I want to make a comparison with a car rental business and say that it would be like valuing Hertz entirely on the basis of the number of cars they own, as opposed to how many they rent out, but cars have a much longer depreciation period, if there are no customers they’re not costing you more money, unlike your computer which you are using for training and sucking up massive amounts of energy, and those cars do maintain decent value even after they’re of little use to the car rental company, unlike the compute here.
Do you think we will recognize any walls? Or is there a point where the output might look different with respect to different paradigms / modalities we throw at it, but we won't be able to understand the quantitative differences as good/bad/scalable?
Everything is a temporary situation on long enough timeframes, especially if it’s exponentially growing. Moore’s law which dictates that compute depreciates quickly has been slowing down a lot in the last few years, coupled with the explosion in demand we’ve found ourselves in a prolonged shortage situation. The bubble will pop, but if you predict when correctly, you will be a rich man.
The key question is on direction of LLMs. Right now, LLMs are taking over human jobs. If the cost of silicon+power < cost of human being doing the same work, what rational reason is there to employ a human being?
If this applies to SWEs, lawyers, business analysts, many research scientists, .... this situation could persist for a long, long time. While capital costs less than the inputs of labor (nominal food, housing, etc.), there is no need for labor.
The key question is about continued progress in models, and of the tooling around them:
- Plateau: Old silicon obsoletes in due course
- Rise quickly: Old silicon maintains value for a long time
Yes, the plan seems to be anti human in the extreme. Why do you need the plebs if they can be entirely replaced by AI? But the question then becomes why does the AI (and before that their security detail in a post money world) need billionaires?
This likely is the tertiary reason as to why llms are so heavily kneecapped. Granted, at this point, projects do exist to remove those arbitrary restrictions, but the effort that goes into it suggests it is a real concern.
Short term, money physically exists and gets spent, so if you wave a magic want of oversimplification and transition all labour to AI instantly, all the money currently in bank accounts and wallets gets spend on the same businesses it was already getting spent on, a lot of which gets spent on stuff from other businesses who have in this scenario also replaced all their labour with AI.
Eventually, perhaps quickly, all this money ends up in the hands of shareholders and landlords. There's a lot of both in the economy; famously retirement funds, but smaller-scale shareholders and landlords also exist. I wouldn't want to guess what the distribution looks like, probably highly variable between countries not just social classes (the definitions of which themselves can vary between countries).
Long term, money exists as a convenient fiction to help us organise transactions of goods and services: while it may be physically possible to eat gold and banknotes, you're not getting any real nutrients out of it when you do. So in a world where goods and services come from machines, the options are too broad to forecast: humanity could be relegated to the same role and economic stature as other primates (both in and out of zoos), or we could get universal UBI denominated in machine labour credits which lets each of us live better lives than the most extravagant billionaires live today.
I don't know. It just seems odd because money was used as an abstraction of labor and if labor disappears it seems like money has no fundamental value. If you can't pay people to do something (because machines are doing all the labor). Then people have no money and money has no value to people. Industrialization resulted in transition to service-based economy but this new wave of machines are being said to replace service work.
I'm just trying to understand if suppose you have fully robotic farms and fully automated slaughterhouses and fully automated McDonald's, who is McDonald's selling anything to and how do these people supposedly buying fully-mechanized burgers have jobs? Something just doesn't add up about this in my head about how this equation balances.
UBI ultimately seems like socialism with extra steps. Mostly is comes across as billionaires desperately begging for an alternative to being nationalized.
Industrialization allowed people to shift human labor from agriculture to factories and such.
Seems like intellectual labor became more possible as people looked beyond subsistence but also more valuable since a greater population could drive demand for more than just subsistence related activities.
If both aren’t done by many humans, what’s left? Sports training and massage therapy? Sports training might not even be safe…
OTOH, my current lifestyle is already weird if I think about it. Developing software for a machine that I cannot make myself, whose raw materials I cannot obtain, using energy I cannot produce on my own — somehow entitles me to get a particular amount of goods and services from others including food, healthcare, entertainment, landscaping, and manufactured goods.
> If both aren’t done by many humans, what’s left? Sports training and massage therapy? Sports training might not even be safe…
Peacock tails.
