Brookings Paper Puts the AI Build-Out at $10.3 Trillion

Brookings posted “Financing the AI buildout” on September 23, 2026, a summary and conference draft of a paper by Stijn Van Nieuwerburgh, the Earl Kazis and Benjamin Schore Professor of Real Estate at Columbia Business School. It is one of three papers on the Panel on AI and Policy at the fall 2026 Brookings Papers on Economic Activity conference, scheduled for 9:00 a.m. Eastern on Friday, September 25. The draft itself is dated September 4, 2026 and carries the standard Brookings notice that each publication represents the sole views of its author. Its central estimate is that investment in data center buildings, power systems, networking and specialized chips will total $10.3 trillion from 2025 to 2032, an average of 3.63 percent of U.S. GDP a year. The paper’s Table 1 sets that against earlier booms as average annual capital expenditure over each period: canals at 0.66 percent of GDP from 1836 to 1841, railroads at 2.24 percent from 1870 to 1890, electrification at 0.50 percent from 1905 to 1925, highways at 1.13 percent from 1956 to 1973, and telecom and fiber at 1.10 percent from 1996 to 2003. The AI figure is a projection; the others are history.

The estimate is built bottom up, and the method is the part an operator can check. Using Cleanview project-level data from July 2026, the paper counts about 57 GW of operating U.S. data center capacity and a planned pipeline of about 509 GW, which it says plainly is not a completion forecast. Its central scenario completes 182.8 GW during 2025 to 2032, including 19.0 GW tied to projects already operating in 2025 or 2026, completes a further 117.2 GW after 2032, and never builds 226.9 GW. Setting aside the 19.0 GW already operating, the remaining three buckets sum to 507.9 GW, which computes to the paper’s approximate 509 GW pipeline. Costs come from a representative 200 MW AI training campus at roughly $8.2 billion: about $2.2 billion for the facility, $0.4 billion for incremental power infrastructure and $5.6 billion for the installed IT system, with GPU racks at $4 million each. The abstract scales that to about $41 billion per GW. On financing capacity, the paper reports that capital expenditures at Oracle, Microsoft, Amazon, Meta and Alphabet rose from about $97 billion in 2020 to more than $400 billion in 2025 and are projected to exceed $800 billion in 2026, above their combined operating cash flow for the first time. Its figure puts 2026 at $800.5 billion of capex against $707.1 billion of operating cash flow, which computes to about 113 percent.

The trillion-dollar total is mostly chips. By the paper’s own breakdown, the facility and power infrastructure are about one-third of a campus’s initial cost and the compute hardware about two-thirds, and hyperscalers tend to point their own capital at the IT equipment. The buildings, shells and power are the layer being moved to third parties. The paper describes developers, REITs, infrastructure funds and private equity building or buying facilities and leasing capacity to hyperscalers, financed by bank mortgages, syndicated lines, private credit and securitization, and it says data centers have become a major institutional real estate asset class. That is the CRE exposure, and the paper’s worked example is Meta’s Hyperion financing in Louisiana. For the financed phase of roughly 2.0 GW and about $30 billion, the paper says Meta sold an 80 percent equity stake to Blue Owl for about $2.5 billion, and the joint venture, Beignet, raised $27 billion of debt in October 2025, rated A+ by S&P, one notch below Meta. The paper puts debt to assets near 90 percent, the debt service coverage ratio at 1.12x, and the yield at 6.58 percent, at least 100 basis points over what it estimates Meta would have paid unsecured, or more than $5 billion of extra interest over the financing’s life.

The lease terms are where a landlord’s underwriting changes. According to the paper, Meta holds a sequence of five four-year leases beginning in 2029 and running to the bonds’ 2049 maturity, may terminate some or all of the campus at each renewal, and in that case must cover any shortfall between sale proceeds and a contractually specified minimum value. Under current GAAP, the paper says, neither the future leases nor that residual value guarantee sits on Meta’s balance sheet before it takes effect, even though one of the two will. It cites Moody’s estimate of roughly $970 billion of hyperscaler lease commitments, about $660 billion of them for leases not yet on balance sheet, and says Meta and BlackRock’s 960 MW Sopaipilla project in El Paso, Texas used a nearly identical structure for $12.3 billion of A+ bonds in July 2026. It adds two market signals. The 36-month rolling beta of data center REITs, measured on FTSE Nareit index returns, has risen from roughly 0.5 to around 1, which the author reads as data centers behaving less like defensive infrastructure and more like part of the same risk complex as AI stocks. And under its central assumptions, the 2025 to 2032 buildout needs about $3.7 trillion of annual revenue by 2032 to earn a 10 percent unlevered return at a 50 percent operating margin, roughly $6.9 per billed GPU-hour at 80 percent utilization.

The artifact worth building is an exposure register for any data center an owner holds, lends against or is bidding, one row per asset, assembled from documents rather than marketing. Inputs: the lease and every amendment; any residual value, construction or completion guarantee; the ownership chart down to the special-purpose entities; a rating agency pre-sale report where the debt is rated; the tenant’s most recent 10-K and 10-Q lease commitment footnotes; and the interconnection and power agreements. Output: columns that answer the paper’s questions for one building. Who legally owns the asset, and who bears demand, refinancing and residual value risk, which the paper argues are no longer the same party. The tenant’s termination and renewal dates. The guaranteed minimum value and who stands behind it. Debt to value and debt service coverage at the property level, not the tenant’s corporate leverage. The share of rent from one tenant. And a line for what is not disclosed, because the paper notes that publicly rated debt is a small share of data center financing and that syndicated bank loans and private credit, which are much larger, publish little. Every populated cell cites a page of a contract or a filing.

The reviewer is the acquisitions or credit lead, with counsel on the guarantee and termination language, and the approval gate is specific: no valuation, loan sizing or exit cap assumption treats a terminable lease with a residual value guarantee as equivalent to a fixed 20-year term until someone has modeled the renewal dates and read who pays the shortfall. An assistant is useful for the volume here, pulling lease commitment tables out of five hyperscalers’ filings, extracting renewal and termination clauses from a lease stack, and flagging where an entity chart has an unexplained layer. It should not decide the probability of a renewal, value a stranded 200 MW shell, or convert the paper’s national scenario into a view on one market. PSV ran no model on this paper, tested no product, and promises no underwriting, valuation or return outcome.