Columbia Business School professor Stijn Van Nieuwerburgh prepared a paper for a Brookings conference estimating that US AI infrastructure construction through 2032 will require more than $10 trillion, or about 3.6% of GDP annually. This scale exceeds the relative cost of any previous infrastructure wave in American history. The author's warning goes beyond cost accounting: financing the buildout is shifting to less transparent schemes involving banks, private credit, and special purpose vehicles, and if return expectations are revised, risks could spread to the financial system as a whole, not just AI companies.

image

What happened

Columbia Business School professor Stijn Van Nieuwerburgh prepared a paper on AI buildout financing for a Brookings conference, and its findings were summarized on September 24, 2026, in a separate article by Reuters correspondent Howard Schneider. The central figures of the work are as follows: over seven years, the US will need to bring 183 GW of new data center capacity online, and the author measures payback through industry revenue, which must reach approximately $3.7 trillion per year by 2032. For reference, the current level is cited: combined revenue of OpenAI and Anthropic is about $100 billion, meaning revenue would need to grow by roughly 80% annually for the buildout to pay back. The second block concerns the financing structure: AI capital expenditures are no longer covered by cash reserves of Amazon, Meta, and Alphabet and are distributed through bank loans, private credit, and special purpose vehicles (SPVs), whose transparency the author directly compared to the state of the financial system before the 2007–2009 subprime mortgage crisis.

Context

To gauge the scale, the figure is useful to compare with past US infrastructure waves: 3.6% of GDP annually is more than electrification, railroads (2.2% of GDP at the peak of their construction), interstate highways, or the internet ever cost. Until recently, AI construction followed a different model: hyperscalers paid for data centers from their own cash flows and cash reserves, and the question of external funds barely arose. The paper records a regime shift: the capital investment cycle now depends less on product success than on access to cheap capital, so any revision of return expectations hits the entire chain of participants. In discussing the topic, the telecom bubble of the 1990–2000s is cited as a historical precedent, when excessive network construction ended in a collapse of expectations and industry consolidation.

Why this matters for the industry

For the industry, the main shift is in the source of money: construction growth has exceeded the ability of major players to finance it from their own cash flows, so risk has shifted to external creditors, and if return expectations are revised, dollars will leave not only AI companies but also the financial system. This changes the assessment of the sustainability of the entire capital investment cycle: the market is moving from the narrative that “hyperscalers will buy everything” to the question of “who will provide the money and at what interest rate,” and attention is shifting to the structure of SPV and private credit deals. Under the pressure of payback arithmetic, vendors will more aggressively monetize inference: selling priority access, packaging enterprise contracts, and tightening limits, so API pricing and predictability become a subject of negotiation. The strategy of “building on subsidized API prices and waiting for demand” loses its foundation, and compute buyers gain an argument about the risk of overbuilding capacity; for startups, it is practical not to bake cheap compute into unit economics as a constant and not to sign long-term contracts expecting today’s terms to hold through 2032.

Why this matters for users

A simple arithmetic from the paper is useful for readers: to pay back investments exceeding $10 trillion, AI industry revenue must grow from about $100 billion to roughly $3.7 trillion per year by 2032, i.e., about 80% annually; running loud financing claims through this proportion makes it easier to distinguish fundamental demand from a bubble like the telecom bubble of the 1990–2000s. For those building products on others’ models, a new class of risk emerges — the financial reliability of the supplier: delays in bringing capacity online, a change in asset ownership, or contract revisions could hit service availability and price even without technical problems. Reasonable steps to take now — inventorying dependencies (which models, from which providers, in which regions) and accounting for inference cost, limits, and provider SLA statuses in observability, and on a multi-year horizon — eval comparisons on alternative and smaller models, so provider switching becomes a planned procedure. If pressure on monetization intensifies, it is more likely that priority access will become more expensive, free tiers will be reduced, and API limits will be tightened, but this is a scenario, not a completed fact.

What is still unknown / limitations

The paper is a scenario assessment by a researcher, not a statement of completed credit events: until they occur, the effect works through expectations. The cited figures depend on assumptions about construction rates, energy capacity, and revenue growth, and the comparison with the subprime mortgage crisis is the author’s interpretation. The paper’s content was accounted for via the Reuters summary, so exact wording and methodology should be checked against the original work. Discussion on Hacker News (item 49872317) had not developed at the time of preparation: one point of rating and zero comments. Finally, in available APIs, prices, and limits on the day of publication, nothing changed — this is a signal of the planning horizon, not urgent changes.

Sources

Author

Look at AI, editorial team