As in, things where the effort itself is the point, to show off that you are capable of surviving when you consume resources so extravagantly on something other than (or even detrimental to) mere survival.
This includes stuff like hand-made art, being in a literal cult, extreme sports, and also refusing modern medical interventions/seeking out infections.
You may ask how someone can get paid for those things; I don't know, but we did manage to monetise talking to each other (ads on Facebook) and being locked in a house with some strangers while everyone's under surveillance cameras for a few weeks (TV show Big Brother).
> I'm just trying to understand if suppose you have fully robotic farms and fully automated slaughterhouses and fully automated McDonald's, who is McDonald's selling anything to and how do these people supposedly buying fully-mechanized burgers have jobs? Something just doesn't add up about this in my head about how this equation balances.
Well, people need to eat, so either the customers are on government support, or it comes from passive income, or from savings.
The people without those options, do it the old fashioned way: pick berries, throw rocks at animals, rub sticks for fire to cook them, or starve. Mostly starve, as the maximum number of humans who can survive as hunter-gatherers is 100-1000x smaller than the current global population.
> UBI ultimately seems like socialism with extra steps.
I agree. It's very much "from each according to their ability, oh wait we're all strictly worse than machines I guess that's from each nothing, to each according to their needs".
> Mostly is comes across as billionaires desperately begging for an alternative to being nationalized.
Perhaps, but that feels like claiming they're playing 5D chess, when Zuckerberg only plays Settlers of Catan with sycophants who let him win.
The overwhelming majority of the labor force remains service, manual labor, and other such stuff that LLMs will have no real effect on. So the economy will be fine, but I do agree with you from a different angle. The entire goal of LLMs seems self destructive. If they're successful then the endgame is completely removing the barriers to entry to producing software and other digital tech. But if we do reach that endgame then the value of tech is going to plummet because there will be absolutely no barriers to entry to compete, or even just individuals homebrewing up what they need on demand.
Like imagine there was something you could buy where you insert some lumber, give it some passable description of furniture, and it outputs it. And you paid $20/month for access to this. And this was all being bankrolled by the furniture industry? I mean, sure guys - it's much appreciated, but I don't think I've ever seen anybody so enthusiastic about digging their own grave. I think it's already obvious that the gazillion dollars of API calls isn't going to materialize - it seems the handful of companies that trialed that are already reversing course hard. And in the future where LLMs are successful, that'd be even more true.
Llms either reach the point where they can quickly design and build physical robots to take on that service industry or they stop exponential growth.
Both of those are devastating for their valuation. Stopping growth means open modes catch up in a year or so. Continuing means end of the current economy.
There’s a good chance physical humanoid robots will always be more expensive than humans, especially in this new hypothesised reality where there’s an enormous labour surplus.
China is advancing robotics at a crazy pace, and hitting typical Chinese prices. This [1] robot is available starting at $6k. And of course what matters isn't the up-front cost, but the maintenance. If their maintenance costs are lower than human wages+taxes, then robots win.
The biggest practical issue will be that if these robots ever did start replacing people on a large enough scale, then you'd have a lot of angry, desperate people with a lot of time on their hands. So that alone will probably work as the primary mediating factor.
Quite an interesting time to be alive because the future is so completely impossible to predict. The world just a decade from now could look entirely different, or it could just be self driving cars all over again.
It's not rational relative to the short-term incentives of a typical corporation or investment vehicle. PE, VC, fund managers aren't paid to give a fuck about the social contract. Literally not in their job description.
> Is wanting low unemployment in our society not rational?
Only conditionally on there being bad consequences for high unemployment.
I don't particularly trust politicians, but there's a whole host of hypothetical scenarios about futures where work is essentially optional. Unfortunately, they're all either in the sci-fi or religion sections of the book store:
Despite people occasionally investigating UBI, the efforts to research UBI seriously have the same problems that Marx had with literal Communism, in that there's an obvious difference between any partial transition as compared to a global transition, and we don't have a completely disconnected parallel world to be a petri dish for us to test the economic outcomes on.
> Capitalism allows individuals to take decisions in a free market.
