Buffett's Escalator: Hitting the Core of the AI Investment Frenzy

Investment guru Michael Burry, who predicted the subprime mortgage crisis and shorted NVIDIA, Jack Clark, co-founder of Anthropic, and tech observer Dwarkesh Patel engaged in a debate on the topic:

Is AI the greatest technological revolution in human history or a capital bubble on the verge of bursting?

Jack Clark put forward a very interesting but counter - intuitive observation: "In 2017, the mainstream consensus in the AI circle was the tabula rasa strategy, which involved letting AI start from scratch and learn through repeated trial - and - error in games like StarCraft and Dota 2, ultimately evolving into a general intelligence system."

Both DeepMind and OpenAI bet on this path and developed game - playing AIs that outperformed humans.

However, history has proven that this path is a dead end.

The real game - changer is another approach: Large - scale pre - training + Transformer architecture + Scaling laws.

The Transformer model proposed in the paper "Attention Is All You Need" made large - scale training more efficient, while the scaling laws revealed a simple yet powerful truth - the more data and computing power you invest, the smarter the model becomes, and this relationship can be precisely modeled.

Jack Clark said, "AI is at its worst state ever! The capabilities of Claude or GPT - 4 today represent the floor, not the ceiling. If you last interacted with AI a few months ago, your judgment of cutting - edge technology is severely off - mark."

This rapid iteration has led AI labs to return to agent development, but this time they are standing on the shoulders of pre - trained large models.

DeepMind's SIMA2 can explore 3D environments, and ClaudeCode can program autonomously. Their underlying capabilities are endowed by pre - trained models, just as every future StarCraft AI has read the original Chinese version of The Art of War.

However, just as the discussion was immersed in the excitement of technological breakthroughs, Michael Burry poured cold water on it. He told a story about Warren Buffett:

Warren Buffett owned a department store in the late 1960s. When the department store across the street installed an escalator, he had to follow suit. In the end, no one benefited from this expensive project. The profit margin didn't increase sustainably, the cost structure didn't improve, and the two stores remained in the same competitive position.

This escalator theory hits the nail on the head regarding the AI investment frenzy.

While all tech giants are driven by FOMO (fear of missing out) to buy GPUs and build data centers, no one can truly gain a sustainable competitive advantage.

"Because your competitors are doing the same thing. Eventually, both sides will only turn AI capabilities into a necessary cost for doing business, rather than a source of profit."

Burry's data is even more shocking: "NVIDIA has sold $400 billion worth of chips, but the end - user revenue from AI products is less than $100 billion." This 4:1 infrastructure - to - application revenue ratio is, in his view, a typical sign of a bubble.

Worse still, chips are updated annually. Data centers built at a huge cost today may become stranded assets in two or three years.

He specifically cited a statement by Microsoft CEO Satya Nadella: "I've withdrawn from some projects and slowed down the construction pace because I don't want to be stuck with four - to - five - year depreciation on this generation of chips." In Burry's eyes, this is solid evidence that even the most optimistic builders are starting to worry about capital traps.

The metric that Burry cares about the most is ROIC (return on invested capital), which is the gold standard for measuring a company's growth potential. In the past, software giants like Microsoft and Google had extremely high ROICs because software has almost zero marginal cost, and once developed, it can generate continuous cash flow.

However, AI has changed everything. These companies are turning into capital - intensive hardware enterprises, requiring continuous large - scale capital investment to buy GPUs, build data centers, and pay for electricity.

Nadella said in an interview that he hopes to maintain ROIC during the heavy capital expenditure cycle through AI, but Burry responded, "I don't see that happening. Even for Nadella, it just sounds like a hope."

An even more hidden problem is stock - based compensation (SBC). Burry calculated that about half of NVIDIA's reported profits are eaten up by SBC.

"When half of the employees are worth $25 million, does their productivity improvement really count? After deducting the real SBC costs, the profit margins of these AI star companies will shrink significantly."

Dwarkesh tried to refute: "Why is ROIC more important than absolute return? AI has expanded the potential market of tech companies from advertising ($400 billion) to the labor force (trillions of dollars)."

But Burry insisted: "If a company invests in low - return projects by borrowing or burning through its cash flow, it's just bloated. The price - to - earnings ratio will eventually drop to 8 times, the level of traditional enterprises with no growth prospects."

