Even legends get it wrong sometimes
TLDR
Burry’s AI thesis mistakes early-stage experimentation for fake demand. Tokenmaxxing, benchmarking, and inefficient usage are not signs that the AI boom is hollow; they are exactly what adoption looks like before workflows mature, costs fall, and usage expands. Enterprises may optimize spend and route around expensive models where they can, but they are not going to compete with frontier labs or eliminate the need for massive compute. We’re still in the early innings of the AI boom: efficiency will improve, but usage should compound faster, keeping aggregate demand for compute on an upward trajectory.
Overview
Burry is bearish on AI and Nvidia. He’s provided his point of view in The Heretic’s Guide to AI’s Stars Part III: Tracepalooza & the Bezzle and here
Cassandra Unchained@michaeljburry
Great commentary yet again from BTIG’s Jonathan Krinsky
•Party Like It's 1999. In 1999, the best performing Nasdaq 100 stock was Qualcomm (QCOM, not rated), up 2600%. The best rolling 52-wk return for QCOM during the entire dot-com bubble was 2600%, so SNDK is beating that by
3:52 PM · May 6, 2026 · 234K Views
1 Reply · 221 Reposts · 1.29K Likes
Burry’s core point is that a meaningful portion of today’s AI demand is temporary, experimental, overfinanced, and being capitalized by the market as if it were durable, recurring demand. He thinks companies are “tokenmaxxing,” harvesting traces, benchmarking models, overbuying infrastructure, financing data centers through increasingly fragile channels, and mistaking a temporary training phase for a permanent demand curve.
Why he’s wrong
Burry is wrong because he is too focused on the “visible” excess of the current phase and not focused enough on the demand that has not even started yet.
Let’s walk through his thesis point by point.
Tokenmaxxing
Burry says:
“Tokenmaxxing is not merely heavy AI use, and it is certainly not sustainable AI use. It may be productive, but not 10x, not 15x. All of it is showing up as demand, and being mistaken for something it is not.”
He is right that tokenmaxxing is real. He is also right that some of it is silly. But this is where he starts to overreach.
I can guarantee that every developer, myself included, has tokenmaxxed (at least as much as their budget enables them to). That first time AI debugs a hard-to-find issue that was hiding deep in a utility class and the only way to catch it was to mock the third-party API you fetch data from and realize sometimes they send back null instead of zero is euphoric. That’s your light bulb moment. Like, woah, I’m never debugging anything again. Your next logical next step from there is to tokenmaxx. We’ve all done it. Unfortunately, for all of us lazy devs (the good ones are always lazy), LLMs aren’t at a point where we can send an ill-formed prompt and have the AI produce any meaningful output beyond another TODO application.
There are developers, engineers, and tech workers who are pushing AI tools to extremes. Burry is right, a lot of this usage is not efficient. Some of it is performative. But a lot of it is just the natural result of people playing with, exploring, developing intuition for one of the most important inventions: LLMs.
Burry’s claim that tokenmaxxing is wasteful is equivalent to saying that developers should not be playful and explore with LLMs. That’s the exact opposite of what we want our engineers and scientific researchers to do. The very essence of engineering is to explore and develop intuition.
OpenAI solved Erodos problem exactly because they tokenmaxxed - is Burry saying that is wasteful? Within math, Terence Tao sees a very prominent role for AI going forward. He says, “[The] most productive near-term adoptions of AI in mathematics will primarily come... from using medium-powered tools to accelerate and scale up more mundane and time-consuming, but still essential, research tasks," allowing [experts] to ‘readily and reliably [assess], [confirm], and [convert]’ the output into standard mathematical workflows.”” (source)
Tokenmaxxing is not the same thing as fake demand. It is not a pyramid scheme. It is not fraud. It is not 2008 housing. It is not people pretending empty houses are worth more because a rating agency blessed the paper. It is developers taking a powerful new tool and pushing it to its logical extreme.
That matters because every important technology starts this way. Early users do weird things. They overuse it. They abuse it. They waste money on it. They use it inefficiently. They run experiments that look insane from the outside. Then the tooling gets better, the workflows get better, the cost curve improves, and the strange behavior becomes normal infrastructure.
A short seller’s view on tokenmaxxing: people are intentionally overusing AI to climb to the top of company’s internal leaderboards which are fake and drive no value.
An engineer’s view on tokenmaxxing: I’m intentionally pushing this tool to its limit so I can understand where it performs best and where I still need to do manual work. I’m going to produce 100x - 1,000x more work in the next year because of this.
