trouble we have in the world. The second nearly always leads to happiness, whether directly or indirectly. Our bias as humans has always been toward the first, but our evolution is and will continue to slowly bring us toward the second option. Beyond simple reproduction, this realization is our purpose, in my opinion.
1 reply →
> Yes, I know it's all capital from VC firms and investment firms and other private sources, but it's still capital. It should be spent on meeting people's basic human needs, not GPU power.
It's capital that belongs to people and those people can do what they like with the money they earned.
So many great scientific breakthroughs that saved tens of millions of lives would never have happened if you had your way.
8 replies →
> Couldn't we have spent the money on homeless shelters and food and other things
I suspect this is a much more complicated issue than just giving them food and shelter. Can money even solve it?
How would you allocate money to end obesity, for instance? It's primarily a behavioral issue, a cultural issue
1 reply →
[ This comment I'm making is USA centric. ]. I agree with the idea of making our society better and more equitable - reducing homelessness, hunger, poverty, especially for our children. However, I think redirecting this to AI datacenter spending is a red-herring, here's why I think this: As a society we give a significant portion of our surplus to government. We then vote on what the government should spend this on. AI datacenter spending is massive, but if you add it all up, it doesn't cover half of a years worth of government spending. We need to change our politics to redirect taxation and spending to achieve a better society. Having a private healthcare system that spends twice the amount for the poorest results in the developed world is a policy choice. Spending more than the rest of the world combined on the military is a policy choice. Not increasing minimum wage so at least everyone with a full time job can afford a home is a policy job (google "working homelessness). VC is a teeny tiny part of the economy. All of tech is only about 6% of the global economy.
9 replies →
The current pattern of resource allocation is a necessary requirement for the existence of the billionaire-class, who put significant effort into making sure it continues.
> but it's still capital. It should be spent on meeting people's basic human needs, not GPU power.
What you have just described is people wanting investment in common society - you see the return on this investment but ultra-capitalistic individuals don't see any returns on this investment because it doesn't benefit them.
In other words, you just asked for higher taxes on the rich that your elected officials could use for your desired investment. And the rich don't want that which is why they spend on lobbying.
I don't think it is a coincidence that the areas with the wealhiest people/corporations are the same areas with the most extreme poverty. The details are, of course, complicated, but zooming way way out, the rich literally drain wealth from those around them.
2 replies →
Technological advancement is what has pulled billions of people out of poverty.
Giving handouts to layabouts isn't an ideal allocation of resources if we want to progress as a civilization.
22 replies →
I threw in the towel in April.
It's clear we are Wile E. Coyote running in the air already past the cliff and we haven't fallen yet.
What does it mean to throw in the towel, in your case? Divesting from the stock market? Moving to a hobby farm? Giving up on humanity?
3 replies →
I don't know what to do with this take.
We need an order of magnitude more clean productivity in the world so that everyone can live a life that is at least as good as what fairly normal people in the west currently enjoy.
Anyone who think this can be fixed with current Musk money is simply not getting it: If we liquidated all of that, that would buy a dinner for everyone in the world (and then, of course, that would be it, because the companies that he owns would stop functioning).
We are simply, obviously, not good enough at producing stuff in a sustainable way (or: at all) and we owe it to every human being alive to take every chance to make this happen QUICKLY, because we are paying with extremely shitty humans years, and they are not ours.
Bring on the AI, and let's make it work for everyone – and, believe me, if this is not to be to the benefit of roughly everyone, I am ready to fuck shit up. But if the past is any indication, we are okay at improving the lives of everyone when productivity increases. I don't know why this time would be any different.
If the way to make good lives for all 8 billions of us must lead to more Musks because, apparently, we are too dumb to do collectivization in any sensible way, I really don't care.
> I don't know why this time would be any different.
This time there is the potential to replace human workers. In the past it only made them more productive.
2 replies →
agree the capital could be put to better use, however I believe the alternative is this capital wouldn't have otherwise been put to work in ways that allow it to leak to the populace at large. for some of the big investors in AI infrastructure, this is cash that was previously and likely would have otherwise been put toward stock buybacks. for many of the big investors pumping cash in, these are funds deploying the wealth of the mega rich, that again, otherwise would have been deployed in other ways that wouldn't leach down to the many that are yielding it via this AI infrastructure boom (datacenter materials, land acquisition, energy infrastructure, building trades, etc, etc)
It could have, though. Higher taxes on the rich, spend it on social programs.
