Inside the AI power wars

The hyperscalers are not interchangeable. Ontiveros puts Google at the frontier: the biggest energy trading desks, large load flexibility deals, and a name that appears in the proceedings of grid operators like PJM and ERCOT more than anyone else’s. Kann traces that to a history of early power purchase agreements, the first 24/7 clean energy commitment, and shifting workloads geographically for carbon reasons well before AI made flexibility fashionable. Microsoft was the clear second for years; Ontiveros now ranks it fourth, in a market where the few people who know how to procure data center power change jobs the way AI researchers do. Amazon rose on unusual deals such as Talen and Vistra’s Comanche Peak. Meta does the strangest things for a structural rather than cultural reason: with no cloud business it serves only its own workloads and builds for itself rather than for maximum fungibility, which Kann suggests may free it from the four nines the others owe customers. That is how it can run behind-the-meter gas with roughly 10 megawatts of backup on a 100 megawatt site.

Why go behind the meter? Almost none of it is off-grid ambition. These are bridges to a grid connection expected in a few years, and the case is arithmetic. Buildout runs in the tens of gigawatts a year growing around 50% annually, against maybe five or six gigawatts of gas next year and twenty to twenty-five gigawatts of nameplate solar and batteries worth far less once adjusted for peak contribution. Utility commitments are also soft: the story he says every developer tells is a promised gigawatt by 2027 revised to a hundred megawatts by 2029 with the gigawatt in 2042, and no penalty for missing. For a lab that is intolerable, because power is revenue rather than overhead. Anthropic goes from about 1.5 gigawatts at the end of 2025 to wanting more than 10 by 2027, the size of Google in two years, and nobody commits that capital against a date somebody else controls. But Kann declines the framing. He bets far more generation gets added and argues the chokepoint is transmission and distribution, where a substation upgrade or a high-voltage transformer runs years. Pressed, he guesses roughly 35 gigawatts total for 2027 or 2028; Ontiveros’s adjusted figures are lower and he calls both small against fifty gigawatts a year of data centers. They disagree about which constraint binds first, not whether one does.

The reframe is the most useful thing here: ask about the AI labs rather than the hyperscalers, because roughly half of what Amazon and Microsoft build now goes to OpenAI and Anthropic. The labs’ binding limit is credit, not ambition. They are not investment grade, the capital is front-loaded, and a construction loan at high loan-to-cost needs an investment-grade signature. That gap is the product hyperscalers sell into, and Google’s response was the aggressive one. Instead of buying equity the way Nvidia did with the neoclouds, the specialist GPU rental companies, Google lent its own credit to Anthropic’s buildout, backstopping about a gigawatt with third-party developers and carrying on the order of $50 billion of commitments. Ontiveros reads this as commercial rather than generous, since that support is what lets Anthropic buy Google’s own AI chips, its tensor processing units, at around $20 billion a gigawatt.

The same logic explains who loses and what gets bought. CoreWeave scaled contracted power from about 1.3 gigawatts to three and a half, then stalled after the third quarter of 2025 when a bond selloff froze the high-yield market, while rising turbine and interconnection deposits structurally disadvantage anyone without a balance sheet. Ontiveros calls this existential for Nvidia and expects it to start lending its own credit in the second half of 2026. What gets bought follows a tier list ordered by how fast it energizes: premium turbines, then aeroderivative and industrial units, then reciprocating engines, then fuel cells. Bloom inverts the usual bridge logic, fast to deploy when permitting stalls but two days from zero to full output, so it cannot double as backup and choosing it means islanding more or less for good. Clean firm generation sits outside the ranking, because nuclear and geothermal contracts are mostly non-binding and milestone-contingent: options on long-term capacity rather than answers to speed to power, a framing Kann proposes and Ontiveros endorses, then extends into a call that gas turbine orders top out this year.