Comment by dgellow
1 day ago
One aspect Paul Kedrosky mentioned recently is the concept of „duration mismatch“. The price per token goes down over time (either because the AI vendor reduces due to competition pressure, or because customers are now incentivized to use older cheaper models). But datacenters are financed through debt, with the assumption their revenue increases over time. Quoting him: „[AI vendors are] paying for a fixed cost with a depreciating commodity“[0].
So you have on one end the token revenue trending down, on the other end the training cost going up for the next frontier models, and you need to pay back your 10y debt.
"So you have on one end the token revenue trending down, on the other end the training cost going up for the next frontier models, and you need to pay back your 10y debt."
Not necessarily, the bond holders could simply take a massive hair cut and lose shitloads of money. On the topic of bubbles and exuberance, Jeff Bezos made the salient point that there was a massive over-invested biotech boom in the 1990s and tons of sophisticated investors ended up losing lots of money. But humanity still kept the medical advancements made by the boom. Stocks going down didn't un-research drugs, and it won't un-research new GPUs or un-build datacenters.
> Stocks going down didn't un-research drugs
Drugs cost pennies to manufacture after they are researched and make their way through the approval pipeline. There are many generic drug manufacturers who can work off the existing formulas.
The more apt comparison is that LLMs won't be un-trained. Opus 4.8 now exists. Even if Anthropic somehow went bankrupt, that particular asset could, at the very least, be sold for proverbial pennies on the dollar to a "generic" inference provider.
Research does get lost over time. The whole point of the patent system is keeping that from happening; if the drug company goes bankrupt, even if they lose all their internal documentation in the process, hopefully the patents and other public paperwork provides enough information for an unrelated company -- either having acquired the patent rights, or after the patent period ends -- to reconstruct the processes with less investment then the original research.
If a bankrupt AI company maintains enough of a skeleton crew to consolidate and archive its intellectual property it could be sold off to another company, but there are also timelines where it all ends up digital dust in the wind.
1 reply →
Or locked away in litigation for decades… See what became of the Amiga
Datacentres aren't the same as infrastructure or research though. All the hardware in them has a finite, useful lifespan. In 10 years time it'll be totally useless
Hardware fails, and also scales out in terms of efficacy to run it as more power efficient, modern hardware turns up. It requires constant investment to keep it useful, and cost efficient
When AI pops, we'll temporarily have some extra compute capacity that will be horrendously uneconomical to run due to the high grid load and low consumer demand, before they get shutdown. There's simply no real use for them at this scale
Those data centers are specifically for AI workloads. Let’s say everything crashes and we now have all the data centers, what do you do with them? GPU are pretty specialized hardware, without AI a data center full of outdated graphics cards isn’t really too valuable.
It’s really not obvious the infrastructure we are building for AI stuff is something that will benefit humanity over time.
Without talking about the fact that bubbles are extremely destructive. Bezos is obviously someone who came out ok from the dotcom bubble but we are talking about something that destroys a lot of value globally. That has real, direct consequences, not just investors losing some money. The US economy is currently only growing because of the AI bet
AI data centers are being already used at max capacity, aren't they? I have a hard time imagining people would suddenly use AI less than they do as of today, let alone collectively drop it altogether. So the worst case scenario is that they'd need to be auctioned off way under what they'd be worth now, but still for someone to use them for AI.
Inference is much cheaper than training a new model, so running them just for inference is a completely different thing than having to price in the fact that at the moment all of these companies need to compromise between compute for inference and compute for training new models. If no new models were to be trained, and all the compute was inference only, that would change everything when it comes to the overall compute cost of AI.
Dotcom infra buildup is a bad comparison, in that it wasn't even close to being all utilized. The infra was completely overproportional to the day to day usage.
3 replies →
You sell the GPU's to remote gaming companies.
Replace servers with regular compute.
11 replies →
> Those data centers are specifically for AI workloads. Let’s say everything crashes and we now have all the data centers, what do you do with them?