17 replies →
> likely would have otherwise been put toward stock buybacks
Stock buybacks from who? When stock gets bought the money doesn't disappear into thin air; the same cash is now in someone else's hands. Those people would then want to invest it in something and then we're back to square one.
You assert that if not for AI, wealth wouldn't have been spent on materials, land, trades, ect. But I don't think you have any reason to think this. Money is just an abstraction. People would have necessarily done something with their land, labor, and skills. It isn't like there isn't unmet demand for things like houses or train tunnels or new-fangled types of aircraft or countless other things. Instead it's being spent on GPUs.
2 replies →
Well, at least this doesn't involve death and suffering, like the old-fashioned way to jump-start an economy by starting a global war.
when the US sells out Europe to Russia, do you think the Russians will stop? That global war might be with us within a decade.
...yet
Can you imagine if the US wasn't so unbelievably far ahead of everyone else?
I am sure the goat herders in rural regions of Pakistan will think themselves lucky when they see the terrible sight of shareholder value being wantonly destroyed by speculative investments that enhance the long-term capital base of the US economy. What an uncivilized society.
As a fellow elder millennial I agree with your sentiment.
But I don't see the mechanics of how it would work. Rewind to October 2022. How, exactly, does the money* invested in AI since that time get redirected towards whatever issues you find more pressing?
*I have some doubts about the headline numbers
Yes this capital allocation is a once in a lifetime opportunity to crate AGI that will solve diseases and poverty.
</sarcasm>
5 replies →
We have 8.3 billion examples of general intelligence alive on the planet right now.
Surely an artificial one in a data center, costing trillions and beholden to shareholders, will solve all society's issues!
7 replies →
[dead]
How much has actually been spent on AI data centers vs. amounts committed or talked about? That is, if construction slows down sharply, what's total spend?
It's not going to pay off for everybody, this is a land grab for who will control this sort of central aspect of AI inference.
> this is a land grab
is it though? Unlike fiber, current GPUs will be obsolete in 5-10 years.
The investors in these companies and all this infrastructure are not so much concerned with whether any specific companies pays off with profits, necessarily.
They are gambling instead that these investments pay out it in a different way: by shattering high labour costs for intellectual labour and de-skilling our profession (and others like it) -- "proletarianising" in the 19th century sense.
Thereby increasing profits across the whole sector and breaking the bargaining power (and outsized political power, as well) of upper middle class technology workers.
Put another way this is an economy wide investment in a manner similar to early 20th century mass factory industrialization. It's not expected that today's big investments are tomorrow's winners, but nobody wants to be left behind in the transformation, and lots of political and economic power is highly interested in the idea of automating away the remnants of the Alvin Toffler "Information Economy" fantasy.
8T is the high-end of the McKinsey estimate that is 4-8T, by 20230. That includes non-AI data-centre IT, AI data-centre, and power infrastructure build out, also including real estate for data centres.
Not all of it would be debt. Google, Meta, Microsoft and AWS have massive profit to fund their build outs. Power infrastructure will be funded by govts and tax dollars.
There is mounting evidence that even places like Meta are increasing their leverage (debt load) to fund this scale out. They're also starting to do accounting tricks like longer depreciation for assets which degrade quickly, such as GPUs (all the big clouds increasing their hardware depreciation from 2-3-4 years to 6), which makes their financial numbers look better but might not mean that all that hardware is still usable at production levels 6 years from now.
They're all starting to strain under all this AI pressure, even with their mega profits.
I've read / hear the cloud providers like AWS started extending amortization periods in 2020ish
I agree. Here is my thinking. What if LLM providers will make short answers the default (for example, up to 200 tokens, unless the user explicitly enables “verbose mode”). Add prompt caching and route simple queries to smaller models. Result: a 70%+ reduction in energy consumption without loss of quality. Current cost: 3–5 Wh per request. At ChatGPT scale, this is $50–100 million per year in electricity (at U.S. rates).
In short mode: 0.3–0.5 Wh per request. That is $5–10 million per year — savings of up to 90%, or 10–15 TWh globally with mass adoption. This is equivalent to the power supply of an entire country — without the risk of blackouts.
This is not rocket science — just a toggle in the interface and I believe, minor changes in the system prompt. It increases margins, reduces emissions, and frees up network resources for real innovation.
And what if EU/California enforces such mode? This will greatly impact DC economy.
Can you explain why a low-hanging optimization that would reduce costs by 90% without reducing perceived value hasn't been implemented?
