Michael Burry (Trades, Portfolio), of Big Short fame, has made some waves for calling out AI companies, alleging that they are understating depreciation by extending the useful life of assets artificially boosts earnings. In case you’re unfamiliar, Burry was one of the first to spot the housing crisis that set off the Great Recession. I have a lot of respect for Mr. Burry, so I wouldn’t write his comments off. That said, there are some important caveats worth discussing.
Separately, I’ve been lightly researching AI chip obsolescence because it’s been a popular topic on Reddit and other places. Burry’s comments immediately piqued my attention. A common argument is that AI chips become obsolete after just a few years because newer chips offer stronger computing power and/or energy efficiency. This specific argument, I believe, is misplaced because older AI chips will still have many relevant uses running older and/or lighter AI models rather than the bleeding-edge models. These lighter models represent a hot potential growth area, making older chips not just viable but valuable. I can’t say for certain that Burry is worried about obsolescence, but he has mentioned useful lifespan.
Another concern: How long will the chips last before physically breaking down? I have not found a conclusive answer, but so far, my impression is that they typically last longer than 3 years. If the lifespan average falls under three years, Burry’s argument gains a lot of strength. If the chips can last significantly longer, however, that’d create headwinds for his thesis.
So far, Burry has been rather vague concerning the matter, so it’s hard (impossible?) to evaluate his argument. Details are forthcoming, but for now, we can consider possible angles and perhaps more importantly, critique common arguments and popular beliefs. With that in mind, I believe it’s important to take a deeper look at the potential lifecycle for AI chips. Until I see Burry’s argument, I can’t really refute it, but I believe the discussion below is important, and if nothing else, will provide readers with useful food for thought. It should also act as a primer for when Burry makes his November 25th release.
Michael Burry Stirs Up Chip Depreciation Controversy: Important Context To Consider
All of this is important for investors owing to concerns over the AI race and potential bubble. Stock markets and the economy have, in many ways, been propped up by AI investments. Valuations have been pushed high with AI companies leading the charge. Outside of AI investments, the American economy seems to be teetering on the verge of recession. If Burry is correct, it’ll inject a lot of skepticism into the markets, potentially causing corrections. It could also call into question the viability of current strategies and use cases around AI.
Twenty years ago, basic home computers often started to become quite sluggish (and perhaps short on storage) within just a few years. Upgrades were common as chips advanced at a rapid pace and the advances greatly improved the home computing experience. Often, obsolescence happened quicker for laptops rather than desktops, but still, useful lifecycles could be quite short. These days, it’s pretty common to see people hold onto laptops for substantially longer because the computer still meets their needs.
I mention this anecdotal experience simply to illustrate that hardware can remain quite useful past optimal lifespans. Bleeding-edge chips will lose their bleeding-edge much more quickly, but that doesn’t mean they’re useless even when pulled from the frontier. On a roughly annual basis, new gaming GPUs come out that offer better performance than the last model, and if you want the best graphics, you’ll need to upgrade pretty quickly. The same will prove true for data center chips: the next generation will typically perform substantially better.
For developers working on bleeding-edge frontier AI models, even seemingly minor boosts in performance, energy efficiency, and other metrics can make a big difference. Those developers working on the biggest and most cutting-edge AI models will probably move to upgrade as quickly as possible. The race to develop AI right now is white hot, with multiple massive competitors pouring vast resources into development. With so much poured in and the potential upside so high, quick upgrades are not just viable but logical.
Yet while a lot of public attention and media emphasis is on bleeding-edge frontier models, lighter (and older) models still remain relevant for a variety of tasks, and when it comes to leveraging AI for use in daily life, these lighter models can be just as effective, if not more so, than the bleeding-edge models. For example, many AI models can be run on a scaled-down model version locally. This is useful for smaller developers, and also businesses that, for security needs or otherwise, want to keep their data locally hosted and perhaps completely offline.
Many AI tools, including chatbots and various agents, simply don’t need all of the power of the frontier models. Cost concerns with frontier AI models, including how expensive they are to run, are prominent. The most advanced models consume the most energy and need the most chips, making them especially expensive to run. Some sources report that a six-month training run with GPT-5 costs $500 million. To be clear, training runs are especially intensive, but the point is: running the latest models is very expensive.