The discussion about whether AI really improves productivity has been lost in a fog of data.

An internal survey of developers at Anthropic shows that 60% of those using Claude claim to have increased their productivity by 50%. However, an independent study by METR found that after using AI tools in familiar codebases, developers actually took 20% longer to merge pull requests.

Jack Clark candidly admitted: "The data is contradictory and scarce. People's subjective feelings may be completely different from objective reality. A self - reported surge in productivity may actually mask a stagnation or even decline in real productivity. Anthropic is developing new monitoring tools and hopes to present research results in 2026 to clarify the truth."

Dwarkesh raised a deeper question: "If AI can really increase developers' efficiency by 10 times, why is the competition among the three major AI labs (OpenAI, Anthropic, and Google) more intense than ever? Either they can't build a moat by using their own products, or the productivity gains from AI are much smaller than they seem."

The most ironic thing is the silence in the job market.

Dwarkesh said, "If you showed me Gemini 3 or Claude 4.5 in 2017, I would have thought it could make half of white - collar workers unemployed. But so far, the impact of AI on the labor market is barely noticeable, like something you can only see under a'spreadsheet microscope'. The Industrial Revolution led to an extension of compulsory education to delay young people's entry into the labor market, but the AI revolution has not caused any similar social upheaval."

If there's one thing that shocked everyone in this discussion, it's Google's lag.

All eight authors of "Attention Is All You Need" are Google employees. Google has access to massive amounts of data from Search, Gmail, and Android, owns TPU chips, and even developed an internal large - language model early on.

However, this tech giant with all the technological accumulations has watched helplessly as OpenAI triggered the AI revolution with a single chatbot, ChatGPT. Burry couldn't believe it: "It's incredible that Google is chasing a startup in the AI field."

What's even weirder is that this revolution was initiated by a chatbot. ChatGPT's use cases were limited from the start - search, student cheating, and programming - yet it sparked a multi - trillion - dollar infrastructure race.

Burry made an analogy: "It's like someone built a prototype robot, and then every company in the world started investing in the robot's future."

The competitive landscape is also confusing. Dwarkesh observed that the leading position in the AI field is extremely unstable - Google was leading in 2017, OpenAI took the lead a few years ago, and now the big three take turns on the podium every few months. It seems that there are certain forces (talent poaching, information flow, reverse engineering) constantly erasing the snowballing advantage of any single lab.

Jack Clark believes that although all labs are using AI for development assistance, there is a "weakest - link effect": "Code generation speed has increased by 10 times, but code review speed has only doubled, so there's no qualitative change overall."

At the end of the discussion, the three guests surprisingly reached a consensus: "Energy is the ultimate limiting factor for AI development."

We also wrote about the same view in the article "Google Is Reaching for the Skies to Generate Electricity, but Is AI Really Short of Power?"

Burry's advice to policymakers is extremely radical: "Spend $1 trillion, bypass all protests and regulations, and install small nuclear reactors across the country. Build a brand - new national power grid and protect each facility with a nuclear energy defense force. This is not only for AI but also for the country's economic security. Only with cheap and abundant energy can the United States keep up with China in the competition and have a chance to pay off its national debt through economic growth."

Jack Clark strongly agrees: "AI will play an important role in the economy and fundamentally depends on the underlying infrastructure. Just like large - scale electrification and road construction in history, we need to do the same for energy. Large - scale AI data centers are ideal test customers for new energy technologies. I'm particularly looking forward to the integration of AI's energy demand and nuclear technology."

Behind all the debates about model parameters, training algorithms, and application scenarios, electricity is becoming an unavoidable physical constraint.

This round - table discussion didn't provide a definite answer but left two thought - provoking questions:

First, who will ultimately benefit from the value of AI?

If Burry's escalator theory holds true, no company in the AI supply chain can earn excessive profits, and the value will only flow to end - users. This is good for humanity as a whole but a nightmare for investors.

If Jack Clark is right and the rapid iteration of AI capabilities will eventually build a moat, then now is the best time to bet on future giants.

Second, should we trust timelines or data?

Dwarkesh pointed out that the revenue growth rate of AI labs (whether it will be $40 billion or $100 billion in 2026) is more telling than any benchmark test.