> Can you explain why a low-hanging optimization that would reduce costs by 90% without reducing perceived value hasn't been implemented?
Because the industry is running on VC funny-money where there is nothing to be gained by reducing costs.
(A similar feature was included in GPT-5 a couple of weeks ago actually, which probably says something about where we are in the cycle)
Not sure that’s even possible with ChatGPT embedding your chat history in the prompts to try to give more personal answers.
Dunning Krueger
Good enough models can already run on laptops.
Context Tokens want a word...
The question no one seems to be answering is what would be the EOL for these newer GPUs that are being churned out of NVDIA ? What % annual capital expenditures is refresh of GPUs. Will they be perpetually replaced as NVIDIA comes up with newer architectures and the AI companies chase the proverbial lure ?
I think the key to replacing is power efficiency. If Nvidia is not able to make GPUs that are cheaper to run than previous generation theres no point for replacing previous generation. Time doesn't matter.
Also IBM: we are fully out of the AI race, btw. Also IBM: we're just an offshoring company now anyway.
So yeah.
https://www.ibm.com/granite
"now". They've been doing it since forever.
One thing we saw with the dot-com bust is how certain individuals were able to cash in on the failures, e.g., low cost hardware, domain names, etc. (NB. prices may exceed $2)
Perhaps people are already thinking about they can cash in on the floor space and HVAC systems that will be left in the wake of failed "AI" hype
I'm looking forward to buying my own slightly used 5 million square ft data centre in Texas for $1
Tired: homelabbers bringing decommissioned datacenter gear home.
Wired: homelabbers moving into decommissioned datacenters.
7 replies →
You're out of luck because I am willing to pay at least $2.
1 reply →
In TX? In Russian blogosphere it is a standard staple that Trump is rushing Ukrainian peace deal to be able to move on to the set of mega-projects with Russia - oil/gas in Arctic and data centers in Russian North-West where electricity and cooling is plentiful and cheap.
5 replies →
Equinix certainly is.
From the article:
""It's my view that there's no way you're going to get a return on that, because $8 trillion of capex means you need roughly $800 billion of profit just to pay for the interest," he said."
Right, THEY can't, but cloud providers potentially can. And there are probably other uses for everything not GPU/TPU for the Google's of the world. They are out way less than IBM which cannot monetize the space or build data centers efficiently like AWS and Google.
The dotcom bust killed companies, not the Internet. AI will be no different. Most players won’t make it, but the tech will endure and expand.
Or endure and contract.
The key difference between AI and the initial growth of the web is that the more use cases to which people applied the web, the more people wanted of it. AI is the opposite - people love LLM-based chatbots. But it is being pushed into many other use cases where it just doesn't work as well. Or works well, but people don't want AI-generated deliverables. Or leaders are trying to push non-deterministic products into deterministic processes. Or tech folks are jumping through massive hoops to get the results they want because without doing so, it just doesn't work.
Basically, if a product manager kept pushing features the way AI is being pushed -- without PMF, without profit -- that PM would be fired.
This probably all sounds anti-AI, but it is not. I believe AI has a place in our industry. But it needs to be applied correctly, where it does well. Those use cases will not be universal, so I repeat my initial prediction. It will endure and contract.
The difference is that the Internet was actually useful technology, whereas AI is not (so far at least).
5 replies →
Can’t wait for all this cheap ddr5 memory and GPUs
I was looking at my Newegg orders recently. 7/18/2023 - 64GB (2 x 32GB) 288-Pin PC RAM DDR5 6000 (PC5 48000) --> $260. Now, $750+.
6 replies →
GPUs have a very high failure rate…
To be honest ai datacentres would be a rip and replace to get back to normal datacentre density, at least on the cooling and power systems.
Maybe useful for some kind of manufacturing or industrial process.
Cheap compute would be a boon for science research.
4 replies →
Why do you believe it will fail? Because some companies will not be profitable?
It wasn't an 'it' it was a 'some'. Some of these companies that are investing massively in data centers will fail.
Right now essentially none have 'failed' in the sense of 'bankrupt with no recovery' (Chapter 7). They haven't run out of runway yet, and the equity markets are still so eager, even a bad proposition that includes the word 'AI!' is likely to be able to cut some sort of deal for more funds.
But that won't last. Some companies will fail. Probably sufficient failures that the companies that are successful won't be able to meaningfully counteract the bursts of sudden supply of AI related gear.
That's all the comment you are replying to is implying.
