$10.3 Trillion and Counting: Brookings Paper Warns AI's Money Trail Is Going Dark

Every few months, a number comes along that recalibrates the debate about artificial intelligence’s economic footprint. The latest landed this week, and it is a big one: $10.3 trillion. That is the total US AI infrastructure investment projected from 2025 through 2032 in a new paper by Columbia Business School professor Stijn Van Nieuwerburgh, prepared for the fall 2026 conference of the Brookings Papers on Economic Activity (BPEA) — the elite economics series that has historically been the venue where looming macro risks get their first serious airing.

The headline figure is striking enough. But the paper’s more consequential argument is about where the money is coming from — and why that trail is getting harder to follow.

A Buildout Larger Than Railroads, Canals, or the Interstates

Expressed as a share of the economy, Van Nieuwerburgh estimates the AI buildout will absorb an average of 3.63% of US GDP per year through 2032. By his calculation, that exceeds every comparable infrastructure episode in American history: the canal boom, the railroad expansion of the late 1800s (which he puts at roughly 2.2% of GDP at its peak), rural electrification, the interstate highway system, and the telecommunications buildout that began in the mid-1990s.

The physical dimension of the forecast matches the financial one. The paper projects 183 gigawatts of new US data-center capacity over the next seven years, set against approximately 57 gigawatts installed today. To be precise, that is a projected addition rather than capacity already under construction — but even the direction of travel implies tripling the country’s data-center fleet, with all the power procurement, transformers, cooling systems, and specialized chips that entails.

From Cash Flows to Credit Chains

The first phase of this boom was largely self-funded. Hyperscalers like Amazon, Microsoft, Google, and Meta could cover capex out of enormous accumulated operating cash flows. That era, Van Nieuwerburgh argues, is ending: the investment now underway exceeds what even the major technology companies can fund from their own cash flows, shifting more of the burden to lenders, joint-venture partners, and other outside investors.

The financing chain now runs through AI labs, technology companies, banks, private-credit lenders, and real-estate firms — with special-purpose vehicles (SPVs), securitizations, lease commitments, and loan guarantees layered on top. Each link adds leverage, and each link spreads exposure to parties far removed from the data centers themselves.

Here the paper makes its most quotable comparison. In a briefing with reporters, Van Nieuwerburgh likened the opacity of these special-purpose vehicles to a defining feature of the subprime mortgage crisis: structures that make correlated exposures difficult to observe until a downturn forces them into the light. He is careful to state this does not mean financial distress is imminent — but it does mean that regulators and markets may not see concentration risk building until it has already built.

The $3.7 Trillion Revenue Hurdle

The return case for all this capital rests on a demanding assumption: the paper calculates that AI would need to generate roughly $3.7 trillion in annual revenue by 2032 to deliver the expected return on the invested capital. Set against the roughly $100 billion in combined annual revenue currently attributed to OpenAI and Anthropic — a comparison that measures the hurdle rather than a like-for-like tally, since the target is industry-wide — the implied growth rate is about 80% per year, sustained for seven consecutive years.

Van Nieuwerburgh is not predicting failure. His upside case is straightforward: strong growth in AI applications, heavy utilization of the new facilities, and continued model improvement could produce stable, even spectacular, cash flows. His downside case combines demand that falls short with rapid technology shifts (a chip or architecture breakthrough that strands today’s facilities), obstacles to completing projects, and high leverage. In that world, the financing commitments made in optimism would make the correction considerably more painful.

Why This Paper Matters Now

Three features distinguish this analysis from the routine “AI bubble” commentary.

First, provenance. BPEA is not a bank research note or a venture capitalist’s Twitter thread; it is the peer-reviewed conference series that famously hosted early warnings ahead of past financial crises. Its editors chose to feature three AI-focused papers this fall — including companion pieces on why AI is so politically contentious and on the vanishing advantage of specialization — signaling that mainstream economics now treats AI’s macro-financial risks as first-order.

Second, the paper’s policy ask is deliberately modest and actionable. Van Nieuwerburgh writes that “the most important policy contribution at this stage may therefore be to improve measurement and transparency while the capital structure of the industry is still evolving.” In other words: not to stop the buildout, but to make sure regulators can actually see who owes what to whom before the structure hardens.

Third, timing. The paper lands amid a wave of related warning signs: a Fortune investigation into how little the Federal Reserve knows about who is financing the $3 trillion global AI boom, a US House bill advancing to address data-center economic impacts, and local resistance to new facilities. The systemic-risk conversation is no longer fringe — it is moving through the institutions that actually set policy.

The Bottom Line

The AI investment race has entered a new phase. The sums are unprecedented in peacetime economic history, the funding has shifted from transparent corporate balance sheets to opaque multi-party structures, and the entire return case depends on revenue growth that has yet to materialize at anything close to the required scale. None of that guarantees a bad ending. But as the subprime comparison reminds us, the most dangerous financial structures are the ones nobody can fully map while the music is still playing.