You just run the models and sell the tokens. The demand will still be there even if there will be less money in chasing new frontier model
> GPU are pretty specialized hardware, without AI a data center full of outdated graphics cards isn’t really too valuable.
AI accelerators used in DC are not really "graphic cards" any more, you ain't running gaming on it
2 replies →
In order to not un-build the data centers, they at least have to make more than it costs to operate them, and also not have some attractive liquidation value (the land, maybe).
I could imagine something like “inference is done at home or in China, that’s the price to beat” and it’s not worth keeping all those GPUs cool out in Nevada.
But the parent comment was that one of the bigger costs in these data centers was the interest expense on the borrowed money. A restructuring removes or heavily reduces that amount.
The fiber laid during the dotcom bubble never paid back the investors or lenders, but it's still profitably connecting customers all these years later.
3 replies →
> Jeff Bezos made the salient point...
Big AI investor tells us that investing in AI is good. Oh, the surprise!
Does that invalidate this point? Yes. Because it makes no sense. The big money is not going to R&D but to build infrastructure that will be outdated in 5 years.
No, that's not the right read. He said bubbles and exuberance still produce lasting value for humanity, even when investors lose money.
Big money is going to build infrastructure which is fundamentally required for R&D. They aren't separate, they are the same thing. It sounds like you're complaining that Pfizer isn't investing in drug research, they are buying mass spectrometers and micron fidelity microscopes. Same thing!
Current AI datacenter/model development investment rate is roughly 1T/year. That's a lot. But the US economy is 33T/year. So the investment pays back (roughly) over ten years if, each year, the AI investments increase overall productivity by 0.6%, assuming the AI companies can capture half of the value of that productivity gain.
> „[AI vendors are] paying for a fixed cost with a depreciating commodity“
That's just a confusing way to say you don't think future models will be worth the development costs. Because if future models are significantly better, why would the price of tokens to access those models deprecate?
I'm surprised people think LLMs, a thing which mainly excels at advertising, spam and writing code is going to generate that much economic activity.
Companies whose main core competency is writing code were already making up a big chunk of the economy before AI. Also, less wealthy companies were constrained in their use of software by the inability to afford the salaries of talented programmers (and ripoff practices from software consulting companies who in theory could help). Lowering the cost of building software systems ought to unblock a good amount of economic activity as the technology diffuses.
12 replies →
But what if it kills current ad-tech as we know it (paying to show ads on random sites without any way to verify that the site is legit), and the flow of ad money for legitimate goods turns back to journalism, magazines and other publications?
That would be half a trillion[1] redirected to regular people just from Google Ads.
[1] snatched my number from here: https://pixis.ai/blog/2025-google-advertising-benchmarks-for...
1 reply →
Two of the things you’ve listed are some of the most profitable activities in our economy.
1 reply →
The $1T number seems more promises than reality, which is closer to the $300B to $500B level. Still a big number, but between a third and a half of the value used in the popular media.
The cost of power cost increase alone on industry gonna erase all gains from it.
You can't consider it in vacuum. AI takes limited resources. So far it winded up cost on near every consumer electronics that runs an OS, and it winded up cost of energy that is used by the entire industry and every single customer
It's not just the cost of datacenters, it's cost of infrastructure (that given current direction of US govt will just be paid from people's fucking taxes and bills..) and cost of other industries turning outright unprofitable "thanks" to demands of AI
These are similar numbers to the dotcom bubble. With GDP growth and the percentage of productivity AI contributes staying the same in this scenario this requires regular gains in revenue or growth. If things just stumble, like with most datacenters going unbuilt the bubble will pop.
A few things, I think you’re missing the point here
- most tasks do not require the latest frontier models, even if they are a magnitude more intelligent (we don’t actually know if that will be the case). Current Gemini flash is cheap, fast, and pretty capable with good guidance for most tasks
- now that companies pay API costs instead of a subscription they will be setting restrictions on token use to not have their budget explode (like Uber in this submission), that’s a strong incentive to NOT use expensive models, and limit their thinking budget
- there is competitive pressure from China and others who can offer very decent performances at a fraction of the token price
- the price of tokens for the frontier models is likely to go up, but the price to access older models is what depreciates! The overall price per token is going down now that we are in a new world where companies understand that token maxing is one of the stupidest concept ever created by humankind.