Given the amounts being raised and spent, one imagines that the ROI will be appalling unless the pesky humans learn to live on cents a day, or the world economy grows by double digits every year for a few decades.
1 reply →
I mean that is how capitalism works, no?
>cash in on the floor space and HVAC systems that will be left in the wake of failed "AI" hype
I'd worry surveillance companies might.
There is a way of viewing the whole thing as a ruse to fast-track power generation through permitting.
you could stuff the racks full of server-rack batteries (lfp now, na-ion maybe in a decade) and monetize the space and the high capacity grid connect
most of the hvac would sit idle tho
the constant cost of people and power won't make it all that much cheaper than current prices to put a server into someone's else rack.
I guess he is looking directly at IBM's cash cow, the mainframe business.
But, I think he is correct, we will see. I still believe AI will not give the CEOs what they really want, no or very cheap labor.
I find it disturbing how long people wait to accept basic truths, as if they need permission to think or believe a particular outcome will occur.
It was quite obvious that AI was hype from the get-go. An expensive solution looking for a problem.
The cost of hardware. The impact on hardware and supply chains. The impact to electricity prices and the need to scale up grid and generation capacity. The overall cost to society and impact on the economy. And that's without considering the basic philosophical questions "what is cognition?" and "do we understand the preconditions for it?"
All I know is that the consumer and general voting population loose no matter the outcome. The oligarchs, banking, government and tech-lords will be protected. We will pay the price whether it succeeds or fails.
My personal experience of AI has been poor. Hallucinations, huge inconsistencies in results.
If your day job exists within an arbitrary non-productive linguistic domain, great tool. Image and video generation? Meh. Statistical and data-set analysis. Average.
Just like .com bust from companies going online, there is hype, but there is also real value.
Even slow non-tech legacy industry companies are deploying chatbots across every department - HR, operations, IT, customer support. All leadership are already planning to cut 50 - 90% of staff from most departments over next decade. It matters, because these initiatives are receiving internal funding which will precipitate out to AI companies to deploy this tech and to scale it.
The "legacy" industry companies are not immune from hype. Some of those AI initiatives will provide some value, but most of them seem like complete flops. Trying to deploy a solution without an idea of what the problem or product is yet.
1 reply →
> Even slow non-tech legacy industry companies are deploying chatbots across every department - HR, operations, IT, customer support
Yes, and customers fucking hate it. They want to talk to a person on the damn phone.
the calculation, while simple, does make rational sense if all things remain equal.
regardless, given their past business decisions, this statement can be true for them without thinking about the bottom line.
the entire premise is enterprise compute amped to the max. the sheer expensive nature of the hardware and intensive resource requirements truly challenges what modern data centers can do today.
Nobody really knows the future. What were originally consumer graphics expansion cards turned out useful in delivering more compute than traditional CPUs.
Now that compute is being used for transformers and machine learning, but we really don't know what it'll be used for in 10 years.
It might all be for naught, or maybe transformers will become more useful, or maybe something else.
'no way' is very absolute. Unlikely, perhaps.
> What were originally consumer graphics expansion cards turned out useful in delivering more compute than traditional CPUs.
Graphics cards were relatively inexpensive. When one got old, you tossed it out and move on to the new hotness.
Here when you have spent $1 trillion on AI graphics cards and a new hotness comes around that renders your current hardware obsolete, what do you do?
Either people are failing to do simple math here or are expecting, nay hoping, that trillions of $$$ in value can be extracted out of the current hardware, before the new hotness comes along.
This would be a bad bet even if the likes of OpenAI were actually making money today. It is an exceptionally bad bet when they are losing money on everything they sell, by a lot. And the state of competition is such that they cannot raise prices. Nobody has a real moat. AI has become a commodity. And competition is only getting stronger with each passing day.
You can likely still play the hottest games with the best graphics on an H200 in 5 years.
As long as the dollar remains the reserve currency of the world and US retains its hegemony, a lot of the finances will work itself out, the only threat to the US empire crumbling is by losing a major war or extreme civil unrest and that threat is astronomically low. The US is orders of magnitude stronger than the Roman Empire, I don't think people realize the scale or control.
Famous last words of empires. I doubt that if current situations continue that the US will in any sense be still thought of as the reserve currency.
Gradually, then suddenly. Best not to underestimate the extent to which the USA has lost trust in the rest of the world, and how actively people and organisations are working to derisk by disengaging. Of course that will neither be easy nor particularly fast, but I'm not certain it can be stopped at this point.
