David Cahn on the AI Bubble, Data Centers and Defense

TL;DR

Sequoia partner David Cahn believes AI is in a bubble, and he thinks the more useful question is who survives it. Companies that consume compute benefit from overbuilding, because cheaper compute lowers their COGS and raises gross margins. He argues that the physical work of building data centers, including steel, servers and power, is becoming a competitive moat.

Transcript

I do think we're in an AI bubble. You can see the fragility. Everybody can see the fragility. The thing that I think is more interesting is who's going to survive the bubble. Consumers of compute benefit from a bubble because if we overproduce compute, prices go down, your COGS goes down, and your gross margin goes up. The lesson that punches you i... Read More

Key Insights

• David Cahn believes AI is in a bubble with visible fragility. He thinks the more interesting question is which companies will survive it, since he is skeptical of any claim that everyone will win or everyone will lose.
• Consumers of compute benefit from an AI bubble. If the industry overproduces compute, prices fall, which lowers a company's cost of goods sold and raises its gross margin.
• The physicality of AI was underestimated. In summer 2024 most people discussed compute, models and data, while builders were flying electricians to Texas and trying to buy generator capacity that was sold out until 2030.
• The AI power trade was the best trade of 2025, according to Cahn. Wall Street investors who bet that power would be the main constraint made a lot of money, and the conversation moved from dollars to gigawatts.
• The $600 billion question is a simple equation: about $150 billion in Nvidia chips implies about $300 billion in data center investment, which requires about $600 billion of revenue at a 50% gross margin. In summer 2025 the figure was about $840 billion.
• The ability to build data centers is a moat, in Cahn's view. Complexity compounds when everyone builds at the same time and buys from the same vendors, so investors need to follow the AI supply chain all the way down.
• AI is now one of the biggest contributors to US GDP growth because GDP mostly captures physical activity. The construction boom, steel production and physical data center work are all being counted.
• Cahn's biggest miss was the scale of AI talent deals. Recent graduates seen as AI experts can get $50 million pay packages and well-known names can get a billion dollars, which he reads as a sign of desperation in the ecosystem.

Questions & Answers

Q: Is AI in a bubble according to Sequoia's David Cahn?

Yes. David Cahn says he believes AI is in a bubble and that everyone can see the fragility. He considers the more interesting question to be who will survive it. He is skeptical whenever anyone claims that everyone will win or everyone will lose, because markets always have variability, like a race where some runners are ahead and others fall behind.

Q: Why do consumers of compute benefit from an AI bubble?

Companies that consume compute benefit from a bubble because overproduction pushes prices down. Cheaper compute lowers their cost of goods sold, which raises their gross margin. Cahn makes this point while asking who survives the bubble. Application companies that buy compute could gain from overbuilding, while the companies that funded the excess capacity carry the risk.

Q: What is David Cahn's $600 billion question?

The $600 billion question, which Cahn wrote in 2024, is a simple equation. About $150 billion invested in Nvidia chips implies roughly $300 billion in data center investment. If the companies using that compute need a 50% gross margin to pay it back, about $600 billion of revenue has to be generated. Redone in summer 2025, the figure was about $840 billion.

Q: Has the $600 billion AI revenue question been answered?

Cahn says it has not. The question underneath it is whether the customer's customer is healthy. The customers are clearly healthy: companies are buying and building data centers, and those stocks have gone up. What remains unproven is whether there is a real end user generating enough revenue to justify all that compute spending.

Q: Why is building data centers a competitive moat?

Cahn argues that construction itself is a moat because building physical infrastructure is hard and its difficulty is consistently underestimated. Complexity compounds when everyone does the same thing at the same time and buys from the same vendors. He recommends following the supply chain past Meta and Google to the contractors they call, and then to the suppliers those contractors call.

Q: What prediction did David Cahn get right about AI in 2025?

Cahn predicted that people were underestimating the physical side of AI: steel, servers and power. He says this came true in two ways. First, the AI power trade was the best trade of 2025, and Sam Altman now talks about gigawatts instead of dollars. Second, mainstream media now report that AI's physical buildout is one of the largest contributors to US GDP growth.

Q: Are billion-dollar AI researcher pay packages justified?

Cahn calls them symbolic of desperation in an ecosystem that needs to prove its investments are worth it. The logic is that raising the probability of making a trillion dollars by 1% is worth a large sum. He agrees the math holds, but he warns that human brains reason badly at that scale and are psychologically biased to overestimate how much one person contributes.

Q: Why was David Cahn wrong about Meta in AI?

Cahn had predicted that Meta would do well in AI because it was vertically integrated, and he admits this was wrong over a 12-month horizon. He links Meta's very large pay packages to its underperformance. He still thinks he may be proven right over a longer period, because Zuckerberg's dramatic actions show how much a founder CEO can do when fully focused on a problem.

Summary & Key Takeaways

• David Cahn, a Sequoia partner, says AI is in a bubble with visible fragility, so the question that matters is who survives it. Consumers of compute gain from a bubble because overproduced compute lowers prices, cuts their COGS and lifts gross margins. He describes the hardest lesson in venture as the fact that no investor can make a company succeed.

• His 2024 thesis was that AI should be understood through atoms, not bits: steel, servers and power. Generators were sold out until 2030 and electricians were being flown to Texas. The AI power trade became the best trade of 2025, and Sam Altman now talks in gigawatts rather than dollars.

• Cahn's $600 billion question asks whether the customer's customer is healthy. About $150 billion of Nvidia chips implies roughly $300 billion of data center investment, which needs about $600 billion of revenue at a 50% gross margin. Redone in summer 2025, the figure is about $840 billion, and he says the question is still unanswered.