Tokenmaxxing is a de minimis part of aggregate demand
He is wildly overestimating how much of NVIDIA’s demand is actually coming from this behavior. No one can derive the numbers to understand how much of AI demand is due to tokenmaxxing, but 99% of the population (including large public companies) can’t afford to tokenmaxx to any meaningful degree.
We’ve spent too much time on tokenmaxxing. It’s irrelevant to the bigger picture. Please, let’s never discuss this again.
The Bezzle
Burry says:
“[About that bezzle, born of the word embezzlement]. The bezzle is not that AI is fake. Rather, the bezzle is the temporary demand - benchmarking, trace-harvesting, leaderboard hopping - that is being financed and tallied as if it were permanent demand, or an indication of even greater demand.”
Burry goes on to say that companies are using their own internal tools, which they’re also tokenmaxxing. For example, JPMorgan’s LLM Suite, Intuit’s custom-trained LLMs, etc. Burry’s point is that enterprises are using frontier models today to benchmark, harvest traces, and train their own internal AI wrappers, but over time they will move “up the pyramid” by replacing third-party models with proprietary tools, smaller models, cached workflows, and cheaper internal systems. As that happens, the frontier labs’ external token revenue gets compressed: adoption may keep rising, but the profit pool shrinks because companies learn how to compete with, substitute, and route around the expensive model providers.
I actually don’t even follow his point on this. He is basically arguing that companies are going to vertically integrate and compete with frontier labs. I don’t have a crystal ball, but that seems highly unlikely. Even if it was accurate, who cares? Sure, that’s bad for VC funds invested in Anthropic and OpenAI, but it’s not the end of the world. Aggregate demand remains the same.
Enterprises will absolutely optimize their AI spend, but that is not the same thing as competing with frontier labs. Is he really claiming that Intuit and JP Morgan are going to start R&D departments to develop the next generation of models? That seems outlandish. The LLMs we have today are going to pale in comparison to what we have in 5 - 10 years. “Altman has said training GPT-4 cost more than 1 billion.” (source). Zuckerberg offered a [1 billion training runs or if $1 billion pay packages are going to generate positive ROI, but it shows you what the table stakes are. Frontier labs are a big-boy game, and the Fortune 490 companies are in no position to compete.
Microsoft drops Claude
He also spends some time walking through how Microsoft cancelled Claude Code. His basic point goes back to the pyramid idea - where companies are going to vertically integrate and they want to reduce their total spend on frontier labs like Anthropic.
I agree with Burry on this point, but it’s moot. Aren’t all companies trying to reduce costs? And we still haven’t changed aggregate demand. We can shift what tools and providers we use, but aggregate demand is increasing.
Talent Gap
It’s hard to explain the talent gap between Silicon Valley, NYC, and everywhere else. You need to live it and breathe it. There are exceptions to every rule, but this is by and large very true. The idea that a Fortune 490 company not in NYC or Silicon Valley is going to compete on the AI frontier is simply misguided. Even JP Morgan, for all of its might, fortress balance sheet, and NYC headquarters, is not going to compete. They’re not even trying. Their own internally developed tool, LLM Suite, is model agnostic. “LLM Suite provides secure, scalable access to advanced large language models from multiple providers” (source). Institutions outside of Silicon Valley don’t have the risk appetite to compete. There’s not even a framework for them to spend on talent, capex, etc. Those types of pay packages and risk-appetite were regulated out of the large financial institutions in the Financial Crisis and Dodd-Frank. Of course, that risk still exists on Wall Street, but it’s with firms that are smaller and nowhere near the size of Google, Meta, etc. Maybe things could change with firms like Jane Street setting up their own data center. That’s getting off topic though.
Jevons Paradox
Burry says:
“Jevons Paradox does not rescue this capital expenditures cycle, and tokenmaxxing is the proof. Engineers are literally consuming as many tokens as they can, regardless of cost to the company.”
No. No. No. We already discussed this: <1% of engineers are tokenmaxxing. We are going to consume 1,000x - 10,000x more tokens as we start creating videos and agentic becomes a deeper part of engineering. And if that doesn’t cause token demand to explode even higher, wait until your girlfriend starts sending selfies (a shameless self plug).
The Cisco Analogy
Burry compares NVIDIA to Cisco, Sun, DRAM cycles, and other historical hardware busts.
This is not a crazy analogy. Hardware cycles are real. Customer concentration matters. Supply commitments matter. Overordering matters. Bullwhips happen. Margins can compress. Investors can overpay for dominant companies.
