aceModerate evidence
More
Paper estimates $9 trillion US AI buildout, with debt risks hidden off balance sheets
Study title: Financing the AI Buildout
- Evidence type
- Observational studyPreprint, not yet peer reviewed
- Sources
- 1 sourcePeer-reviewed
- Published
- 3 min read
- Truvace record
- TRV-2026-2479cite or check it
Study title: Financing the AI Buildout
A Columbia Business School economist estimates in a working paper that a 1 GW AI campus costs about $41 billion. A 188 GW US buildout completed by 2032 would need nearly $9 trillion of investment, averaging 3.2 percent of GDP a year over 2025–2032, more than the canal, railroad, electrification, highway and telecom booms.
Read the study abstract
Artificial intelligence is driving a large expansion in data center capacity, power infrastructure, and specialized computing equipment. We estimate that a 1 GW AI campus costs about $41 billion. A 188 GW U.S. buildout completed by 2032 would require total investment of nearly$9 trillion and amount to annual investment of 3.2 percent of GDP over 2025–2032. As hyperscaler capital expenditures outgrow internal cash flow, financing is shifting toward leases, joint ventures, project debt, private credit, securitizat…
The problem
A working paper estimates a 188 GW US AI buildout by 2032 would cost nearly $9 trillion, and that leases, project debt and special-purpose vehicles raise asset-level leverage while hiding obligations from hyperscaler balance sheets.
In brief
Capital spending by Oracle, Microsoft, Amazon, Meta and Alphabet rose from about $97 billion in 2020 to more than $400 billion in 2025 and is projected to top $800 billion in 2026, passing their combined operating cash flow for the first time. The firms are leaning more on leases, joint ventures, private credit and securitization.
The author argues these structures widen funding capacity but make leverage harder to see. He says financial distress is not imminent and that AI demand growth could support the projected investment; the buildout figure is a scenario, not a forecast.
Main points
- A representative 200 MW AI training campus costs roughly $8.2 billion in 2025, with about two-thirds going to IT equipment and one-third to real estate and power infrastructure.
- The central scenario has 188.3 GW of new US capacity operating in 2025–32, 114.9 GW more after 2032, and 213.5 GW of announced capacity never completed.
- Moody's estimates hyperscalers have about $970 billion of lease commitments, roughly $660 billion of them future leases not yet on balance sheets.
- Aggregate leverage for four hyperscalers rises from 20% to 48% (debt to capital) and from 0.7x to 2.6x debt to EBITDA once lease obligations are counted, the author calculates.
- Meta's Hyperion project: $30 billion, 2.0 GW, with $27 billion of debt (about 90% of assets) and a debt service coverage ratio of 1.12x.
- Five-year CDS spreads widened for all five hyperscalers between November 2025 and July 2026; Oracle's rose from 118.8 to 205.5 basis points.
- Data center REIT market beta has risen from roughly 0.5 to around 1, so they now move with the broader stock market.
The full story
The paper frames AI as a physical capital boom. Modern AI racks can draw well above 100 kW versus 5–10 kW for traditional racks, and a continuous 200 MW campus uses as much electricity as 170,000 average US households. At high utilization, capacity completed by 2032 would use roughly as much electricity yearly as the entire US residential sector (1,515 TWh in 2025).
The Hyperion deal shows the structure. Meta sold an 80 percent stake to Blue Owl for about $2.5 billion, and the joint venture Beignet raised $27 billion of A+-rated debt in October 2025 at a 6.58% yield, at least 100 basis points above Meta's likely corporate rate. That implies over $5 billion of extra interest, reflected in Meta's lease payments. Meta holds five four-year leases from 2029 and a residual value guarantee.
The author identifies vulnerabilities in tenant concentration, technological obsolescence and execution delays. To earn a 10% unlevered return, the buildout needs about $3.55 trillion in mature annual revenue, roughly 8.8% of 2032 GDP, or $6.4 per billed GPU-hour at 80 percent utilization.
He proposes priorities of better measurement, clearer disclosure of leases and guarantees, and study of asset valuation and financial stability. He notes that circular deals, such as Nvidia's 25% residual value guarantee on a $500 billion chip credit facility, link exposures across the AI value chain.
Sources
1 on file- NBER Working Paperdoi:10.3386/w35865
Talk it over
Who carries this cost, and what would have to change to prevent it?
Paper estimates $9 trillion US AI buildout, with debt risks hidden off balance sheetshttps://truvace.com/item/financing-the-ai-buildout
Who carries this cost, and what would have to change to prevent it?
The problem A working paper estimates a 188 GW US AI buildout by 2032 would cost nearly $9 trillion, and that leases, project debt and special-purpose vehicles raise asset-level leverage while hiding obligations from hyperscaler balance sheets.
Answer in the threadHow do we know? What can this kind of evidence show, and what can’t it?
The evidence This rests on an observational study, from NBER Working Paper. It is a preprint, not yet peer reviewed.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat The 188 GW figure is a scenario rather than a forecast, sensitive to large multiphase projects and projects without reported completion dates, and the author says it is premature to conclude AI infrastructure poses systemic risk.
Answer in the threadHave you seen AI used this way where you study, work or live? Did it go the same way?
Answer in the thread
The questions are the same on every record. The details under them are taken from this record as written.
What readers say