Michael Burry Stirs Up Chip Depreciation Controversy: Important Context To Consider
If you look at GPT-5’s pricing (shown above), on the surface, it actually looks substantially cheaper than GPT-4 (shown below). My initial thought was that OpenAI was simply trying to encourage people to use the latest model and thus set up their price structure to encourage that. However, a very insightful post by Zack Saadioui, which I recommend you check out, offers some crucial insights. The short of it is, when you use GPT-5, the request is sent to a central router, which from there, decides which model to use based on the difficulty of your request. As shown below, there are a variety of GPT-5 models of increasing sophistication, and the lightest models are much cheaper.
Michael Burry Stirs Up Chip Depreciation Controversy: Important Context To Consider
Further, GPT-5 uses so-called reasoning tokens. These tokens basically align with the internal thought process of the model, and the more thought processes you use to process your request, the more tokens you use. Further, these tokens are counted as output tokens, which are more expensive than input. Long’ish story short: the more complex your request, the more it’s going to cost. If you need the most fully powered model and thus compute power, you’re going to pay quite a bit more. If your request is relatively simple, it’ll be directed to the lighter-weight, cheaper GPT-5 models.
Going forward, I believe we’ll see this broken down at the hardware level with the most advanced models processing the most complex requests using the most reasoning tokens to run the most cutting-edge hardware. Yet ChatGPT (and other providers) should be able to use older hardware to run lighter nano models to process simpler requests. We talk about planned obsolescence with consumer goods, but for AI companies, we’re more likely to see planned downscaling with lighter and older models simply using older chips.
Another point worth mentioning, by the way, is that Chinese developers have made major breakthroughs on limited hardware, showing how lighter models and slower GPUs can be used to produce results. I suspect the current efforts of developers in China show (at least in part) the future of older GPUs being taken off frontier development.
My general thesis that chips have a substantially longer lifespan than 3 years could quickly be disproven if we find out that chips are failing at a high rate after just a few years. The chips might have remained useful if they were still functioning, but a burned-out chip is probably little more than recycling material at this point. This is a question I’ve been trying to answer definitively with light research over the past few weeks. So far, I’m finding mixed messages.
This study found that the last time a specific AI chip is used to train a frontier model (bleeding edge) is about 2.7 years. But I’m not as concerned about frontier models as outlined above. A commenter on this forum notes that general data center chips last 4 to 6 years (I’ve seen this mentioned elsewhere for standard data centers), but suspects that AI chips will last longer. That said, a Google employee has claimed that with high utilization, chips may last only 1 to 3 years, but potentially up to 5 with more moderate use. If this is true, Burry’s worries about short lifespans may be accurate even outside of useful lifespans.
CoreWeave, so far, may offer the most conclusive evidence. In a study with a 1024 GPU cluster in operation, the Mean Time to Failure was 3.66 days, which as I understand, means that when running a cluster, the first GPU will burn out after 3.66 days. This means that over the course of a year, about 100 GPUs will burn out. Thus at a constant rate, by year three, around a third of GPUs will have burned out. This number is pretty high, but it should also be noted that researchers are finding ways to extend chip life and mitigate physical damage. Further, it’s unlikely that any cluster will actually run 24/7 365 days a year.
There is one last point I want to quickly touch on. Newer chips may prove more energy efficient than previous chips, thus lowering operating costs. At some point, the calculus could shift to newer chips being cheaper overall due to energy savings. However, computing power rather than energy efficiency seems to be the chief concern, for now, since major players are racing to build the most powerful models. Energy, while limited, simply isn’t as crucial a concern at the moment as compute is. Further, demand for compute is so high that it would likely take many years for more energy-efficient chips to satisfy demand overall, which means older chips will likely remain in use.
We’ll find out more on November 25th when Burry releases more details. I suspect there will be a lot of buzz and likely industry leaders will push back rather quickly. Hopefully, they bring forward hard evidence, including the lifespan of their chips and how they justify the depreciation, financially speaking. Markets may suffer some turbulence, but even if there is an AI bubble, I’d be surprised if this were the pin to pop it. That said, if Burry’s argument is convincing, investors heavily exposed to AI should, at the very least, take time to evaluate their positions and risks.