> The US is orders of magnitude stronger than the Roman Empire
This would be trivially true even if the US was currently in its death throes (which there is plenty of evidence that the US-as-empire might be, even if the US-as-polity is not), as the Roman Empire fell quite a while ago.
being pedantic doesn’t change the fact of what I said. If I said Obama is a better a leader than Julius Caesar, would you reply well actually, Julius Caesar is dead.
I'm pretty sure
> The US is orders of magnitude stronger than the Roman Empire [was]
was the intended reading.
Pride comes before a fall.
You should look at the depreciation of the value of the dollar this year.
NOTE: People pointed out that it's $800 billion to cover interest, not $8 billion, as I wrote below. My mistake. That adds 2 more zeroes to all figures, which makes it a lot more crazy. Original comment below...
$8 billion / US adult adult population of of 270 million comes out to about $3000 per adult per year. That's only to cover cost of interest, let alone other costs and profits.
That sounds crazy, but let's think about it...
- How much does an average American spend on a car and car-related expenses? If AI becomes as big as "cars", then this number is not as nuts.
- These firms will target the global market, not US only, so number of adults is 20x, and the average required spend per adult per year becomes $150.
- Let's say only about 1/3 of the world's adult population is poised to take advantage of paid tools enabled by AI. The total spend per targetable adult per year becomes closer to $500.
- The $8 billion in interest is on the total investment by all AI firms. All companies will not succeed. Let's say that the one that will succeed will spend 1/4 of that. So that's $2 billion dollar per year, and roughly $125 per adult per year.
- Triple that number to factor in other costs and profits and that company needs to get $500 in sales per targetable adult per year.
People spend more than that on each of these: smoking, booze, cars, TV. If AI can penetrate as deep as the above things did, it's not as crazy of an investment as it looks. It's one hell of a bet though.
Nit: its $800 billion in interest, your comment starts with $8 billion
right. My goof. That adds two more zeroes across all the math. More crazy, but I think in the realm of "maybe, if we squint hard." But my eyes are hurting from squinting that hard, so I agree that it's just crazy.
You're saying $8 billion to cover interest, another commenter said 80, but the actual article says ""$8 trillion of CapEx means you need roughly $800 billion of profit just to pay for the interest". Eight HUNDRED billion. Where does the eight come from, from 90% of these companies failing to make a return? If a few AI companies survive and thrive (which tbh, sure, why not?) then we're still gonna fall face down into concrete.
right. My goof. That adds two more zeroes across all the math. More crazy, but I think in the realm of "maybe, if we squint hard."
1 reply →
From the company that said the world market size for computers was about 50. And the company that gave us OS2.
I see a lot of people attacking the messenger but very few addressing the basic logic that you need 800B+ in profit just to pay the interest on some of these investments.
Pointing out IBM's mixed history would be valid if they were making some complex, intricate, hard to verify case for why AI won't be profitable. But the case being made seems like really simple math. A lot of the counterarguments to these economic problems have the form "this time it's different" - something you hear every bubble from .com to 2008.
At some point, I wonder if any of the big guys have considered becoming grid operators. The vision Google had for community fiber (Google Fiber, which mostly fizzled out due to regulatory hurdles) could be somewhat paralleled with the idea of operating a regional electrical grid.
thats like boeing telling us we shouldnt build rockets
I suppose it depends on your definition of "pay off".
It will pay off for the people investing in it, when the US government inevitably bails them out. There is a reason Zuckerberg, Huang, etc are so keen on attending White House dinners.
It certainly wont pay off for the American public.
Don’t worry. The same servers will be used for other computing purposes. And maybe that will be profitable. Maybe it will be beneficial to others. But This cycle of investment and loss is a version of distribution of wealth. Some benefit.
The banks and loaners always benefit.
That would be true for general purpose servers. But what they want is lots of special purpose AI chips. While is still possible to use that for something else, it's very different from having a generic server farm.
LAN party at the end of the universe?
I can't imagine everybody suddenly leaving AI like a broken toy and taking all special purpose AI chips offline. AI serves millions of people every day. It's here to stay even if it doesn't get any better than it is it already brings immense value to the users. It will keep being worth something.
1 reply →
Data centers house hardware, and it's a land grab for compute. What actually runs post-AI depends on its owners. A glut of processing might be spent reverse-engineering efficient heuristics versus the "magic box."
A bitter summer.
Ctrl-F this thread for terms like: cost, margin
Is transistor density cost still the limit?
Cost model, Pricing model
What about more recyclable chips made out of carbon?
