America's $10 Trillion AI Buildout Is Straining on Financing, Power and Permits, Not Demand

The Financial Times reported this weekend that America's AI infrastructure buildout, which headlines now size at about $10 trillion, is running into financing pressure, construction setbacks and local opposition. The framing in the summaries we reviewed is a test: can developers deliver the capacity they have promised, and will investors keep paying for it.

That is a story about supply, and the demand side is not what is wobbling. This post lays out what is documented, what is not, and where the pressure lands for people who buy compute rather than build it.

What Actually Happened

The Seeking Alpha summary of the FT piece says America's race to build computing infrastructure for AI "is encountering financing pressures, construction setbacks and local opposition," and that these difficulties are testing whether developers can deliver the capacity promised to technology companies and AI businesses, and whether investors will keep funding the expansion. That is as much of the FT article as we could read. The full piece is paywalled, so we are not going to claim details from it that we have not seen.

The $10 trillion headline number has a plausible source. Crypto Briefing, in an October 18 article, cites a 2026 paper by Stijn van Nieuwerburgh of Columbia Business School that projects $10.3 trillion of US AI infrastructure investment between 2025 and 2032. We could not confirm that the FT used the same paper, so treat the match as likely and unverified.

Per Crypto Briefing's summary of the paper, that spending works out to about 3.63% of US GDP per year. The comparison point is the 19th-century railroad buildout, which peaked at roughly 2.2% of GDP. Hyperscaler capital expenditure from Microsoft, Amazon, Meta, Alphabet and Oracle is projected at $4.2 trillion through 2029, and the paper expects that spending to exceed those companies' operating cash flow by 2026. Even the largest balance sheets would then rely mainly on outside capital.

The Constraints Are Physical and Financial

Crypto Briefing lists three bottlenecks: power supply, shortages of skilled workers, and construction delays. It reports that Google has cited expected grid-connection delays of five to ten years driven by data-center demand, that the pressure is sharpest in regional grids such as PJM and ERCOT, and that the research flags a potential power shortfall of up to 68 GW for data centers.

None of those constraints are fixed by a bigger check. A developer with money still waits on an interconnection queue, a transformer order or a county board.

The Reuters reporting from September 24 shows how that becomes a financing problem. Oracle issued a force majeure notice in connection with Jupiter, a data center campus that Blue Owl's STACK Infrastructure unit is building in New Mexico to support OpenAI. A person familiar with the notice told Reuters that Oracle invoked it because of delays securing power, and that the program is being delayed by a year. Blue Owl has about $3 billion of equity in the project and earns a lower return during construction, with returns rising once the data center is complete, so a later completion date pushes back the higher returns.

Blue Owl's public line, per Reuters, is that "Blue Owl, STACK Infrastructure and Oracle remain fully aligned on Project Jupiter" and that the notice "does not change the financial commitments to this multiyear project." Oracle shares fell 3.5% on the Thursday the notice was reported, and Blue Owl closed 3.6% lower.

Reuters also reported that the event was reverberating through other proposed financings, including SB Energy's campus for OpenAI in Ohio, and that SB Energy decided that week to delay its IPO. Earlier this year, Reuters noted, Blue Owl struggled to syndicate financing for a $4 billion data center project in Lancaster, Pennsylvania, anchored by CoreWeave.

Where the Money Gets Cautious

The quote in the Reuters piece that sums up the tension comes from Sean McDevitt, a partner at Arthur D. Little: "The financing side of the AI buildout is starting to ask much harder questions than the demand side." He added that underlying demand still appears very strong, and that investors and lenders are focused on how risk is allocated.

Reuters adds context on rates and volume. US interest rates are at levels last seen two decades ago, and Wall Street is scrutinizing deals that looked settled months ago, so the presence of a highly rated company like Oracle is not enough on its own to settle investors. Moody's projects AI-related capex by the six biggest US technology companies at around $1 trillion in 2027, which raises the question of whether markets can absorb that much issuance without wider pricing pressure or tighter terms.

Contract language is moving too. Rajat Rana, a partner at Quinn Emanuel, told Reuters that a data center contract from January and one signed today are "massively different." Lawyers say force majeure provisions are becoming more common in data center agreements, which concerns lenders. The practical effect is that delay risk, which used to sit with the developer, is being negotiated into who pays when a project slips.

We covered the market's price for this risk earlier in the week, when CoreWeave's bonds traded at roughly 15% yields. The Jupiter notice is the same pricing signal arriving through a contract rather than a bond screen.

The Revenue Gap Sits Under All of It

Crypto Briefing's summary of the Columbia paper says the AI industry would need roughly $3.7 trillion in annual revenue by 2032 to support the projected investment, while current revenue estimates from major players sit around $100 billion. It adds that the International Monetary Fund warned as of October 2026 of potential stock-price corrections if profitability on AI investment trends downward.

We would handle the $100 billion figure carefully. Anthropic alone reported a third-quarter revenue of $16.8 billion, per Reuters as relayed by Fortune India in our October 17 post, so the paper's baseline is likely defined narrowly or dated. We have not seen the paper, and the definition matters. The direction of the argument, that required revenue is a large multiple of current revenue, does not depend on the exact baseline.

On the community side, Reuters cites research firm Data Center Watch counting at least 45 projects worth $68 billion that faced opposition from community groups in the second quarter of 2026, after 75 projects worth about $130 billion were disrupted in the first quarter. Reuters uses different verbs for the two quarters, so we would not read a trend into the drop. It names Meta's El Paso, Texas project and Related Digital's planned $16 billion Oracle-focused campus in Saline Township, Michigan among developments that have faced resident opposition.

What Builders Should Take From It

• Treat power and permits as compute supply constraints. If your roadmap assumes capacity coming online on a date, ask the provider where its power is secured and what interconnection status the site has. A signed contract is weaker than an energized site.
• Read delay and force majeure language before you sign. Reuters reports these provisions are spreading. Know whether a slipped site gives you a credit, an exit or nothing.
• Keep a second provider warm. A one-year slip at a single campus, as reported for Jupiter, is long enough to miss a product cycle. Multi-provider routing costs engineering time now and saves a scramble later.
• Expect lender caution to show up in prices and terms. Higher financing costs for developers tend to move into contract pricing for customers. Budget for capacity prices that rise or stay flat rather than fall on schedule.
• Separate demand risk from delivery risk in your own planning. Nothing in the reporting we reviewed says AI usage is falling. The risk is that the capacity you expect is late or more expensive.
• Watch the revenue line. The Columbia argument turns on AI revenue climbing toward the multi-trillion-dollar level the buildout implies. If it does not, the pressure shifts from construction delays to the repricing of projects already underway.

The useful reading of this week's coverage is narrow. Money, power and permits are the binding constraints on the buildout right now, according to the reporting in hand, and each one moves on a different clock. If you buy compute, the date on the contract matters less than whether the site behind it is energized.

Developer312 covers the AI business signals builders actually need to act on. Get the weekday briefing at developer312.com.