Capitalism provides a set of incentives that shape how people make decisions. Anyone can be selfish, but selfishness in capitalist society has a particular shape. To ignore the external incentives when looking at human behavior is horribly naive and shortsighted, but is frequently done by capitalism-apologists who seek to disregard any criticism of their favorite incentive system.
Are current datacenter deployments structured in such a way that the memory can later be moved to newer GPU dies? Or is it all packaged together as on consumer graphics cards?
I assumed the latter and therefore that the memory is depreciating along with the GPU cores it's soldered onto PCBs with.
... or is it a different argument being made, perhaps that depreciation for GPUs has slowed because rising demand will keep them in service longer?
Removing RAM chips off old cards is uneconomical, until it isn't. With things going the way they are, if you've got a card with soldered on RAM that could be transplanted to a newer card, I think you'll start seeing that happening.
> Time on 4 year old H100 servers costs more now than when they were new (!!)
There are several confounding factors.
We’ve seen massive inflation since then. So some growth in cost was expected.
More importantly, the current Tech industry almost always starts by selling things at a loss. The increased cost could simply be the industry choosing to not subsidize that particular service anymore.
But also, I don’t think that’s a realistic comparison. Rented out GPUs are likely not a similar use profile as compute used for training LLMs. The latter is likely closer to the cryptocurrency GPUs that are running at full tilt 24/7.
> Rented out GPUs are likely not a similar use profile as compute used for training LLMs. The latter is likely closer to the cryptocurrency GPUs that are running at full tilt 24/7.
This is untrue.
H100's are used for training (well were, but are now outdated because B100/B200s are much faster).
Most of the reason people rent H100s is for smaller training runs.
If you are doing inference you usually buy managed capacity at Baseten or something, and that is often priced differently (although it comes down to an extra margin on longer term H100 prices basically).
Inference utilization is often actually higher than training now because so much effort has been spent on optimizing that stack.
I also feel that the GPU/NPU value does not lose money as fast anymore.
What I am wondering though is how long can you run such a system at basically full load without interruption before it starts to just physically degrade.
If I have a H100 and I let it run for 4 years at full throttle does it still have the same theoretical value as it had at the start or are the chips just burning out.
I think I remember that back when the cards used for crypto mining were sold en masse on ebay the advice was to stay away from them because they are more likely to fail?
Others say that moderate load means a lifespan of ~5 years
Not sure what that means but I would assume that a datacenter will start replacing a node once the error rate hits a certain threshold without really investigating why it failed, so the practical lifespan may be shorter than 5 years even if it would technically still be usable enough
Temperature is a big factor, as well as current density.
But there's also the # and magnitude of thermal cycles (which translate into mechanical stress, leading to metal-fatigue like effects on contact points etc), attack from chemicals in the air, cosmic radiation, ESD damage & more. Some may matter, some not.
That's why "new" > "used" in case of electronics. Especially since you don't know the (ab)use history of used parts.
> I also feel that the GPU/NPU value does not lose money as fast anymore.
That's because the rate of improvement in silicon manufacturing has been continually declining for a few decades, which has a compounding effect. Just compare the technological improvements in successive decades. 1976->1986->1996->2006->2016->2026.
That's why "in real terms" performance has only been very slowly improving if you compare apples to apples (and not e.g. apples to oranges by reducing precision, like nvidia tends to do, or by comparing chips with x W to an MCM with x*2 W and saying the latter is much faster). The "just halve the number of bits in each generation" strategy has also run out now, there's no more bits to halve.
Depreciating doesn't just mean it could depreciate in value relative to the performance of newer GPUs, but also that its lifespan is limited by reliability issues and failures.
My local inference rig now costs three times what I bought it for. If I'd gotten the max ram I could at the time I would have made $10k after selling the excess to my current spec.
How someone can look at an asset class thats appreciated an order of magnitude in the last two years and say it will depreciate in value when the tailwinds are even stronger now is beyond me.
Undervolting is not running at max utilization by definition almost.
…but the real question whether you want to undervolt your asset if you’re renting it out is why bother? You probably expect to replace it anyway after it’s spec lifetime, for sure want to replace it when a more efficient solution is available since datacenters are power and volume constrained and customers care about performance much more than hardware longevity (otherwise they’d buy instead of rent).