But Cisco is a terrible example. We’re nowhere near that.
The headlines will tell you this: internet demand increased → Cisco’s customers double ordered → Cisco ordered more equipment → Demand plummeted → Cisco crashed. Yeah, true. We need to a highlight a super important part of the story, “Chief Financial Officer Larry Carter said [Cisco] plans to "scrap and destroy" the majority of the inventory because most of it "can't be sold because it was custom-built for Cisco."“ (source). Now we have the full story.
GPUs are general purpose. That’s why Nvidia entered a licensing agreement with Groq, to ensure they have a competitive edge as specialized hardware rolls out. GPUs are never being thrown out. In fairness, there’s a bear case to make that GPUs can be written down because advancements will render them less valuable. Everyone building data centers is thinking about and trying to understand how to compete while also not building something that will be outdated in a few years. But, that’s a way different story than a huge write down occurring. Also, none of that matters for the pick-and-shovels play. If anything it’s better for them.
Where Is the Funding Coming From?
This is one of Burry’s best points, but it’s pointed at the wrong audience. The AI buildout is expensive. Very expensive. He explains how essentially retirees and life insurance companies are funding the capex boom via debt financing. That’s top notch detective work on Burry’s part, genuinely. That was news to me.
If life insurance companies, retirees, annuity holders, private credit funds, and offshore reinsurers are becoming large buyers of AI data-center risk, that deserves scrutiny. Should retirees be indirectly financing data centers through life insurance company balance sheets? I do not know.
That’s a question for retirees, life insurance companies, and regulators. But, if you cut that funding off, the cost of tokens will increase, not decrease. That’s econ 101. More expensive funding = more expensive products. Again, better for the pick-and-shovel plays.
TPUs, Smarter Tokens, Local Compute, and the Real NVIDIA Risk
There’s one major headwind to Nvidia and other pick-and-shovel plays that not enough people are talking about. Burry never talks about any of this and that demonstrates that he’s thinking about this from an accounting perspective rather than an engineering and economics perspective.
Nvidia’s massive tailwind is also its biggest long-term risk: everyone wants more compute, but everyone also wants cheaper compute.
That creates the obvious bear case.
What if TPUs, custom ASICs, local inference, smaller models, better routing, smarter tokens, distillation, caching, and llama.cpp-style local compute dramatically reduce the need for expensive centralized NVIDIA clusters? That is a real risk.
There is a scenario where custom chips make major progress. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. Meta wants its own silicon. OpenAI will almost certainly want more control over its hardware stack. Apple will push local inference. Startups will attack every layer of inference optimization.
If those efforts make step-function improvements, NVIDIA’s margins and market share could come under pressure.
But I think the probability that this destroys the centralized AI data-center demand curve over the next decade is low. In my opinion, there’s not a reasonable scenario where we say, “Hey, we don’t need large-scale data centers en masse. I’m happy with the performance of today’s models and I don’t want there to be step-up improvements anymore”. No way.
Takeaway
The mistake in Burry’s thesis is that he treats visible waste as the main event. It is not. The waste, benchmarking, leaderboard chasing, trace harvesting, and inefficient token usage are symptoms of early adoption, not evidence that the AI demand curve is fake.
The real question is whether AI usage expands faster than efficiency improves.
My answer is yes. Agents, video, enterprise workflows, scientific research, robotics, defense, and software engineering are still barely deployed relative to where they are going. The market may misprice individual companies. Nvidia may face margin pressure. Some data-center financing may end badly. But the idea that AI demand is mostly a bezzle misunderstands what is happening.
We are not at the end of AI demand. We are still in the messy, expensive, inefficient beginning and that’s a stockpicker’s dream.
An important note on Burry’s position for non-financials folks
Burry is long puts in 2027 - he’s betting the stock will decline by 2027. I have no view on short-term price movements. I’m simply stating that long-term (e.g. 24 months or longer) demand will increase, AI adoption will increase, use cases will grow, etc. and that will cause Nvidia’s long-term stock price to increase.
Can a temporary market dislocation push Nvidia’s price down? Absolutely. If your time horizons are short then that’s a totally different dynamic.
As a brief aside, Burry’s thesis is at the 30,000 foot view; so, I’m very surprised he has short-term bets. Maybe he has a more complex position on the books, enabling him to hold the position longer, but arguing tokenmaxxing and life insurance financing is going to cause the stock to decline in 2027 or sooner seems like a strong stance to take. Time will tell.
For informational purposes only — not investment advice.