Coatue: AI capex could hit $2.8tn by 2028 and will need its own Fannie Mae and Freddie Mac
In its October 2026 Public Markets Update, published on 7 October, Coatue estimates that AI capital spending by hyperscalers, AI labs and neoclouds could reach about $2.8tn in 2028. The deck sets that against the largest US capital pools: about $2.1tn of US single-family mortgage originations expected in 2026, the roughly $1.0tn US defence budget for FY26 and about $0.8tn of non-AI capex by S&P 500 companies in 2025. On those numbers, 2028 AI capex would be about 1.5x mortgage originations, 3x the defence budget and 3.5x non-AI corporate capex. The slide's headline: "AI capex will soon need its own Fannie & Freddie to standardize, finance and scale".
The next slide traces how GPU and data-centre funding has evolved: from hyperscalers simply paying capex, to off-balance-sheet set-ups in which neoclouds own the GPUs under take-or-pay contracts (with GPU financing illustrated at roughly 75% debt and 25% equity, backed by guarantees), to multi-layered data-centre joint ventures funded by credit funds, hyperscaler equity and debt, with facility leases, residual value guarantees and separate GPU-layer compute contracts. Coatue cites publicly reported structures such as Meta's Hyperion data-centre joint venture with funds managed by Blue Owl Capital and Nvidia's capacity agreement with CoreWeave as the basis for the framework.
The same deck shows Goldman Sachs' basket of AI winners up about 405% since ChatGPT launched (30 November 2022 to 6 October 2026), against about 2% for its basket of AI losers.
The figures come from Coatue's published slides. Coatue labels them estimates for illustrative purposes, says the AI capex figures do not represent projected debt-financing demand, and the Fannie and Freddie comparison is Coatue's view rather than any policy proposal.
The Economist's cover leader this week ("Will bonds blow up?") links rising rich-world government borrowing costs partly to the capital demanded by AI data centres. Together with Coatue's framing, compute is becoming a capital-markets story: how GPU and data-centre debt is standardised, who carries the residual value risk on chips, and what that costs will shape the economics of neoclouds, data-centre developers and the labs that rent from them.
Coatue, Public Markets Update (October 2026): https://www.coatue.com/blog/perspective/public-markets-update-2026-10-07
Slides shared on X by Mark Goldberg: https://x.com/Mark_Goldberg_/status/2108705247447466453
Winners vs losers slide shared by Peter Walker: https://x.com/PeterJ_Walker/status/2108993339584565255
The Economist, Will bonds blow up?: https://www.economist.com/leaders/2026/10/08/will-bonds-blow-up