Why do you think it’s a waste? If you’re buying GPUs to rent them you’re almost buying a bond. If you’re leasing them, it’s even more obvious that you’re collecting the spread. The GPUs have a financial lifetime after which the business doesn’t pencil and they get sold for peanuts so you can put a better bond in your volume-power.
Consumer GPUs/CPUs tend to be operated at higher clock rates and voltages, because they need to win benchmarks. If you ever bothered to pay attention to how data centers operate their hardware you would notice that they have always gladly sacrificed 10% of performance if the total cost of ownership is reduced.
Since this entire sub-thread is in the context of used 3090s or consumer GPUs in general, you've failed to bring up anything relevant yet again.
Here is your strategy:
1. Increase power consumption by 50%: This costs you more energy to run the GPU, it also costs you more energy to cool the GPU, it ruins the GPU and since you hit power limits of your infrastructure earlier, you will have fewer GPUs in total.
2. Increase maximum performance by 10%: This is hardly noticeable, since the standard inference use case primarily involves taking advantage of the high memory bandwidth of a GPU. This means prompt processing will be 10% faster, or maybe your segmentation model that ingests video runs at 33 fps instead of 30 fps. You're optimizing for winning a benchmark with what will be used hardware in the future, that's asinine.
3. Throw away old GPUs or sell them for peanuts when they still sell for $1000 on the used market if they are in good condition and for $400 if they are damaged. I think the mistake here is obvious. If your GPUs are sold for peanuts, it's because you didn't take care of them.
Your business strategy is obsolete and based around the idea of pre COVID excess hardware capacity before there was massive AI demand where throwing out hardware made sense, because Moores' law was in full swing. Even Google is still offering their v2 TPUs from 2017 even though they've been long since obsoleted. Now in 2026, there isn't enough memory for consumers and people are snatching up all the hardware they can get their hands on. There were some big initial energy efficiency wins from implementing smaller data types that are no longer possible now that fp4 is the smallest possible floating point type that still makes sense and even if you go smaller, you can go down to two bits at best. The parameters are starting to become so small that 2:4 sparsity is becoming unattractive, because it adds one bit to the parameters.
2:4 sparsity for fp4 means 4+4 bits are compressed to 4+1 bits, but 2 bit parameters mean 2+2 bits are compressed down to 2+1 bits.
If you understand even a little bit about hardware, you notice that the tensor core hardware has already been optimized to the extremes and that there isn't much more you can pull out of it. Unlike CPUs there is hardly any control flow in matrix multiplication. The tensor cores implemented in Nvidia GPUs might be a little bit less efficient than an NPU/TPU based implementation (think Google), but there are no more obvious micro architectural improvements here. With CPUs the micro architecture has become so complex, that there may be ways to increase performance further, but for GPUs and NPUs, there is not much left other than process scaling. Further gains require better manufacturing processes from TSMC. TSMC introduced 3nm in 2022 and only started producing 2nm in 2025. That's a three year gap where barely anything happened and all the gains came from going from bf16 or half precision floating point, to fp8 and fp4.
Burning through hardware at high power consumption and mediocre performance increases is clearly not the way to go.
I have been hearing that memory suppliers are _intentionally_ not scaling up new factories like crazy because they assume the demand won't be there on the long term and they don't want to have spare unused capacity. Probably because Samsung and SK have a near duopoly on it as well...
At some point the market will be saturated with supply and prices will come down for older gen hardware. It can take years though, but it happened to fiber cable and fiber doesn't even depreciate like chips.
Will it continue to appreciate to infinity? Maintain its value forever? Or will something else happen?
The same argument you’ve made would work for tulip bulbs, dotcom prices, or whatever. Prices go up until they don’t. Exponentials don’t last forever and the intrinsics of technology assets depreciate: things wear out and are also replaced with better things.
Car rentals are a great comparison, but not for the reason you think. Cars depreciate value similarly to GPUs. The depriciation lifecycle timeframe is actually similar between hyperscaler GPUs and mainstream corporate car rental companies ike Hertz. They sell their cars after 2-3 years or 20-40k miles. There is a huge market for used cars. Hertz runs their used car sales out of their rental retail offices and a lot of overhead is shared. So take the difference in cost to buy new in bulk from the manufacturer from the retail sales price for a 2-3 year old car. As long as Hertz can make more money renting it out in that time, that's revenue positive.
