Wednesday, November 18, 2026

INFORMOPEDIA

World news and deep-dive explainers — climate, economy, technology, science, culture, health

Breaking
Technology

Why the AI Industry Is Buying Gas Power Plants — and What It Gives Up in Return

Informopedia Explains: the company behind Claude, one of the world’s most widely used AI systems, has just bought itself a power station. Not a share in one, not a contract to buy electricity from one — the physical machinery to generate its own power, fuelled by natural gas.

Anthropic has signed a $3.4 billion contract with OnePWR Solutions, a Texas-based energy company, for gas-fired engines to supply its AI data centres with off-grid electricity. The Financial Times, which broke the news, called it the largest known single purchase of power-generation equipment by an AI lab. It is not an isolated deal. Across the industry, the companies building the largest AI systems are quietly becoming energy companies too — and the fuel they are choosing, more often than not, is gas.

That choice sits oddly with everything these companies have said about the climate. So why are they doing it?

The grid cannot keep up

Start with the scale of the demand. A single frontier AI data centre can now consume as much electricity as a small city, and the industry’s appetite is still growing. Data centre power demand is set to double by 2030, according to figures cited at this month’s COP31 climate talks. In many places, the grid simply cannot deliver that much power, that fast.

Connecting a large new data centre to the grid can take years. The queue of projects waiting for a connection is measured in gigawatts, and in some regions the wait stretches beyond the end of the decade. For an AI lab racing to train the next model, years is not a timeline — it is a defeat. A gas plant, by contrast, can be ordered, delivered and running in months. It sits next to the data centre, it answers to no queue, and it produces power on demand.

There is also the question of reliability. Solar and wind are now the cheapest sources of electricity in most of the world, but they produce power when the weather allows, not when a training run needs it. A data centre training a frontier model cannot pause for a cloudy afternoon; it needs steady, uninterrupted power, twenty-four hours a day, for weeks at a time. Batteries can smooth out hours. Gas can run indefinitely. Nuclear can too, but a new reactor takes a decade or more to build, and the AI industry is not prepared to wait.

What gas actually costs the climate

Gas is cleaner than coal, and that comparison is doing a lot of work in the industry’s favour. A gas plant emits roughly 400 to 500 grams of carbon dioxide per kilowatt-hour — about half of coal’s output, but far above wind, solar, nuclear or hydro, which typically come in under 50 grams on a lifecycle basis. Gas is a cleaner fossil fuel. It is not a clean one.

And the carbon dioxide is only part of the story. Natural gas is mostly methane, and methane leaks at every stage — from the wellhead, through pipelines, to the plant itself. Methane traps far more heat than carbon dioxide over the short term, so a leaky gas supply chain can erase much of the advantage over coal. The climate impact of a gas plant depends on what it replaces and on how tightly those leaks are controlled — two things a buyer of gas engines has limited power to guarantee.

The scale of what is being built makes the arithmetic stark. Last month, the environmental nonprofit Sierra Club estimated that upcoming AI data centres could release 2.1 billion metric tons of carbon dioxide by 2030 — enough, it said, to wipe out nearly 15 years’ worth of greenhouse gas reduction efforts. That figure is a projection, not a certainty, and it depends on how many of the announced data centres actually get built. But it captures the direction of travel: the industry that promises to solve problems with intelligence is, for now, solving its own energy problem with combustion.

Why the money works

None of this is happening because AI companies cannot afford clean power. Anthropic reported an $81 billion annualised revenue run rate for October, and its September-quarter compute and infrastructure costs reached $9.9 billion — 58% of its revenue. The company is about to list on the Nasdaq at a valuation of up to $1.4 trillion. A $3.4 billion gas order is, in that context, a rounding error. The constraint is not money. It is time.

That is the uncomfortable core of the story. The AI industry is not choosing gas because it is cheap or because it is clean. It is choosing gas because it is fast, and because the race to build the biggest models rewards speed above almost everything else. Renewable power plus storage plus grid upgrades would be cleaner, and in the long run cheaper — but it would also be slower, and slower loses the race.

What happens next

The gas plants being ordered now will run for decades. That is the lock-in problem: a power plant built in 2026 is still generating electricity — and carbon dioxide — in 2050, the year by which most of these same companies have promised to reach net zero. The industry’s answer, when it gives one, is that the plants can later be converted to cleaner fuels, or that their emissions will be offset. Both claims are easier to make than to deliver.

There is a counter-argument worth taking seriously. AI data centres are also, in principle, the most flexible large electricity consumers ever built: a training run can, in theory, be scheduled for when the wind is blowing and paused when it is not. Some labs are experimenting with exactly that. But the economics of a $3.4 billion gas order point the other way — towards power that is always on, whatever the weather, whatever the carbon cost.

The question the gas purchases raise is not really about engineering. It is about priorities. The AI industry spent years presenting itself as a clean technology — weightless, digital, green. The power plants tell a different story: that the intelligence revolution runs on infrastructure, that infrastructure runs on energy, and that when the energy is not available cleanly and quickly, the industry will take it dirty and fast. Whether that trade-off was necessary, or merely convenient, is the argument the next decade will settle.

Related: COP31 is stuck on the one word that matters — fossil fuels — and the invisible climate leak in every air conditioner. For the hardware behind the boom, see what a GPU is — and why one company’s chips became the most important hardware on earth.

Source: International Energy Agency — Data Centres and Data Transmission Networks

Avatar photo
Informopedia News Desk

Join the discussion

Your email address will not be published. Required fields are marked *