Relative to the current usage demand for tokens is effectively unlimited. If the price of tokens go down people will send more tokens to compensate. We are very very far away from a cost per token where people run out of things they want to send through an LLM.
If you have a good model router, you can route to older, cheaper models that run on older hardware, for simpler tasks. That helps labs extend the economic life of their hardware investments. They will likely fight it at first though as they see it as reducing ASP.
This is why I'm building role-model, a routing protocol and a router runtime: https://role-model.dev/
Running cheaper models on newer hardware is always going to beat running them on older hardware.
The other part of that is that while price per token may be going down, tokens per task is going up
For ~equivalent tasks/results, or because we’re expecting more or better from tasks?
The real measure should be cost per ~equivalent task result, not cost per token nor tokens per task.
For better performance of ~equivalent tasks. That's what all the harness tooling people are using does: (often) increasing output quality by significantly increasing token counts.
Right. Which means tokens are actually being priced well under cost once you factor in all this datacenter/GPU capex. Also worth noting the datacenters are not purely for training. They're for inference too.
I really wouldn’t be surprised if we saw some of these data centers scrapped in the next few years
do GPU chips really depreciate physically? There are no moving parts, I dont think memory chips or GPU chips deteriorate naturally.
I think its only accounting depreciation.
I have been using my laptop for a decade, what is stopping datacenters from using the purchased GPU chips for a decade?
Chips age and fail with age. You can check hot-carrier injection, bias-temperature instability and electromigration as they are the main aging mechanisms. All if these are a linear function of time but exponentieal of temperature. 90-100C these chips are running at are really tough, so they are likely to fail at couple of percent to 10% range in 2-3 years depending on the margins they have in the design.
The solder joints are notorious to fail at a high rate too.
If those don't go the caps and coils will eventually.
7 replies →
There are data centers that use and rent out 10 year old server GPUs.
They can't run larger modern models. They can't run smaller models as fast as newer servers. So their remaining market is applications where customers are okay with older, smaller models and slower performance.
They have to price the service lower than competitors due to the lower performance. The older GPUs are less efficient so it costs them more to keep them running. They're paid off, but they're taking up valuable power, space, and cooling in a data center.
Eventually there is a tipping point where it's better to replace that space and power budget with something new that has more demand.
The parts are sold off on the open market. There's an equilibrium demand for the parts from other data centers keeping older servers running and from hobby people who are okay with a jet engine sounding toaster of a GPU running in their home.
except for you know the enterprise customers who won't change their code and will pay to run old inefficent hardware just to keep from dealing with upgrades?
2 replies →
As long as the demand for GPUs keeps increasing, there are more data centers being built to house them.
When you have waitlists for many many months for Blackwell GPUs, keeping the old ones around as long as customers are willing to pay for them is great.
If I as a customer have a use case for a machine learning model I developed awhile ago, so an insect identification model, I had an ML researcher/eng develop it back in 2019, and it runs fine on a 2018-era T4 GPU (NVidia 2080 era), why mess with it?
3 replies →
In addition to the physical depreciations other comments mentioned I'd also mention that old chips will settle into a low price and then actually go up on a per unit basis if you're trying to buy a significant amount of them. With a limitation on fabrication facilities continuing to pump out older cards is an opportunity cost to the manufacturers that would prefer to be producing newer cards. If you were in a place where you suddenly wanted to buy 10,000 3080s, as an example, I'm not certain if the market could actually fulfill that demand and no one with the ability to increase the available supply to meet that demand actually wants to do so.