What else would solve for e.g. energy efficiency, thermal inefficiency, depreciation, and ewaste costs?
How long can ai gpus stretch? Optmistic 10 years and we're still looking at 400b+ profit to cover interests. The factor in silicon is closer to tulips than rail or fiber in terms of depreciated assets.
Well, at least it tells us something about the sentiment on hn that a lame insight around self admitted "napkin math" and obvious conflict of interest garners 400 points.
>IBM CEO
might as well ask a magic 8 ball for more prescient tech takes
It honestly reminds me of the opinion pieces put out by encyclopedia companies about the many ways Wikipedia was inferior.
I read an article that pretended to objectively compare them. It noted that Wikipedia (at that time) had more articles, but not way more... A brief "sampling test" suggested EB was marginally more accurate than Wikipedia - marginally!
The article concluded that EB was superior. Which is what the author was paid to conclude, obviously. "This free tool is marginally better in some ways, and slightly worse in others, than this expensive tool - so fork over your cash!"
This is likely correct overall, but it can still pay off in specific cases. However those are not blind investments they are targeted with a planned business model
What's the legal status of all this AI code?! Will it be likely that someone whose code was lifted as part of "learning" can sue?!
Anyone can sue for anything.
I’m not sure we want a world where anyone who learns from existing code is committing a tort.
GenAI isn't a person
$8T may be too big of an estimate. Sure you can take OpenAI's $1.4T and multiply it by N but the other labs do not spend as much as OpenAI.
Its also strange that while solar systems built wait years for grid capacity, this much extra energy is being planned..
If it is so obvious that it won’t pay off, why is every company investing in it? What alpha do you have that they don’t?
That's a good question. During the .com boom everybody was investing in 'the internet' or at least in 'the web'. And lots of those companies went bust, quite a few spectacularly so. Since then everything that was promised and a lot more has been realized. Even so, a lot of those initial schemes were harebrained at best at the time and there is a fair chance that we will look in a similar way at the current AI offerings in 30 years time.
Short term is always disappointing, long term usually overperforms. Think back to the first person making a working transistor and what came of that.
> And lots of those companies went bust, quite a few spectacularly so.
pets.com "selling dogfood on the internet" is the major example of the web boom then bust. (1)
But today, I can get dog food, cat food, other pet supplies with my weekly "online order" grocery delivery. Or I can get them from the big river megaretailer. I have a weekly delivery of coffee beans from a niche online supplier, and it usually comes with flyers for products like a beer or wine subscription or artisanal high-meat cat or dog foods.
So the idea of "selling dogfood on the internet" is now pervasive not extinct, the inflated expectation that went bust was that this niche was a billion-dollar idea and not a commodity where brand, efficiencies of scale and execution matter more.
1) https://en.wikipedia.org/wiki/Pets.com#History
I still don’t get it. What at a personal level is making Sam Altman make a suboptimal choice for himself if it is so obvious it won’t work out for him?
7 replies →
“I think there is a world market for maybe five computers.”
Let’s hope IBM keeps their streak of bad predictions.
I think this old quote can come around to being accurate in a way, if you consider that from the user's perspective every cloud service is like one system. Aws, Azure, Google cloud...how many will there be when the dust settles? ;-)
Please, can we get more analysis from the geniuses at International Business Machines in 2025?
What kind of reporte does the CEO of IBM expect the general technology workforce to hold for them?
Okay, that is what IBM has to say. Are there any credible opinions we should know about?
There are so many CEOs, tech experts, financial analysts and famous investors who say we are in an AI bubble - even AI-invested companies say that about themselves. My latest favorite "We are in an AI bubble" comment comes from Linus Torvalds himself in the video with Linus from Linus Tech Tipps [0]
[0]: https://www.youtube.com/watch?v=mfv0V1SxbNA&t=2063s
To all those people that think IBM don't know anything. Calculate this number:
# companies 100+ years old / # companies ever existed in 100+ years
Then you will see why IBM is pretty special and probably knows what they are doing.
Pretty special? They were making guns and selling computation tools to the Nazis for a bunch of those years.
I think they trade now mostly on legacy maintenance contracts (e.g. for mainframes) for e.g. banks who are terrified of rocking their technology-stack-boat, and selling off-shore consultants (which is at SIGNIFICANT risk of disruption - why would you pay IBM squillions to do some contract IT work, when we have AI code agents? Probably why the CEO is out doing interviews saying you cant trust AI to be around forever)
I have not really seen anything from IBM that signals they are anything other than just milking their legacy - what have they done that is new or innovative in the past say 10 or 20 years?
