Everything About AI
NextFin News - Anthropic has agreed to pay $3.4 billion for gas-fired generators to power its artificial-intelligence data centres, a move that puts one of the world's most valuable private AI companies directly into the business of producing its own electricity. The company secured commitments for 1.7 gigawatts of heavy-duty generators from OnePWR Solutions, a San Antonio-based equipment supplier, in late September, according to people familiar with the matter. Anthropic did not buy OnePWR itself, and none of the supplier's employees transferred to the AI company. The arrangement marks what is believed to be the largest known single purchase of power-generation equipment by an AI laboratory.
The deal is a signal that the AI industry's power problem has moved from a procurement headache to a balance-sheet imperative. Anthropic, which disclosed nearly $900 billion in computing commitments through 2045 in its draft IPO filing, had initially planned to buy electricity from OnePWR. Instead it took over the supplier's existing generator orders from Caterpillar, which OnePWR had placed about two years earlier but lacked the financing to complete. By stepping in, Anthropic is buying time in a market where the queue to connect a new data centre to the grid can stretch to seven years.
The Deal: Buying Generators, Not a Utility
The structure of the transaction matters as much as the size. Anthropic did not acquire OnePWR Solutions and did not inherit its workforce. What it bought was the equipment pipeline — 1.7 GW of heavy-duty generators plus the engineering and construction capability needed to install them. OnePWR had ordered the generators from Caterpillar roughly two years ago; the supplier needed financing to close the purchase, and Anthropic provided the exit by taking over the orders directly.
This is a purchase of scarce industrial hardware, not a power purchase agreement. Anthropic now owns the right to the machines and the team that will install them, but it still has to secure the fuel, the site and the permits. That split is telling: in a market where turbines are backordered into the 2030s, controlling the equipment is the binding constraint. The electricity itself can be bought on the spot market or through a fuel contract; the generator cannot be bought at all if the order book is full.
OnePWR is a specialist in exactly this kind of behind-the-meter power. In a February 2026 announcement, the company said it held more than 400 megawatts of power-generation equipment in inventory ready for installation, with more than 800 MW of additional capacity scheduled for delivery over the following 24 months. Its chief executive, David Coker, framed the value proposition in a way that now reads as a description of Anthropic's strategy:
"Access to equipment alone doesn't solve today's power constraints – execution does. By combining deploy-ready generation with an integrated engineering, procurement and construction team, we reduce execution risk and deliver power when and where customers need it."
The company's commercial chief, Antti Ekqvist, went further: "An increasing share of planned capacity is bypassing the grid altogether through behind-the-meter solutions. Power demand across U.S. industrial and commercial sectors continues to climb, and traditional infrastructure is struggling to keep pace."
For Anthropic, the purchase is a hedge against the single biggest bottleneck in AI expansion: power delivery. The company has committed to roughly $900 billion of computing spend through 2045, most of it locked into contracts with cloud providers and chip suppliers. Its revised IPO filing puts total purchase and lease commitments at $895.4 billion, of which about $448 billion is expected to be incurred before the end of 2030 and roughly $70 billion falls due before September 2027. In the third quarter it reported revenue of $17 billion, up more than 12 times from the same quarter a year earlier, alongside a net loss of about $23.3 billion. That loss is not a sign of weakness so much as a map of the capital intensity ahead — every dollar of revenue requires several more dollars of infrastructure still being built.
Why Gas, and Why Now
Natural gas is the fastest available bridge between today's AI demand and the grid of the future. Renewables are cheaper per kilowatt-hour, but they are intermittent; a data centre running AI training or inference around the clock needs firm, dispatchable power that can ramp with workload spikes. Combined-cycle gas turbines can reach more than 60% efficiency and, in combined heat-and-power configurations, above 80%. They can be sited next to the data centre, avoiding transmission losses and interconnection queues that in many US markets now run four to seven years.
The economics have deteriorated sharply for anyone waiting. The cost to build a gas-fired power plant has risen 66% over the past two years, according to analysis from an energy-research firm. Gas turbines alone — which make up as much as 30% of a new plant's cost — are expected to be 195% more expensive by the end of this year than in 2019, and manufacturing constraints mean waitlists now stretch into the early 2030s. In that environment, the rational move is to lock in equipment today even at inflated prices, because tomorrow's price is likely higher and tomorrow's delivery slot may not exist. Anthropic's $3.4 billion is, in effect, a premium paid to jump the queue.
Anthropic is not alone. Hyperscalers across the industry have moved to own or contract dedicated generation rather than rely on the grid. Microsoft signed a 20-year power agreement with Chevron for Project Kilby, a natural-gas plant in West Texas expected to deliver approximately 2.67 GW of capacity near Pecos in Reeves County, with a total capital outlay estimated between $7 billion and $9 billion. Amazon expanded its nuclear arrangement with Talen Energy to receive up to 1,920 MW of carbon-free power from the Susquehanna plant through 2042, though additional phases remain under regulatory review. Meta is bringing two 200 MW off-grid natural-gas plants online by the end of 2026 to serve its Socrates data-centre campus in New Albany, Ohio. Anthropic separately agreed to purchase about 470 MW of natural-gas generators from Houston-based ERock. The pattern is consistent: the largest AI companies are no longer willing to outsource power risk to utilities.
The scale of the underlying demand explains the urgency. The International Energy Agency estimates global data-centre electricity consumption could exceed 1,000 terawatt-hours a year by 2026, roughly equal to Japan's total electricity use, up from 460 TWh in 2022. In the United States, data-centre power demand is forecast to climb from 31 gigawatts in 2025 to 41 GW in 2026 and 66 GW in 2027, according to commodities researchers at Goldman Sachs, with capacity utilisation assumptions implying roughly 95 GW of data-centre capacity by the end of 2027. Hyperscaler capital expenditure is forecast to exceed $600 billion in 2026, a further 36% increase over 2025, with roughly three-quarters tied to AI infrastructure. These are not incremental loads; they are new cities appearing on the grid.
