AI could burn through a ‘mountain’ of gas as US data centres outconsume most countries

As the generative artificial intelligence (GenAI) wave sweeps across the world, the focus of competition among technology giants is quietly shifting — from the initial “chip race” to a contest for a more basic but indispensable resource: energy.

In recent quarterly results briefings and investor meetings, Amazon, Microsoft and Google have repeatedly highlighted a common problem. The constraint is no longer the shortage of GPUs and other core computing components, but access to electricity. A new report by BloombergNEF (BNEF) says large cloud data centres are increasingly bypassing the public grid and turning directly to natural gas-fired generation, as the US transmission network faces years of construction delays and approval bottlenecks.

By 2035, US data centres whose demand is directly driven by the AI industry are expected to increase their natural gas consumption by 15 billion cubic feet a day, according to the report. That increase alone would exceed the current total gas consumption of any country other than China, Russia, Iran and the US. Henry Eaton, BNEF’s gas market analyst and lead author of the report, said the estimate carried “a considerable” margin of error, with risks on both the upside and downside that could not be ignored.

Bypassing grid congestion with behind-the-meter generation

The shift reflects the physical limits of North American infrastructure. For decades, major cloud hubs in the eastern US, including Northern Virginia, have relied on the national grid. But strict environmental reviews and the high cost of upgrading ageing lines have significantly extended waiting times for data centres seeking high-voltage connections. Industry tracking data shows that the average wait in Northern Virginia, the world’s largest single data-centre market, has now risen to as long as seven years. The US median is approaching five years, well above the eight-to-11-month target set by the Federal Energy Regulatory Commission (FERC).

For data-centre developers routinely investing tens of billions of US dollars, waiting has become the most expensive hidden cost. Missing a market window could leave tens of billions of dollars tied up in idle investment. An alternative known as “behind-the-meter” generation has therefore rapidly gained popularity: building dedicated gas-fired power plants near data centres or laying gas pipelines directly to them. The model allows projects to be completed in months rather than years, while providing the round-the-clock, uninterrupted power supply most prized by technology giants.

The BNEF analysts said this infrastructure shift, driven by demand for computing capacity, was structural. Software and internet companies that once focused only on computing and data flows are being forced to take on the role of asset-heavy energy wholesalers, dealing not only in advanced chips but also in land surveys, drilling-rig leases and the operation of heavy-duty gas turbines.

What 15 billion cubic feet means for the resource race

The scale of the trend can be understood from the report’s most significant figures.

Total natural gas consumption by the US power sector is forecast to rise from an average of 36 billion cubic feet a day in 2025 to 54 billion cubic feet in 2035 — an increase of about 50%, or 18 billion cubic feet a day. Of that, the additional direct gas consumption from data centres alone is expected to reach 15 billion cubic feet a day, more than double BNEF’s forecast of 6.9 billion cubic feet a day in December last year. Natural gas is expected to account for 69% of the power required by new grid-connected facilities.

The physical consequence is clear: the mismatch between supply and demand is worsening. Although the US has become one of the world’s largest LNG producers following the shale-gas revolution, the ability of upstream oil and gas fields to increase output is not limitless. Producers also face bottlenecks in the supply of drilling equipment and a shortage of skilled workers. The report says US producers currently expect to increase daily output by 35 billion cubic feet between 2025 and 2035. Meeting the forecast demand would require a further increase of 11 billion cubic feet a day.

LNG exporters and data centres competing for the same gas

The challenge is compounded by the fact that the US natural gas industry is entering a period of maximum expansion on two fronts. It must meet rapidly rising domestic demand driven by the AI boom, while LNG export terminals built with heavy investment over the past decade continue supplying higher-priced gas to European and Asian markets. The demand increase linked to those exports is expected to reach 21 billion cubic feet a day — even more than the increase from data centres.

This is the “cannibalisation” scenario that many market participants fear. As data centres reserve large volumes of pipeline gas to meet construction deadlines, they will compete with international buyers seeking long-term supply contracts for limited production at the wellhead. This cross-industry competition for resources could place sustained structural upward pressure on pipeline gas prices paid by US households and businesses over the medium to long term.

Natural gas was once viewed as a relatively peripheral, cyclical commodity whose demand was driven mainly by winter heating. But as the AI industry turns it into a strategic resource essential to keeping computing capacity running, the cost structure of future large AI models could become increasingly exposed to fluctuations in spot gas prices.

Who stands to benefit from the new contest?

Not every part of the industry will come under pressure as the shift moves from “code” to “gas”. Capital markets are already identifying the sectors likely to benefit:

Upstream energy equipment and engineering services providers: These include heavy-duty gas turbine manufacturers such as GE Vernova, the former General Electric power business spun off and separately listed in 2024, and Mitsubishi Power. Orders for gas turbines from both companies are reportedly scheduled through 2029 to 2031, with capacity remaining in short supply. If data-centre construction cycles are shortened and accelerated, such hardware suppliers will benefit directly from the resulting orders.

Midstream pipeline transport operators: Unlike pure drilling-rig companies, which are more exposed to swings in profitability, interstate pipeline operators such as Enterprise Products Partners often control asset-heavy infrastructure with near-monopoly characteristics. They can collect stable transportation and distribution fees, making them the unseen rent collectors behind the computing infrastructure build-out.

Technology giants able to offer integrated energy solutions: Cloud platforms with strong cash flow and the ability to build small modular reactors (SMRs) or arrange bundled purchases from gas-fired power plants will have an early advantage in the future competition to build infrastructure moats. In the era of computing capacity, the companies that can guarantee the cheapest uninterrupted energy will be best placed to win the race to develop the next generation of models.