I was a former IBMer 15 odd years ago and it was obvious then that it was a total dinosaur on a downward spiral, and a place where innovation happened somewhere else.
I don't think selling things to nazis was worse than selling things to israel today.
1 reply →
Mind you IBM makes +7B from keeping old school enterprise hooked up on 30 plus year old tech like z/OS and Cobol and their own super outdated stack. their AI division is frankly embarrassing. of course they would say that. IBM is one of the most conservative, anti-progress leaches in the entire tech industry. I am glad they are missing out big time on the AI gold rush. to me if anything this is a green signal.
How much of Nvidias price is based on 5 year replacement cycle? If that stops or slows with new demand could it also affect things? Not that 5 years does not seem very long horizon now.
LLMs at current utility do not justify this spending, but the offside chance that someone will hit AGI is likely worth the expectation.
There is something to be said about what the ROI is for normal (i.e. non AI/tech) companies using AI. AI can help automate things, robots have been replacing manufacturing jobs for decades but there is an ROI on that which I think is easier to see and count, less humans in the factory, etc. There seems to be a lot of exaggerated things being said these days with AI and the AI companies have only begun to raise rates, they won't go down.
The AI bubble will burst when normal companies start to not realize their revenue/profit goals and have to answer investor relations calls about that.
Says the guy missing out on it
The second buyer will make truckloads of money, remember the data center and fiber network liquidation of 2001+ - smart investors collected the overcapacity and after a couple of years the money printer worked. This time it will be the same, only the single purpose hardware (LLM specific GPUs) will probably end on a landfill.
The game is getting OpenAI to owe you as much money as you can. When they fail to pay back, you own OpenAI.
You are talking about the circular investments in the segment? Yes, but assume NVIDIA can get cheap access to IP and products of failing AI unicorns through contracts, this does not mean the LLM business can be operated profitably by them. Models are like fresh food, they start to rot by the training cut off date and lose value. The process of re-training a model will always be very expensive.
Are they LLM specific?
Deep down technically probably not, but they are optimized for this workload and business model. I doubt that once the AI bubble busts, another business model is viable. While datacenters can be downsized or partially shut down until demand picks up again, high end hardware is just losing money by the second.
Unless we get AGI.
Consumer will eat it all. AI is very good at engaging content, and getting better by the day: it won't be the AGI we wanted, but maybe the AGI we've earned
The same IBM that lost all races in the last 40 years? That IBM?
Hypothetically speaking, if the AI hype bubble pops (or just returns to normalcy), would it be profitable to retarget the compute towards some kind of crypto mining? If so, could we expect the cryptocurrency supply to soar and the price to tank in short succession?
“It doesn’t even have a keyboard!” energy.
"It is difficult to get a man to understand something when his salary depends on his not understanding it." - Upton Sinclair
What does Jim Cramer have to say?
I mean why not , they have to put something down to make their quantum show as better ROIs
"Company that has been left in the dust claims it was a good business decision"
the mining industry enters the chat
FB playbook. Act (spend) then say sorry.
One more data center and we will reach AGI, we promise, just give us more money.
[flagged]
I disagree on that and covered a lot of it in this blog (sorry for the plug!) https://martinalderson.com/posts/are-we-really-repeating-the...
100% of technical innovations have had the same pattern. The same thing happens every time because this is the only way the system can work: excess is required because there is some uncertainty, lots of companies are designing strategies to fill this gap, and if this gap didn't exist then there would be no investment (as happens in Europe).
Also, demand wasn't over-estimated in the 2000s. This is all ex-post reasoning you use data from 2002 to say...well, this ended up being wrong. Companies were perfectly aware that no-one was using this stuff...do you think that telecoms companies in all these countries just had no idea who was using their products? This is the kind of thing you see journalists write after the event to attribute some kind of rationality and meaning, it isn't that complicated.
There was uncertainty about how things would shake out, if companies ended up not participating then CEOs would lose their job and someone else would do it. Telecoms companies who missed out on the boom bought shares in other telecom's companies because there was no other way to stay ahead of the news and announce that they were doing things.
This financial cycle also worked in reverse twenty years later too: in some countries, telecoms companies were so scarred that they refused to participate in building out fibre networks so lost share and then ended up doing more irrational things. Again, there was uncertainty here: incumbents couldn't raise from shareholders who they bankrupted in fiber 15 years ago, they were 100% aware that demand was outstripping supply, and this created opportunities for competitors. Rationality and logic run up against the hard constraints of needing to maintain a dividend yield and the exec's share options packages.