But Burry insists that until the application - layer revenue exceeds $500 billion or millions of jobs are replaced by AI, it's all just a matter of faith.

History will provide the answer, but until then, we're all placing our bets in this multi - trillion - dollar gamble.

Full - text translation

Host Patrick McKenzie: Michael Burry accurately predicted the subprime mortgage crisis when everyone else was buying. Now, he's skeptical as trillions of dollars flow into AI infrastructure. Jack Clark is the co - founder of Anthropic, one of the leading AI labs building the future. Dwarkesh Patel has interviewed everyone from Mark Zuckerberg to Tyler Cowen to figure out where this is all headed. We've brought them together in a Google Doc, hosted by Patrick McKenzie, to ask the question: Is AI the real deal, or are we witnessing a historic misallocation of capital in real - time?

Patrick McKenzie: You've been hired as a historian of the past few years. Briefly describe what has been built since "Attention Is All You Need" (the Transformer paper). What about the situation in 2025 would surprise an audience from 2017? Which well - informed predictions didn't come true? Tell this story as if you're speaking to someone in your field (research, policy, or the market).

Jack Clark: Back in 2017, most people bet that the path to truly general systems was through training agents from scratch on increasingly difficult task curricula and creating a system with general capabilities in this way. This was reflected in the research projects of all major labs, like DeepMind and OpenAI, where they tried to train super - human players in games like StarCraft, Dota 2, and AlphaGo. I think this was basically a "tabula rasa" bet - starting with a blank agent and "baking" it in some environment until it became smart.

Of course, as we now know, this didn't really lead to general intelligence - but it did lead to super - human agents within the distribution of tasks they were trained on.

At this point, people started trying another approach: large - scale training on datasets and building models that could predict and generate from these distributions. This approach ultimately proved extremely effective and was accelerated by two key things:

1. The Transformer framework from "Attention Is All You Need", which made this large - scale pre - training much more efficient;

2. The roughly parallel development of "scaling laws", or the basic insight that you can model the relationship between the capabilities of pre - trained models and the underlying resources (data, computing power) you invest.

3. By combining the insights of the Transformer and scaling laws, a few people correctly bet that you could obtain general systems by scaling up data and computing power on a large scale.

Now, interestingly, things have come full circle: people are starting to build agents again, but this time, they're infused with all the insights from pre - trained models. DeepMind's SIMA2 paper is a great example, where they created a general agent for exploring 3D environments, relying on an underlying pre - trained Gemini model. Another example is ClaudeCode, a coding agent whose underlying capabilities come from a large pre - trained model.

Patrick: Since large - language models (LLMs) are programmable and widely available, including relatively restricted but still powerful open - source software (OSS) versions compared to 2017, we've now reached a point where any further development of AI capabilities (or any other interesting things) doesn't need to be built on a cognitive base worse than what we currently have. I think this is one of the things that insiders understand best but policymakers and the outside world understand least: "What you see today is just the floor, not the ceiling."

Every future StarCraft AI has read the original Chinese version of The Art of War, unless its designers assess that it would perform worse in defending against a Zerg Rush.

Jack: Yes, one thing we often tell policymakers at Anthropic is "This is the worst it will ever be!", and it's hard to convey to them how important this ultimately is. Another counter - intuitive thing is how fast the capabilities are improving - a current example is how many people are using Opus 4.5 in ClaudeCode and saying various versions of "Wow, this thing is so much better than before." If you last interacted with an LLM in November, your judgment of cutting - edge technology is severely off - mark now.

Michael Burry: From my perspective, in 2017, AI wasn't LLMs. AI was artificial general intelligence (AGI). I don't think people considered LLMs as AI at that time. I grew up reading science fiction, which predicted a lot of things, but none of them depicted "AI" as some search - intensive chatbot.

Regarding "Attention Is All You Need" and the Transformer model it introduced, these were all done by Google engineers using Tensor. As early as the mid - 2010s, AI wasn't a new concept. Neural networks and machine - learning startups were common, and AI was frequently mentioned in conferences. Google already had large - language models, but they were for internal use.

One of the biggest surprises for me is that given Google's dominance in search and Android, as well as its advantages in chips and software, it didn't take the lead all the way.

Another surprise is that I thought application - specific integrated circuits (ASICs) and small - language models (SLMs) would