Interactive: DeepSeek AI bursts the bubble for US gas-fired power growth

On a chilly day in late January, obscure Chinese artificial intelligence company DeepSeek AI put the US natural gas market on its heels, breaking news that questioned the industry’s almost gospel-like narrative that AI-related power demand will soon fuel a historic rise in US gas-fired power burn.

The news that DeepSeek had created a large language model, roughly equivalent to ChatGPT, at just one-tenth of the cost and a fraction of the computing power sent shale gas and independent power producers’ stock prices tumbling and helped to propel a selloff in the NYMEX gas futures market.

Since the release of ChatGPT in late November 2022, electric utilities, market analysts and even natural gas producers have been projecting ever-larger US power grid load growth tied to the buildout of datacenters to support artificial intelligence. By late 2024, US utilities were projecting datacenter electricity demand to reach 900 TWh by 2035 – up from an estimated 185 TWh in 2023. For shale gas producers, the rapid expansion of US electricity demand would mean dramatic and perhaps unprecedented growth in gas-fired power generation.

Electric utilities and regulators should "certainly" be revisiting their load forecasts, Alp Kucukelbir, chief scientist at AI software company Fero Labs Inc., said in late January. "For the purposes of training these types of large language models, we now have evidence that that could take, indeed, much less energy and much less money," he said.

According to an analysis published in fourth quarter 2024 by S&P Global Commodity Insights, Kucukelbir could be right.

Historically, power demand forecasts have overestimated growth, largely because they didn’t account for improvements in energy efficiency -- like the ones achieved by DeepSeek AI. Even the US Energy Information Administration has overestimated electricity demand growth in the past. From 2017-2023, the agency consistently predicted larger-than-realized growth in electricity sales.

For major datacenter developers like Amazon, Alphabet, Microsoft and others, there is a strong incentive to improve computing, cooling and power distribution efficiency – not just to lower costs, but also to reduce the environmental impacts. Considering also the possibility that grid-connection queues could delay growth in new datacenter power loads, Commodity Insights is forecasting much slower growth than US utilities have proposed.

Under the CI planning scenario, datacenter demand would grow by just 280 TWh by 2035, adding about 5.7 Bcf/d in incremental gas burn if 100% of the additional electric load were supplied by gas. Assuming wind and solar power supply at least some of the additional load, the bottom-line impact for gas would be even smaller. If the aggregate utility forecast is accurate and the projected 455 TWh of datacenter demand growth by 2035 is supplied 100% by natural gas, demand for gas would increase by just over 12 Bcf/d – just a fraction of the growth expected from LNG export demand over the next decade.

According to the Commodity Insights analysis, most of the US datacenter-related gas demand growth will come before the end of this decade. Beyond 2035, growth would be marginal.

Although short-term demand growth for gas-fired power generation may be marginal compared to utilities’ projections, market impacts could be significant – especially considering the concentration of datacenters planned in Northern Virginia and the Columbus, Ohio area.

By later this decade, gas prices in both regions could see increased volatility. Located in Transco Zone 5, Northern Virginia’s gas market is already exposed to seasonal price spikes during high demand periods – a trend that could be exacerbated over time. Similar price spikes could become more common around the Columbus, Ohio area, too – potentially affecting hubs like Tennessee Zone 4 Ohio and Texas Eastern M2 receipts.