The Second-Order Squeeze: Gas Prices, Pipelines and the Grid
Owning the generator solves only the first-order problem. The second-order constraint is the fuel and the pipe. Data centres are already large enough to move regional gas markets. During Winter Storm Fern in January 2026, US natural-gas demand hit a record seven-day rolling average of 165.6 billion cubic feet per day, and the Henry Hub spot price surged to an all-time high of $30.72 per million British thermal units on January 23, before retreating to $4.40 within days. A 20% rise in national demand produced a fourfold price move — and that was before the AI build-out has fully arrived.
The interconnection queue tells the same story from the other side. At the start of 2025, about 23.8 GW of natural-gas capacity was planned to come online between 2026 and 2030. By January 2026 that pipeline had grown by roughly 24 GW, with 16 GW added in the first quarter alone — a total of about 64 GW of gas projects queued through 2030. In ERCOT, Texas's grid operator, interconnection requests from data centres reached about 356 GW as of March 2026. Natural-gas demand from data centres in the US may reach 6.1 billion cubic feet per day by 2030, potentially a nearly 20% increase in annual average power burn. Behind-the-meter generation is the escape valve from a queue that would otherwise stall AI deployment for years.
This creates a new kind of industrial geography. AI infrastructure will cluster where gas is abundant and pipelines exist — Texas, Pennsylvania, Ohio, the Gulf Coast — rather than where fibre and cheap labour historically drew data centres. Power availability, not talent density, is becoming the siting decision. The states that win the next wave of AI investment will be the ones with pipeline capacity and permissive siting rules, not necessarily the ones with the deepest tech talent pools.
The Counter-Thesis: Gas Is a Transition Trap
The strongest argument against this strategy is that it locks AI companies into a fossil-fuel path that conflicts with their own climate commitments and could strand assets before they pay off. Microsoft has pledged to be carbon-negative by 2030 and has specified hydrogen-ready turbines in new orders, betting that gas infrastructure can transition to green hydrogen by 2028-2030. Amazon and Google face the same tension: their customers and regulators expect decarbonisation, yet their AI growth requires firm power that renewables alone cannot yet guarantee.
There is also a financial risk in paying today's inflated prices. Gas-turbine prices up 195% from 2019 are, by definition, a cyclical peak driven by a scramble for scarce equipment. If manufacturing capacity expands and the queue clears — or if nuclear, geothermal or long-duration storage becomes commercially deployable at scale — the generators Anthropic is buying at 2026 prices could look expensive in hindsight. A buyer paying peak-cycle prices for equipment with a 20- to 30-year life is making a bet that demand stays ahead of supply for most of that period.
Regulatory risk cuts the same way. Amazon's expanded nuclear arrangement with Talen requires approval of additional phases by the Federal Energy Regulatory Commission; similar scrutiny could follow for gas-fired projects that bypass grid oversight entirely. Several states have already moved to restrict new fossil-fuel generation for data centres, and a political shift could turn today's permitted projects into tomorrow's stranded assets. The behind-the-meter model is attractive precisely because it avoids the grid — but it also avoids the regulatory compact that has historically guaranteed cost recovery for utility-scale investment.
The answer is that Anthropic is not betting on gas as a permanent fuel; it is buying optionality on time. The generators are modular and can be repowered, blended with hydrogen, or retired early if a cheaper firm-power source materialises. The alternative — waiting for the grid — has no fallback: a data centre with no power generates zero revenue. In a market where a single large AI training run can be worth hundreds of millions of dollars, the option value of guaranteed power dwarfs the risk of overpaying for hardware.
Who Benefits and Who Is Exposed
The immediate beneficiaries are the equipment and fuel supply chain. Caterpillar, as the original manufacturer of the generators, stands to see order books extend into the next decade. OnePWR Solutions demonstrated that there is a market for intermediaries who can hold inventory and de-risk execution for hyperscale buyers. Natural-gas producers and midstream pipeline operators gain a new, large, relatively inelastic customer base. Utilities that can offer firm power at scale retain relevance; those that cannot risk losing their largest industrial customers to behind-the-meter competition.
The exposed are the laggards. AI companies without secured power face higher costs and slower deployment. Regions with gas constraints or moratoria on new fossil generation will see AI investment flow elsewhere. And ratepayers in markets where data centres have already signed long-term contracts may find themselves subsidising grid capacity that the largest users no longer rely on — the fixed costs of the grid remain, but the customers best able to pay them are increasingly generating their own power.
What to Watch
Three signals will test whether this is a durable shift or a cyclical scramble. First, gas-turbine lead times and prices: if waitlists shorten and turbine prices fall back toward historical norms, the urgency premium in these deals was cyclical. Second, interconnection queue times: if grid upgrades accelerate and queues compress below two years, behind-the-meter generation loses its main advantage. Third, the fuel-price pass-through: if Henry Hub repeatedly spikes above $15-20 per mmBtu during peak periods, the economics of gas-fired data centres will face their first real stress test.
The falsifying signal for the structural thesis is specific: if, by 2028, more than half of new hyperscale data-centre capacity is connected to the grid rather than built behind the meter, and gas-turbine lead times have fallen below two years, then the 2026 generator rush was a temporary equipment shortage, not a regime change in how AI companies procure power.
Bottom line: Anthropic's $3.4 billion generator purchase is not really about gas. It is about converting an uncertain grid connection into an owned, dispatchable asset — and every AI lab with a power problem will be asked whether it can afford not to do the same.
Explore more exclusive insights at nextfin.ai.
