Comment by ValentineC

I didn't suggest it was. I pointed out that some of the subscriptions offered by the Chinese labs probably are. Not the per token API prices.

And Chinese electricity prices are some of the lowest

Yeah, this argument is bullshit. You can head over to Openrouter and look at the token cost for deepseek-v4-flash and deepseek-v4-pro. They are very competitive on the open market

> 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.

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 GPU are pretty specialized hardware, without AI a data center full of outdated graphics cards isn’t really too valuable.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

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.

> 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.

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.

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.

Running cheaper models on newer hardware is always going to beat running them on older hardware.

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.

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.

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.

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...

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.

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 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!

GPU do depreciate indeed, but here the depreciating commodity is the token, not the hardware. You sell cheaper token with the same hardware

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.

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 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.

> 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…

The “you wouldn’t download a car” meme applies here

They'll ban them because (unless run locally or self-hosted) they are just data capture tools for the China.

Worst indeed, hardly anyone trades with them.

So, have you ever been to China and could hadely found anything familay?

- Oh, they must have been blocked from entering the Chinese market!

But none of that is true. You could see global brands everywhere here — Tesla, Unilever, KFC, Apple, and so on.

---

Or have you ever actually done cross-border trade? Or any international business collaboration? If you had, you’d definitely realize that what’s really stopping you is U.S. legislation. At least, that was the case with our former U.S. partner

[dead]

You don't have to remove the safeguards if you can prompt your way around them.

There's a subreddit for people wanting to sex-talk to various models. It just so happens that the same prompt they use to 'jailbreak' SOTA models for sex talks also works if you want to have model write malware, or tell you how to design a highly illegal device.

Search for "heretic"+Gemma/qwen/DeepSeek for examples where exactly this has been done.

From what I understand it’s mostly TSMC and the memory providers being out of capacity over the next few years.

So it’s not even about datacenters.

Here’s a Reuters article about TSMC: https://www.reuters.com/world/asia-pacific/broadcom-flags-su...

So this is actual committed contracts with all kinds of companies such as Apple, NVidia, AMD.

Also, the whole reason they can’t build data centers faster is precisely because of this.

> they can't build datacenters at anywhere near the rate they want to

That was because the supplies the datacentre needed were constrained - supply-constrained, not end-user demand constrained, so would be in agreement with the GP comment (and the article I read didn't imply anything about lying).

Google also needs fabs to build their TPUs.

Today's frontier models will be tomorrows low-end option. I think whatever model you are using today will be less expensive to use a year or two from now.

They aren't going down, but in the meantime they'll cover their ass by bribing their way into the S&P 500 and then use your 60 year old mother's 401k and teacher's pension to fund their risky capital expenditure.

I meant “roll your own” LLM for use not build new ones.

A key point here is open in terms of being able to download and use it, not open as knowing what data and instructions were fed into it when training.

A paranoid part of me thinks that these models are all inherently biased and instructed to be pro CCP, with specific gaps in their training data related to undesirable historic events and political ideas.

Here's a free startup idea: operate an open-weight model service, and offer "Verified AI Integrity," which signs the input tokens, the seed for the randomness in selecting outputs, and the model ID, proving that the result of the call to AI was completely "organic" and was not interfered with.

Your main audience would be snake oil salesmen trying to prove their AI products are unbiased and not under the thumb of any outside influence. This doesn't address the biases of the model itself, but that's not your business. Your business is selling tokens and security certificates. If you can get the right angel investor, you could maybe have your new standard required for some government applications.

Yes, you can. There are multiple inference providers out there. The problem is, it’s hard to beat the Chinese providers in cost. And you also have to compete with frontier model providers’ subsidized offerings.

There are plenty of US-based inference providers available, including AWS, that serve Chinese models at competitive prices (vs frontier US models). They also have lots of usage. Not necessarily for coding, but for other enterprise tasks.

Have you heard of openrouter? There's 1000 of these companies already. Do something else.

It's called AWS. Bedrock is right there. Price or data policy is never the issue. The models themselves are the problem -- most large US companies are not going to touch them.

Source: directly involved in these discussions. You can downvote as much as you'd like but you can't ignore the facts.

> Deepseek has some models in Bedrock.

Just looked into it, seems like at most they have just 3.2, not 4: https://aws.amazon.com/bedrock/pricing/

Looking around their catalogue more, most of their models seem quite outdated, aside from the OpenAI and Anthropic ones (but those get more expensive). I wouldn't willingly pick Bedrock and would instead throw money at OpenRouter, that has both a bunch of providers, as well as almost any model for you to try.

Makes me think of how my Claude.md files specifies to use the built in framework code-generators (rails). Those generators are deterministically right every time.

I wonder how often the Agent actually follows the guidance. I do see them follow it when I look. But it doesn't seem so every time.

A lot of the time if you're copying code from one place to another what you actually want to do is abstract it so you can reuse it in both places.

The LLM can easily do this type of stuff, just tell it and it'll happily do it. This is exactly what I mean when I tell people they need to work closer with the AI, tell it how to do things. Don't just tell it what to do and get frustrated when it does it differently than you would.

A good way to achieve this without writing huge prompts is tell it to plan the change first. Just give it some vague low-effort directions. It'll usually get most things right, you tell it what you want different and once you're happy you tell it to go ahead.

AI may get so commoditized for certain use cases that you will not even be able sell inference at a profit. AI might be bundled in with other services, just like cursor bundles in their own AI model for auto complete with their editor. I.e. cameras might have AI for image recognition bundled in etc.

I agree with all of this.

So my question remains the same: How are the players investing 100s of billions in buildout going to hope to make this back? Market capture looks bleak, inference looks like a race to the bottom. End users look like they could be beneficiaries. Where do the big boys go?

> Speculation here but I think openAI/antrhopic api inference is insanely profitable, it just needs more volume to amortize the training costs.

Well, they just rent their hardware, so I'm not so sure. But they'll both be public soon and we should get that breakout in their cost structures, somewhat.

The majority of Deepseek providers on OpenRouter for v4 pro are in the US. Especially interesting is that they are in the same ballpark for pricing.