Chips do wear out and need to be replaced (entropy do be like that and durability is not a primary concern for chip design) so you'll need to refresh your stock and, even if you don't need cutting edge models, the price of all chips at scale will go up over time. It may feel unintuitive since, when the PS3 was released PS1s were extremely cheap - but if you're struggling to understand this effect from your experiences in the consumer market you're actually looking at the price factor that starts making antiques increase in value since at a certain point they become scarce goods. The market price for an NES is higher today than it was in 2003 because the price had already bottomed out from demand from the general consumer market but the demand remaining (speedrunners and the like) is now fixed or growing while the supply is inevitably shrinking.
They do degrade physically, but the bigger thing is they stop being competitive quickly. Each year or so we see doubling of GPU speeds for the same amount of power.
If you build a 100MW data center with GPU compute and three years laster a new data center opens with the same cost for GPUs and same electricity cost you do, but can do twice as much compute, you quickly lose business unless the market is just so constrained customers can't afford to be picky. But the moment there's slack in the market you'll see major migrations off of providers that have the same cost but half, or quarter of the same performance.
So when you see someone talking about GPUs fully deprecating in value in 1-3 years this is what they're talking about. Right now it's not a big deal because there's no slack in the market. But once there is, the bottom will drop out.
Gradually, and especially when hot. Modern chips are pretty close to the physical limits of how small they can be made, and that means atomic/chemical effects like electromigration are accounted for and determine the lifetime. Every extra 10 degrees Celsius of temperature doubles the speed of chemical reactions.
When they stray too close to the line ... you get Intel's 13/14th gen chips that wear out after 1-2 years instead of 10-20 years. Intel calls it "Vmin drift" because that doesn't sound scary, but the actual point is that various wear-out mechanisms push the chip outside of its design envelope - increasing the voltage or lowering the clock speed may get it to run for a while longer, but you're living on borrowed time as the various circuits just stop working right and you get unpredictable instruction mis-execution: https://fgiesen.wordpress.com/2025/05/21/oodle-2-9-14-and-in...
sounds like planned depreciation on Intel's part, they definitely do not design server grade chips for longevity since that would harm their own revenues
3 replies →
I used to work in datacenters, during spinning disk era we had technicians from vendors basically every couple of days to replace some broken part. When the massive switch to ssd happened instead of having them every couple of days it was 3 or 4 times per month.
Despite no moving parts things broke anyway and, even if it doesn't break, the vendor can make you change the technology just by playing with maintenance cost of the older one, limiting or removing spare parts from the market.
My understanding is that a lot of AI data centers are still heavily relying on spinning HDDs, which is why seagate, western digital are selling more HDDs than ever before.
2 replies →
Today's data center GPUs are essentially overclocked, and so at limit of how much the chip materials can physically handle, and therefore degrade over time. For example, GH200s operate at 1W/superchip but the actual safe power is somewhere around 650W which will allow them to function for a decade or more. But that leads to around 15% slowdown and that is unacceptable in today's competition. So current GPUs are destined to be depreciating assets.
In future, we might have fixed cost GPUs but not today.
I would presume the reason they are overclocked is because they are trying to make up for the shortage. In time, the shortage of computing components will be remedied, and tokens produced at lower power pulls will be cheaper.
i think its reasonable to give up 15% of speed for a decade more lifetime. This depreciation change alters economics of GPU
1 reply →
I assumed the issue was similar to crypto mining, where given finite amounts of space and power it makes sense to always be running the latest and most powerful GPUs instead of keeping older hardware running. There's definitely a secondary market for these GPUs as well.
Nothing is stopping them, it's just not worth it: Have a look at e.g. vast.ai's pricing (https://vast.ai/pricing).
The V100 (2017 -> 9 years old) can be rented from $0.02 to $0.37/h (right now I can find a V100 with a Xeon Gold 6140 and 48GB RAM for $0.165/h). Let's assume the guy you rent it to pins it at its 250W TDP and let's ignore the running costs of CPU/RAM/etc... Then you draw 1/4 kwh for that compute hour. The industrial electricity prices in the US vary between 7.5 and 25 ct per kwh (depending on state, time of day, etc...), so at 100% efficiency, assuming nothing ever breaks, and the CPU consumes 0W you earn about 14ct/h.