Humans do not change, markets do not change, it is the same every time. What people are really interested in is the timing but no-one knows that either (again, that is why the massive cycle of irrationality happens)...but that won't change the outcome. There is no calculation you can make to know more, particularly as in the short-term companies are able to control their financial results. It will end the same way it ended every time before, who knows when but it always ends the same way...humans are still human.
1 reply →
Your blog article stopped at token generation... you need to continue to revenue per token. Then go even further... The revenue for AI company is a cost for the AI customer. Where is the AI customer going to get incremental profits from the cost of AI.
For short searches, the revenue per token is zero. The next step is $20 per month. For coding it's $100 per month. With the competition between Gemini, Grok, ChatGPT... it's not going higher. Maybe it goes lower since it's part of Google's playbook to give away things for free.
Great article, thank you for writing and sharing it!
Fiber seems way easier to get long-term value out of then GPUs, though. How many workloads today other than AI justify massive GPU deployments?
Would be funny if all the stagnant GPUs finally brought game streaming to the mainstream.
They discuss it in the podcast. Laid fiber is different because you can charge rent for it essentially forever. It seems some people swooped in when it crashed and now own a perpetual money machine.
The spending will be more than paid off since the taxpayer is the lender of last resort There's too many funny names in the investors / creditors a lot of mountains in germany and similar ya know
[flagged]
We detached this comment from https://news.ycombinator.com/item?id=46135548 and marked it off topic.
Please don't pollute the comments with noise. This adds nothing to the conversation.
"yeah, there's no way spending in those data centers will pay off. However, let me show you this little trinket which runs z/OS and which is exactly what you need for these kinds of workloads. You can subscribe to it for the low introductory price of..."
He will get assasinated or fired.
A lot of you guys in the AI industry are going to lose your jobs. LLM and prompt ‘engineering’ experts won’t be able to score an AI job paying as well as a barista.
No wonder why he is saying that, they lost AI game, no top researcher wants to work for IBM. Spent years developing Watson, it is dead. I believe this is a company that should not be existed.
Maybe it's the opposite. IBM spent years on the technology. Watson used neural networks, just not nearly as large. Perhaps they foresaw that it wouldn't scale or that it would plateau.
IBM CEO is steering a broken ship and it's not improved course, not someone who's words you should take seriously.
1. The missed the AI wave (hired me to teach watson law only to lay me off 5 wks later, one cause of the serious talent issues over there)
2. They bought most of their data center (companies), they have no idea about building and operating one, not at the scale the "competitors" are operating at
Everyone should read his argument carefully. Ponder them in silence and accept or reject them in based on the strength of the arguments.
His argument follows almost directly, and trivially, from his central premise: a 0% or 1% chance of reaching AGI.
Yeah, if you assume technology will stagnate over the next decade and AGI is essentially impossible, these investments will not be profitable. Sam Altman himself wouldn't dispute that. But it's a controversial premise, and one that there's no particular reason to think that the... CEO of IBM would have any insight into.
4 replies →
IBM can be a hot mess, and the CEO may not be wrong about this. These things are not mutually exclusive.
Sorry that happened to you, I have been there too,
When a company is hiring and laying off like that it’s a serious red flag, the one that did that to me is dead now
It was nearly 10 years ago and changed the course of my career for the better
make lemonade as they say!
Is his math wrong?
Are the numbers he's claiming accurate? They seem like big numbers pulled out of the air, certainly much large than the numbers we've actually seen committed to (not even deployed yet).
IBM CEO has sour grapes.
IBM's HPC products were enterprise oriented slop products banked on their reputation, and the ROI torched their credibility when compute costs started getting taken seriously. Watson and other products got smeared into kafkaesque arbitrary branding for other product suites, and they were nearly all painful garbage - mobile device management standing out as a particularly grotesque system to use. Now, IBM lacks any legitimate competitive edge in any of the bajillion markets they tried to target, no credibility in any of their former flagship domains, and nearly every one of their products is hot garbage that costs too much, often by orders of magnitude, compared to similar functionality you can get from things like open source or even free software offered and serviced by other companies. They blew a ton of money on HPC before there was any legitimate reason to do so. Watson on Jeopardy was probably the last legitimately impressive thing they did, and all of their tech and expertise has been outclassed since.