Power, Not Chips, Now Gates the AI Build

Oracle holds rights to up to 2.8 gigawatts of Bloom Energy fuel cells. For one New Mexico campus, it also holds a force majeure notice. That pairing is the AI buildout in miniature: the scarce input is no longer silicon but electrons, and the permission to use them.

The thesis is simple. The advantage in AI infrastructure is passing from whoever can buy the most chips to whoever can secure power and local consent: developers who bring their own generation, offer load the grid can switch off, or win the neighbours' blessing before they pour concrete. Three developments from the past week make the case.

Self-supply gets paid for

Bloom Energy reported record third-quarter revenue of $1,235.8 million, up 138.1% from a year earlier, and raised its full-year guidance to between $4.4 billion and $4.6 billion. CFO Simon Edwards said demand gave the company confidence "to raise our 2026 outlook for the third time this year." CEO KR Sridhar described a move away from a century-old model of centralized power toward firm, onsite generation, and said what began with hyperscalers is spreading across market segments.

The economics are Bloom's own claim, relayed by Motley Fool: in the company's cost model, a 1-gigawatt AI data center run on its systems could spend $3.6 billion (27%) less on non-chip construction than one built around standard AC power. Treat that as a vendor's model. Oracle's behaviour is better evidence. When it expanded the Bloom partnership, Motley Fool noted, its executive talked about speed rather than cost. "By rapidly deploying Bloom's reliable, efficient fuel cell energy, we are quickly meeting the demands of our customers across the United States." said Oracle Cloud Infrastructure's Mahesh Thiagarajan.

Speed is what a grid interconnection queue cannot sell.

Flexibility as a ticket to connect

The second route is to ask the grid for less, at the right moments. Reuters reported on Oct. 8 that technology companies and utilities are exploring ways to make data centers more flexible in when and how they draw power. OpenAI has agreed to cut its grid draw by up to 1 gigawatt from a planned 3.2-gigawatt Georgia facility during periods of grid stress. Federal regulators have ordered grid operators to consider new connection rules, including faster pathways for facilities that offer flexibility.

The prize is large. A Duke University Nicholas Institute study, as Reuters reported, put capital savings from greater data center flexibility at $40 billion to $150 billion over the next decade. The American Public Power Association's summary adds that new gas combined-cycle investment falls by 10% in the flexible scenarios, and that retail prices fall by roughly 0.5% to 2.8% because the system is relieved during the highest-cost peak hours.

Reuters is candid about how far this has travelled. The "demand response" strategy has so far mostly lived in pilots and one-off agreements. BMI analyst Alexander Kheder told Reuters that scaling curtailment deals across hundreds of new facilities "will demand significant capital expenditure and coordinated policy frameworks."

Consent is now a line item

The third piece of evidence is Loudoun County, Virginia. Amazon has dropped plans for four data centers on a 122-acre former George Washington University campus, Tom's Hardware reported, and will repurpose most of it for mixed-use and residential development after community pushback. The GW Hatchet reported the plan was for 800,000 square feet of data centers under the name "Bridgefield Tech Center," and that residents circulated a moratorium petition citing resource use, traffic, noise and pollution.

This was not a developer failing to find power. It was a developer failing to find permission, in a county with the highest concentration of data centers in the world, according to the GW Hatchet: 54 million square feet already built.

Loudoun Now's account of the settlement is blunt. Amazon will donate $5 million to the county's Housing Fund, forgo legal action and withdraw its application, while the county fast-tracks a rezoning in which data centers are barred. Board Chair Phyllis Randall said she told Amazon at the first meeting, "I don't want to see any data centers, and I want to see housing." AWS's Josh Weissman, asked whether Amazon would buy the campus again, answered: "In hindsight, would we have bought the campus that we did? I mean, probably not." Amazon had paid $427 million for it, Loudoun Now reported.

Delegate John McAuliff was less diplomatic. "Loudoun is done with data centers. I hope other companies follow suit." The county is not finished with the subject, either: its supervisors were set to consider a temporary pause on new projects but pushed the item to a November meeting, the GW Hatchet reported.

The objection that bites

Bring-your-own-power is no cure, and the evidence is in the same file. Motley Fool, relaying a Sept. 24 Bloomberg report, said Oracle sent a force majeure notice tied to Project Jupiter, the New Mexico campus Bloom's fuel cells are due to power, because the natural gas pipeline meant to supply it has been delayed. Bloom's own risk list includes "air permitting delays which affect the ability to finance the use of Energy Servers for customers," and any actual or perceived slowdown in the adoption of AI. And Nvidia's 800-volt DC roadmap, with full-scale production due alongside its Kyber racks in 2027, listed Eaton and GE Vernova as partners but not Bloom, Motley Fool noted.

Those are real risks to Bloom as a stock. They do not rescue the old model, though. A delayed pipeline and a contested air permit are the same bottleneck wearing different clothes: physical supply and public consent. Self-generation shifts the queue from the utility to the gas network and the permitting office. It does not abolish it. Demand also is not waiting. The Energy reported that AEMO's latest data showed underlying demand across Australia's main grid hitting a record in the September quarter, and that the number of data centre projects in its implementation phase rose to 10, despite recent jitters over data centre energy supply.

Nor is flexibility a free lunch. Kheder's warning about capital and policy coordination stands.

What follows

Reuters, citing the Electric Power Research Institute, reported that US data center electricity use could rise from about 177 to 192 terawatt-hours in 2024 to between 383 and 793 TWh by 2030. That range is too wide to plan around, and the spread is the point: how much of it gets built depends less on chip supply than on who can show a grid operator a way to connect, and a county board a reason to say yes.

Expect the winners to look less like chip buyers and more like utilities with servers: developers who arrive with generation, a curtailment clause and a housing fund. Amazon, which paid $427 million for land it will not build on, has just supplied the price of learning that in reverse.