Key Takeaways:
The AI buildout is running out of money before the revenue arrives, and the funding chain has already moved from free cash flow to debt to equity issuance — the same sequence that ended the 19th century railroad boom.
The capital timing mismatch is the most urgent problem in AI, ahead of compute and power, according to Ben Thompson, founder of tech analysis firm Stratechery. Speaking on the "Invest Like the Best" podcast on Aug. 19, Thompson warned that AI infrastructure spending has raced down the capital curve: tech companies burned through free cash flow, then debt markets in roughly a year, and now Google is issuing equity while Nvidia assembles a $500 billion financing program to tap pension and insurance float.
"We start with free cash flow, and then tech companies burned through the debt market at an astonishing rate, in about a year," Thompson said. "Now Google is issuing equity, and Nvidia is putting together a $500 billion program to connect to pension and insurance float. After that, where does the money come from?"
The stakes are enormous. Industry capital expenditure is running at roughly $800 billion this year and is projected to reach $1.3 trillion next year, with new compute capacity absorbed almost as soon as it comes online. Thompson drew a direct parallel to the railroad boom, which consumed a comparable share of U.S. GDP: construction was a decade-long project financed with short-term debt, and the world simply ran out of money. The bubble burst, but the railroads stayed — and kept contributing to GDP for a century. The likely lasting legacy of this cycle, he argued, is power infrastructure.
"If all of this collapses and we find ourselves with a lot of excess electricity, that's actually a pretty good world," Thompson said. "We've always lived in energy scarcity, and energy is the foundation of everything."
Nvidia's margins look intact, but the risk has moved off the balance sheet
Nvidia has kept its reported margins high, but Thompson said the company is effectively discounting through off-balance-sheet risk. Its financing deals — providing a 25 percent backstop in exchange for new cloud companies committing to buy compute through 2030 — shift risk onto Nvidia's books even as the price discount stays hidden from gross margin.
"They're maintaining their apparent margins through various deals, like providing a 25 percent backstop in exchange for new cloud companies committing to buy compute until 2030," Thompson said. "Why can these companies get lower financing costs? Because Nvidia is taking the risk. And taking risk has a cost."
The discount is real in discounted cash flow terms even if it never appears on the income statement, he said. Nvidia's quarterly free cash flow has grown 18-fold over three years to $48.5 billion, and it held $30.2 billion in marketable equity securities last quarter, up from $12.9 billion a year earlier. The company has also committed up to $105 billion to back an OpenAI data center in Ohio and invested $30 billion in OpenAI in February.
Thompson identified Google and Amazon, not AMD, as Nvidia's true long-term rivals. Both are building in-house chips and selling them externally — Google has sold about 20 percent of its TPU capacity to Anthropic, and Amazon's Trainium is expected to follow. Their edge is a lower cost of capital.
"It's a capital-cost war," Thompson said. "The hyperscalers' core advantage over new cloud companies is cheaper capital."
Google is becoming Berkshire Hathaway
Thompson offered a striking analogy: Google is following the Berkshire Hathaway playbook. Berkshire's See's Candies generates outsized margins but has a limited ceiling, so Warren Buffett poured those profits into BNSF Railway — lower margin, but absolute profit dwarfing See's entire history. Google Search, he argued, is the modern See's Candies: near-zero marginal cost and no reinvestment need, funding an AI bet that requires enormous cash.
"If AI's total addressable market is all white-collar work and eventually everything robots can touch, then even at lower margins the absolute profit could dwarf search," Thompson said. "We may look back and call Google Search the See's Candies."
Under that framing, Google's equity issuance is not a sign of weakness but a rational trade — diluting a slice of a much larger pie. Berkshire's participation in the Google raise carries deep symbolism, Thompson said: "It's almost literally the railroad's money flowing to Google."
Compute scarcity may rescue Intel
Thompson argued that TSMC's conservative capacity expansion — rational in surplus periods but costly now — has created a severe compute shortage and transferred the risk to big tech, which is forfeiting revenue it could have earned. That pain is creating the economic incentive for large companies to prop up Intel and Samsung's logic business despite the friction of working with them.
"Nobody wants to go to Intel normally. TSMC is too easy to use," Thompson said. "But when the shortage is severe enough, and the revenue you're giving up is large enough, you're willing to do it." He predicted a major partnership announcement with Intel in the near term, calling it a milestone event.
Who wins, who loses
Thompson's competitive read across the majors: Amazon has the deepest moat, using its "first-best-customer" model to incubate AWS, Graviton, and Trainium before selling them externally. Apple's advantage is owning the user entry point, letting suppliers come to it, though he flagged the risk that it treats the phone as the permanent center of AI. Microsoft is running the IBM 1990s playbook — a middleware and implementation layer — but faces an existential threat from coding agents like Codex and Claude Code. Meta's advertising business is quietly one of the biggest AI monetization beneficiaries, he said.
For investors, the divergence is stark. Nvidia shares have ridden 12 straight quarters of revenue growth above 55 percent, but the off-balance-sheet financing and the rise of Google and Amazon silicon argue for a lower sustainable margin. Google's equity raise, with Berkshire aboard, frames the search franchise as a cash engine for a far larger AI prize. And if Thompson's Intel prediction lands, it would hand the foundry a rare catalyst after years of ceding share to TSMC.
This article is for informational purposes only and does not constitute investment advice.
