Patience runs through every scenario the IMF sketched for AI this week, including the generous one. I read the fund's Bangkok message as continued near-term demand for AI infrastructure with a productivity payoff that trails the spending, which makes the build-out a long-term foundation.

The Wall Street Journal reported the fund's three cases on Saturday. The optimistic one pairs productivity gains with contained inflation. In the middle case the build-out runs into chip supply limits, so inflation and interest rates rise before the productivity gains arrive. The bust takes as much as 1.3 percentage points off global growth in 2028 (Douglas and Landers, 2026). The fund did not say which it considers most likely.

Managing Director Kristalina Georgieva framed the timing problem as one of diffusion. People overestimate what new technology delivers in the short term and underestimate the long-term gains, she said, and some upset looks inevitable even if the gains arrive. IMF research suggests well-run adoption could add half a percentage point to world growth every year (Wee, 2026), and AI investment as a share of gross domestic product is on course to exceed what railroads and electricity grids took.

Spending at that scale needs a funder who can wait several years for a return.

The BIS and the IMF both see the weak point in financing

The Bank for International Settlements, or BIS, compared the boom to the railway mania of the 1840s and the dot-com years in its June annual report. Fortune noted that the BIS stopped short of calling AI a bubble. The concern is competition: firms racing for share can over-commit to projects with uncertain returns, and a disappointment in payoffs could trigger a "sudden pullback in financing," in the report's words as Fortune quoted them.

The Qatar Tribune's account of the report adds two details. AI funding runs through hedge funds and private credit vehicles, and the five largest cloud and AI infrastructure companies spent beyond their earnings and free cash flow in 2025 and 2026, pushing some into debt markets.

The IMF's Global Financial Stability Report, released October 13, names an AI correction as one of three salient risks, next to higher government borrowing costs and a major cyberattack. Computing's summary highlights circular financing, in which established AI companies fund less secure businesses. Hedgeweek's summary highlights the fund's concern about leverage built through derivatives and margin loans. Athanasios Vamvakidis, a deputy director in the fund's Monetary and Capital Markets Department, told Reuters that a correction would reach the real economy through lost wealth and weaker demand.

Stijn Van Nieuwerburgh's working paper for the National Bureau of Economic Research puts a one gigawatt AI campus at about $41 billion and a 188 gigawatt U.S. buildout through 2032 at nearly $9 trillion, or 3.2 percent of gross domestic product each year (Van Nieuwerburgh, 2026). The Journal quoted a larger version of the estimate, over $10 trillion and 3.6 percent a year, so treat the figure as a range. His paper finds funding shifting toward leases and special purpose vehicles, which raise leverage at the asset level and hide contingent obligations.

A slow payoff sorts vendors by how they funded the build

Goldman Sachs' Sung Cho, a co-head of public tech investing, said in a December 2025 interview that 90 percent of AI capital spending to date came from hyperscalers' operating cash flow and 10 percent from corporate debt (World Finance, 2026). The BIS described 2025 and 2026 differently, with the largest companies spending beyond free cash flow. I cannot reconcile the two accounts from public sources, and the gap matters because the financing mix decides how a delay lands.

A vendor funded from leases and credit has a repayment schedule that keeps running while the payoff slips.

The Journal reported this week that two Australian data center companies shelved plans to sell shares to the public, citing market volatility, and that investor nerves over borrowing costs and AI lab profitability hit the bonds of infrastructure providers such as CoreWeave. The same funding test applies to the graphics processing unit brokers I analyzed in May.

Unrelated forecasts converge on 2027, though none can time it

Joachim Klement of Panmure Liberum told Bloomberg his base case is that the AI trade disintegrates as soon as 2027, which would send stocks sharply lower. Peggy Noonan's Wall Street Journal column draws on a different set of voices. It cites Demis Hassabis, who runs Google DeepMind, calling recent AI hacking incidents warning signs of more serious ones in 2027 and after.

Cyber risk sits on the IMF's own list of three. The fund found that software providers had almost no time to fix vulnerabilities before exploitation in more than 70 percent of cases recorded in 2026, up 20 percentage points from the prior year (Hedgeweek, 2026). This month, hackers used a Chinese AI program to attack South Korea's biggest banks and take personal information on 68,000 people (Douglas and Landers, 2026).

Patience also has to survive an energy shock.

The fund's 3.4 percent growth forecast for 2027 assumes the Strait of Hormuz reopens by March (Lawder and Shalal, 2026), and public debt stands at its highest level since World War II (Wee, 2026). A vendor that needs cheap capital to wait out a slow payoff is exposed to both.