And remember: V100s hours are sometimes sold at 1/10th the price.
If I pick average conditions you need to start thinking of whether it is worth it to rent them out: Usually it isn't unless you have them anyways and just sell idle capacity.
It's barely worth it to run them in a pure "is it profitable" sense, if we also account for the opportunity cost of taking up a slot in your datacenter it seizes to be worth it really quickly.
Chips do deteriorate and fail naturally at datacenter scale or in timescales of decades, though not exactly like on financial reports. Leak current increases or electro-migrations occur at junctions or whatever those words mean.
And yeah, it does feel like GPUs will start losing values slower going forward with Moore's Law being dead for a while. It used to be that 3-5 years old GPUs were more useful as space heaters than GPUs, but that's much less of the case today.
> There are no moving parts, I dont think memory chips or GPU chips deteriorate naturally
I believe they do, but I too would love to know more details because there are several ways this can happen. Electromigration, package failures, VRAM failures, dielectric breakdown... Hopefully there will be studies soon similar to that old Google paper on HDD failures!
Currently it's a pretty big ask to look at the several hundred billion transistors and the interconnects between them to find what broke.
Though, those capabilities are maybe just a few years out, funnily it's taking AI to make it potentially doable.
GPU do depreciate indeed, but here the depreciating commodity is the token, not the hardware. You sell cheaper token with the same hardware
When everything is said and done it'll be datacenters in American competing with ones in China that have several times lower electricity prices. Token prices will drop to a level that will be unprofitable for American data centers and they will need to close.
Thats the main issue here.
1 reply →
the hardware itself is still useful, but random failures happen every so often, so if you're trying to run a fixed sized fleet then your fleet shrinks when you can't get spares any more
Your laptop doesn't have a 100% duty cycle. If you ran it like a data center it would indeed wear out much faster.
Transistors do wear out. Not going to elaborate as it is easy to ask GPT
When it was profitable to mine crypto with GPUs people used to sell these miner GPUs on the used market after about two years.
These were about half of the cost of an used GPU just used for gaming. By that pricr, I'd say a GPU kept busy has twice as high a chance of failure after two years of use.
Not great, not terrible.
Yes, even if the hardware is untouched. As technology advances, the power cost per compute cycle goes down. A gpu using old tech costs progressively more to operate compared to the newer models. So its value goes down over time = depreciation.
As for duty cycles, the chips are perfectly happy at 100% operation. Cooling and power componants fail, not the chips. But it costs manpower to repair such things and manpower is inconveniant these days. A gpu with any sort of fault just gets dumped.
Using a shittier model is just more work for the user, I’m not sure why anyone does it, unless they’re playing with it like a toy.
Local privacy respecting inference can be worth it. I use a local model to log everything I do all week to automate my timesheet. I also have it do a bunch of other data tasks. I won't say that larger SOTA models wouldn't do these tasks better than a local model but PII is a concern and my employer wouldn't approve of me just setting tokens on fire everyday to do what I could do myself.
> I use a local model to log everything I do all week to automate my timesheet.
Isn’t that just more work than logging it yourself?
1 reply →
I sometimes let Claude Opus create plans, DeepSeek v4 pro implements and writes tests. Claude reviews and corrects.
Saves like $2-3 per session. Same quality code.
“Same quality code.” [x] - Doubt
> more work for the user
Model routers allow this to happen automatically without any more work by the user.
> a shittier model
A ton of tasks don't require the most expensive frontier models, etc.
> I’m not sure why anyone does it
1. Faster solutions from the LLM - also reduces employee costs of having the employee waiting on the LLM
2. Avoiding things like the half-billion dollar per month bill for a single company’s LLM use recently reported in Axios
What you call a shittier model is what was considered frontier and fantastic one generation ago…