Same with GPUs. There is also a huge market for used GPUs from 1-2 generations ago. The A100 is a six year old chip at this point and is still running strong, especially for inference. Like cars, chips can be refurbished and repaired. A hyperscaler or even mid level player here isn't going to hold onto chips for their entire usable lifespan.
> if there are no customers they’re not costing you more money, unlike your computer which you are using for training
So are you using the computers or not? I'd argue that if you're using them for training, then it's not wasted capacity. And if you're not using them, then you can turn them off, so you're not sucking up energy.
I don’t know but this dude at my son’s school has a 32GB RTX 5090 and it’s worth more than what he paid for; and he did the same trick with the RTX 4090 before that.
Until shortages are the rule, these assets are appreciating
"depreciating" is not being used in the right sense.
There is depreciation, which is taking the purchase price and dividing it across N number of years (typically 5). That's the D in EBITDA and is mostly used as a profitability calculation.
The depreciation of a GPU also gets mucked up in the current GPU financed market as well. DDTL loans. The people running the GPUs often don't even own the GPU, they lease it, so there is nothing for them to depreciate (D).
The analogy that a GPU is like a used car makes zero sense. There is no oil or tires to change on a GPU. They don't wear out in the same way that a rental car would. They are housed in climate controlled locations with clean power. They just don't fail the way that is portrayed in the press.
Useful life of a GPU is based on profitability. When does opex cost more than profitability?
Some companies, like mine, also have support contracts. Anything goes wrong with the GPU (or any part of the system), Dell comes and fixes it at no extra charge. We just migrate customers and workloads to hot spares while the parts are replaced.
As for compute going down in value... the 122TB of enterprise nvme and 2GB of ram in each server that I bought 2 years ago is now worth vastly more than I paid for it. I'm also renting my GPUs out for more money now due to supply being so tight and demand being so high.
Compute is about to come an appreciating asset in the near-term, and it some ways it already is.
The frontier labs are shifting from pricing grounded in the price of compute, to pricing grounded in the intelligence provided, or more specifically the economic value of that intelligence downstream.
The margins on that allow them to pay a hefty premium on compute and still come out ahead.
As they buy more compute at high prices, they're also pricing out competition from cheaper models. It's already become materially more difficult to get compute to run open weight models at competitive prices as a result of frontier labs in the last year.
Opus 4.7 has all the signs of a smaller model distilled from a newer pretraining run... except a smaller price.
Flash 3.5 raised in price pretty meaningfully over Flash 3
GPT 5.4 got a small price bump over gpt-5.3-Codex/gpt-5.2, then gpt-5.5 doubled pricing over gpt-5.4
Even open weights isn't immune: Kimi K2.6 was originally priced higher despite openly being 2.5 + more post-training, same with GLM 5.1 vs 5
-
All while rental prices are spiking month over month, and NVIDIA Inception discounted prices for buying are higher than undiscounted prices for buying 6 months ago...
I run a consumer AI product and the current reality of trying to get compute vs what it was 6-12 months ago is enough to justify it to anyone who has the money.
I think OpenClaw created a mania that was completely unfounded (Apple Silicon is worth dirt compared to literally anything from NVIDIA including consumer GPUs), but the prediction of compute becoming scarce was correct
Not necessarily. The GPU leases Spacex has made are month to month, so they are taking on all of the risk. If demand goes down, they're the ones stuck with the assets.
In the short term, compute becomes an appreciating asset.
In the medium term, everyone ramps up production. Huawei and other Chinese companies work really hard to develop in-house alternatives. At some point, the hype cycle will peak and less money will flow into datacentres (yes, this will happen. It always does. Even for technologies that change society. The bubble always bursts).
The question is not if this will happen. It will happen. It's just a question of when it happens and how big the magnitude of the cycle is.
I want to make a comparison with a car rental business and say that it would be like valuing Hertz entirely on the basis of the number of cars they own, as opposed to how many they rent out, but cars have a much longer depreciation period, if there are no customers they’re not costing you more money, unlike your computer which you are using for training and sucking up massive amounts of energy, and those cars do maintain decent value even after they’re of little use to the car rental company, unlike the compute here.