<
Welcome to OKRVV B2B Platform.
HOME > NEWS > News

The AI Bottleneck Moved to Power, and AI Labs Became Energy Buyers

Core Tip:The scarce resource in artificial intelligence is no longer the chip. It is electricity, and the equipment that makes it. Two disclosures this month made that concrete: Anthropic, the company behind Claude, agreed to pay about 3.4 billion dollars for gas-

The scarce resource in artificial intelligence is no longer the chip. It is electricity, and the equipment that makes it. Two disclosures this month made that concrete: Anthropic, the company behind Claude, agreed to pay about 3.4 billion dollars for gas-fired generators, while the geothermal developer Fervo reported that its contracted backlog had jumped to 11.9 billion dollars. AI companies are no longer simply customers of the power industry. They are becoming its most aggressive buyers, and the rest of the B2B supply chain should price that in.

According to the Financial Times, Anthropic secured commitments in late September for 1.7 gigawatts of heavy-duty generators from OnePWR Solutions, a Texas company that builds microgrids and on-site power plants, in a deal covering engineering and construction as well as equipment. The machines were not fresh orders. OnePWR had ordered them from Caterpillar about two years earlier but lacked the financing to complete the purchase, and Anthropic stepped in to take over the orders rather than buy the supplier. At roughly two million dollars per megawatt, it is the largest known single purchase of power equipment by an AI lab.

The reason for bypassing the utility queue is lead time. Gas turbines and generators now carry delivery timelines measured in years, and Anthropic's own IPO prospectus warns that such equipment must be ordered years in advance and has been subject to extended delivery timelines, price increases, tariffs and rationing as data-center operators compete for it. The company has told investors it carries roughly 895 billion dollars of compute and infrastructure commitments through 2045. Owning generation directly, as Stifel clean energy head Arash Nazhad put it, is a way for a company to protect itself from forces outside its control.

The other side of the market is scaling just as fast. Fervo, which went public this year, said in its third-quarter report that binding power purchase agreements now cover 1,054 megawatts, up from 658 megawatts in June, representing about 11.9 billion dollars of contracted revenue. The centerpiece is a 15-year, take-or-pay contract with Google for 396 megawatts, with first power expected in mid-2028 and an option to expand toward a full gigawatt. Fervo's first Cape Station geoblock entered commercial operation on September 30, ahead of its contractual deadline, after a 23-month build.

The design point matters for anyone selling into this market. Fervo builds in 33-to-50-megawatt blocks sized to match a typical data-center hall, so a hyperscaler can bring capacity online as each block finishes rather than waiting years for one giant plant. Installed cost has fallen from about 7,000 dollars per kilowatt in the first phase to a targeted 5,500 for the Google project, on a stated path toward 3,000. The contract is grid-connected, meaning Google can resell power it does not yet need, a structure that removes the risk of each party waiting on the other's construction schedule.

This procurement rush is colliding with public resistance. In last week's midterms, voters backed limits on AI data centers in 40 of 41 local ballot measures, and the governors elected in Texas and Georgia had campaigned on pauses. Goldman Sachs, meanwhile, raised its forecast of global data-center capacity in 2030 to 273 gigawatts, from 217 previously. Microsoft has pulled back from a data-center commitment over power supply, and an Oracle project in New Mexico has struggled with permitting. Anthropic pledged in February that AI companies should not leave American ratepayers to pick up the tab for their data centers, a promise that dedicated on-site generation is partly designed to keep.

For suppliers, this is the demand event of the decade. Turbine and generator makers, drilling contractors, engineering and construction firms, switchgear and substation suppliers now count AI labs and their infrastructure funds among their anchor customers. The buyer profile has changed in ways that favor vendors who can move quickly: the balance sheets are enormous, the timelines are desperate, and the technical requirements are largely standardized. Companies that once sold into utility procurement cycles measured in decades are being asked to quote against AI buildout cycles measured in quarters.

For companies that buy AI rather than build it, the lesson is that energy is now vendor risk. A model provider's power pipeline, whether it owns turbines, holds power purchase agreements, or waits in a utility queue, determines where its capacity lands and how reliable that capacity is. Procurement teams evaluating multi-year AI contracts should ask the same questions they would ask of a chip supplier: where does your electricity come from, who else has a claim on it, and what happens to my workloads if your project slips.

The gas turbine is a bridge, and phased firm power is the destination. Anthropic's generators solve a 2027 problem; Fervo's geoblocks are how the industry buys power when the politics of gas get harder, which last week's ballot results suggest they will. The AI labs that treat power as a core competency will outbuild the ones that treat it as procurement. The energy market has already noticed. So should everyone else who sells into it.

Disclaimers:
Some contents of this website are uploaded and reproduced by netizens spontaneously, which does not mean that this website agrees with its views;If it involves content, copyright and other issues, please contact within 30 days, and we will delete the content at the first time!

🔗 Shares 5 💬 Comment 0 💰 Award 0 📚 online submission
More>Related Comments
No comments yet. Why not share yours?