Here’s the number Greg Ip wants every AI believer to stare at: 9% of GDP. By the math of Columbia finance professor Stijn Van Nieuwerburgh, that’s the yearly AI bill households and businesses would eventually have to run up to make today’s data-center spending pay off. It’s about what the country spends on food, and around double what it pays for all its energy. Ip’s advice: “You should be skeptical.”
Van Nieuwerburgh’s paper, presented at Brookings, works from the data centers builders actually plan, less some cancellations, then plugs in 50 cents of cash flow for every revenue dollar and a 10% unleveraged return on the capital. The answer: $3.5 trillion of AI revenue in 2032, or 8.8% of GDP. The soft spot is price. Those plans quietly assume today’s scarcity pricing for compute survives a world with roughly four times the capacity and plenty of competition. The companies, he said, are “basically saying we’re going to keep charging scarcity pricing in 2032.”
What fiber already taught us
That’s the history to worry about. London–New York fiber is the cautionary tale: as new strands went in and each one carried more traffic, bandwidth prices on that route fell 96% from 1997 to 2001, and the slide hastened bankruptcies across the long-haul fiber business. AI’s version is steeper: Epoch AI says the effective price of a given level of capability has fallen 47% a quarter since 2023. The bull case is the Jevons paradox — cheaper and better unlocks so much demand that total spending rises anyway — and Ip grants that, so far, that’s how it has played out. But spending on computers rode the same curve and plateaued once they were everywhere, and productivity doesn’t automatically become revenue: in one study of software teams, AI agents lifted lines of code 30%, yet finished projects barely moved, because people spent the gain reviewing the extra code.
None of which is slowing the spending. Amazon (AMZN) plans $220 billion of capital expenditures this year, mostly on data centers, and on Friday Amazon Web Services chief Matt Garman issued a memo calling the fight for AI dominance “the race our nation can’t afford to lose” and pledging more than $1 billion over five years to the towns hosting them. When the builders start paying the neighbors, you know how big the bet has gotten.
Our read
This is concentration risk, Investments/Risk (IN04). The Journal’s Streetwise noted that most of September’s risers had AI or data-center supply-chain exposure. In a cap-weighted index fund, that means a big slice of your diversified money is already riding on the 9%-of-GDP bet. Our own books hold Nvidia (NVDA) in 31, Microsoft (MSFT) in 29, Alphabet (GOOGL) in 28, Meta Platforms (META) in 20 and Amazon (AMZN) in 19 — all held at weight; nothing added.
The question isn’t whether AI is real; it is. It’s whether you could hold your AI-heavy positions through a fiber-style price collapse without selling at the bottom or pushing back a retirement date. If the honest answer is no, the weight’s too high, however good the story. Add up how much of your stock money now sits in a handful of AI names, across every fund and account — the umbrella’s easiest to find while the sky is still blue.