Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487]
What this covers
Patrick interviews Ben Thompson, author of Stratechery, in a tour of the major technology companies and the conditions shaping the AI buildout. It opens on China, where Thompson says US dependence on Chinese manufacturing inputs such as fabs and actuators is underappreciated and will not be fixed short of a conflict. It then turns to historical analogies: the 1870s railroads, whose financing collapsed while the track kept running, and container shipping, where rates rose from roughly $3,000-4,000 to $17,000-18,000 per container during COVID. Memory chips supply a third example of boom and bust.
The company-by-company middle section is the most concrete. Nvidia's circular financing, including a reported 25 percent backstop and stakes in neoclouds, works as a hidden price cut. Microsoft is described as following the IBM Middleware Playbook, and Amazon as its own first best customer. Google and Meta benefit from advertising's built-in verification loop. Thompson also uses Absolute vs Percentage Profit Logic, via Berkshire's move from See's Candies to BNSF, to suggest Google search may become the See's Candies of a much larger AI business. On costs, he separates casual chatbot use from long-reasoning inference, and he rejects the idea that open-source models are free. He argues this is why consumer AI needs advertising.
The more contested claims come last. Thompson doubts that results in verifiable domains like chess, Go, and math will carry over to unverifiable ones. He entertains intelligence becoming a commodity, and he sees AI improving AI as real, which raises the question of whether Meta and xAI can catch up. He also suggests that capturing the traces of human thought, perhaps through devices like Neuralink, could matter more than text. If the boom ends in a bubble, he expects the lasting residue to be power infrastructure rather than GPUs.
Thompson argues the AI buildout will be constrained by capital and duration mismatches more than by compute, and that its economics follow commodity-market logic in which risk is handed off rather than eliminated, so the winners will be those who can absorb that risk at the lowest cost of capital.
- Capital markets have to fund a decade-long buildout with short-term financing, as with the 1870s railroads, and funding gaps are the near-term risk.
- In commodity markets price is set by the marginal supplier, so companies with depreciated capacity keep running and push prices toward marginal cost.
- Companies that take on buildout risk, such as Nvidia through backstops and TSMC through conservatism, transfer that risk to others and pay for it in foregone profit or lower margins.
Overbuilding during scarcity causes gluts, yet busts still leave lasting productive assets behind.
- Although the railroad financing failed in the 1870s, the railroads kept operating, expanded the West, and made astronomical contributions to GDP that continue today, so financial bubbles can be devastating yet leave lasting productive assets.
“The funny thing is the railroads kept operating and they expanded the West. Their contributions to GDP were astronomical.”
- Memory has consolidated to a three-firm oligopoly that has become much more disciplined about avoiding past overbuild mistakes, without formally colluding, and this discipline ran head-on into a secular shift in memory demand that took time to recognize.
“But it got down to three. And the problem is three, it's not a monopoly, but it's kind of an oligopoly.”
- Because building capacity takes years, such as two years for a ship, if many firms invest simultaneously during scarcity the resulting supply glut drives prices down, and tech companies are measuring payback periods in a time of scarcity that may not hold in abundance.
“The problem is you're measuring your payback period in a time of scarcity. Is that payback period gonna hold in a time of abundance?”
AI inference breaks Google's zero-marginal-cost ad model, making its AI equity issuance look symbolic.
- Google's advertising is nearly a perfect business because it has zero transaction costs and zero marginal cost, since most advertisers never talk to Google staff and purchases are fully automated by computers that cost effectively nothing, but AI inference breaks that cost structure.
“those computers, from a business perspective, cost $0. AI, obviously that changes significantly.”
- Google's equity issuance to fund AI is symbolic because search may become the See's Candies of an AI business with far larger absolute profits, so shareholders accept a smaller percentage of a much larger pie.
“Well, you have a smaller percentage of an astronomically larger pie.”
Society must keep accelerating because getting stuck where we are would be the worst outcome.
- Thompson calls himself a reluctant accelerationist, believing society must push forward because it cannot go back, and that getting stuck where we are would be the worst outcome.
“What I call myself is a reluctant accelerationist. I think we need to push forward because we can't go back.”
Meta's content model gives it a cost advantage over YouTube and Netflix.
- Instagram generates large revenue while Meta pays nothing for content, in contrast to YouTube, which pays creators along the way and shares revenue after the fact, while Netflix pays upfront, giving Meta a cost advantage in content economics.
“Instagram is this unbelievable product that generates all this money for which Facebook pays $0 for content.”