The $14 billion Meta–BlackRock data center in Texas reportedly carries no insurance against a complete loss, because full coverage has become prohibitively expensive. S&P rated the project's debt A+, slightly below Meta's corporate rating, partly because lenders have no direct claim on the physical assets and the lease contains cancellation provisions. That is what the AI buildout looks like when you follow it all the way down: not a capex line, but a lender underwriting the durability of a tenant's payments rather than the resale value of the building.
Last week this column traced SpaceX's turbine foundry and the power constraint tightening around AI. This week the constraint acquired a number, and the number has a name.
Moody's put roughly $2.8 trillion on the future commitments of the five largest hyperscalers — Amazon, Microsoft, Google, Meta and Oracle — none of it currently sitting in their liabilities. Moody's flags the size, complexity and limited transparency of these arrangements and says they require deeper analysis. The figure is not conventional debt, and calling it that would be wrong. It is a stack of contractual obligations that travel through leases, purchase agreements, project debt, guarantees and special-purpose vehicles before they ever touch a cash-flow statement.
It is a stack of contractual obligations that travel through leases, purchase agreements, project debt, guarantees and special-purpose vehicles before they ever touch a cash-flow statement.
The composition matters more than the headline. A Wall Street Journal analysis found nine major tech companies with roughly $3 trillion of commitments absent from their balance sheets, against about $600 billion of reported capex over the prior year. That gap splits into $1.2 trillion of leases for facilities that have not begun operating and $1.9 trillion of commitments to buy chips, equipment, energy and services. Alphabet alone disclosed $811 billion of purchase and contractual obligations, up from $332 billion three months earlier. Meta reported $347 billion of future lease commitments.
Traditional debt markets are showing the strain. J.P. Morgan calculates hyperscaler bond issuance rose from $17 billion in 2024 to $109 billion in 2025, then hit $194 billion in the first half of 2026 alone; it expects $279 billion for the full year and another $220–$300 billion in 2027. At that pace the group approaches 10% of the US investment-grade market by 2030. Nomura estimates the largest tech groups have borrowed roughly $200 billion this year — about 25% of the US Treasury's net issuance of notes and bonds to private investors, and five times their share in 2025.
The pricing is where this stops being abstract. Alphabet's recent 30-year debt yielded almost 6.4%, roughly 115 basis points above comparable Treasuries. A $12.5 billion bond issued to finance a Meta-linked data center yielded 7.53%, a 287.5 basis point spread — the widest for an A-rated or better bond in three years, and about 200 basis points wider than comparable debt. The AAA corporate spread over Treasuries widened from 32 basis points in May to 41 in August; AA went from 47 to 58. Penn Mutual Asset Management estimates bonds from Microsoft, Amazon, Alphabet, Meta and Oracle traded 20–25 basis points wide of the broad investment-grade index in early 2025, reached almost 50 in the first half of 2026, and sat around 35 in July.
Moving a project off the balance sheet does not delete its risk. It redistributes it. Meta could face up to $28 billion under a single residual-value guarantee, with no liability recorded because payment isn't currently considered probable. Moody's has warned that lease accounting can omit the cost of renewing a data-center lease or compensating investors if a company walks away and the facility loses value.
Then there is the physical chain, where the financing story and the electricity story become the same story. Moody's July report is titled Power Without Delivery, and its conclusion is blunt: a completed AI data center is not necessarily a cash-flow-generating asset, and power availability, transmission infrastructure and energization timelines are becoming financing risks. A press release does not produce electricity. A utility queue does not energize a server. A turbine order does not produce power.
Nuclear is where that gap is widest. Investors have poured $4.6 billion into US nuclear startups this year, up from $3.8 billion in 2025. Yet only two new commercial reactors are under construction in the US and Canada, and no other such projects have received regulatory approval. Bloomberg Intelligence doesn't expect widespread commercialization until about 2035.
The quotes from the Texas Nuclear Summit in Houston are sharper than any model. "The reality is starting to sink in," said Scott Levine of Bloomberg Intelligence. "This is going to take longer than expected." Max Hermanson of ClearView Energy Partners put it better: "The money is arriving faster than the projects are maturing. The nuclear renaissance is currently still showing up more on balance sheets than construction sites."
The demand side keeps compounding. Goldman Sachs expects US data center power demand to reach 66 GW in 2027, more than double the 31 GW seen in 2025. Meta, Alphabet and Amazon have all signed deals to use nuclear energy once reactors get built. Valar notched a $1 billion round in August; Antares raised $470 million in July; Radiant raised $300 million in December. Aalo Atomics is poised to close a Series C north of $500 million, and its CTO Yasir Arafat says bankers are waiting: "They want to see technologies proven in the field. They haven't seen the tech proven." The last US nuclear project — a pair of reactors in Georgia — arrived seven years late and more than 100% over budget.
Gas is getting the capital that nuclear can't yet absorb. Parallax launched with a $117 million raise led by Founders Fund, Eclipse, Lux, Diffusion, Greylock and General Catalyst to build gas turbines designed specifically for AI data centers. Greylock's framing is the whole thesis in two sentences: "Intelligence can now be delivered in seconds, while the data centers and power infrastructure behind it can take years to bring online."
Europe is funding the grid layer rather than the generation layer. London's Reactive Technologies secured €19 million in venture debt from the European Investment Bank, split between €7 million for R&D in Oulu, Finland, and €12 million for EU deployment. "The rapid growth of renewable energy and large power consumers such as data centres are making Europe's grids more complex to manage," said CEO Marc Borrett. The EIB's Karl Nehammer added that the financing supports "the secure integration of more clean power."
Fusion is the long-dated option on all of this. Commonwealth Fusion Systems raised another $1 billion in July, bringing its total to $4 billion, and in April became the first fusion company to apply to PJM Interconnection. Helion's Polaris prototype heated plasma to 150 million degrees Celsius in February; its Orion plant, a 50-megawatt machine intended to supply Microsoft as early as 2028, rode a Series G from $465 million announced in June to $500 million closed in September.
At New York Climate Week, two founders were asked whether they'd prefer the AI buildout to proceed at its current pace or a more climate-responsible one. Both said faster, without hesitation. Both ran energy startups. Climate tech venture deal value has risen four consecutive quarters, cresting $14 billion in the first quarter — the best fundraising environment in years, driven by data center construction, grid infrastructure and dispatchable power.
The undercurrent was that the party won't last forever, but might last long enough. The question this column keeps returning to is narrower and more uncomfortable: how many megawatts were forecast, bought and sold, versus how many actually get energized. That gap decides whether trillions in commitments become on-time cash flow or something that reprices first at the project level — long before it ever reaches an index.