US data centers will burn more natural gas than almost every country by 2035

US data centers will require 15 billion additional cubic feet of natural gas per day by 2035, according to a BloombergNEF forecast reported by Bloomberg News on September 14 — more than double BloombergNEF's own December projection of 6.9 billion cubic feet per day. The figure exceeds the total gas consumption of every country except China, Russia, Iran and the US, and BloombergNEF expects gas to supply 69% of power for newly built, grid-connected AI data centers through 2035. A separate August BloombergNEF analysis counted 99 gas plants in development across 22 states specifically to power data centers, backed by companies including Amazon, Microsoft, OpenAI and Anthropic, which at typical industry rates would emit about 318 million metric tons of carbon dioxide a year.

US data centers will need 15 billion additional cubic feet of natural gas a day by 2035, more than the entire gas consumption of every country on Earth except China, Russia, Iran and the US itself. Fifteen billion cubic feet a day. That's how much more natural gas American data centers will burn every day by 2035, according to a new forecast from BloombergNEF published this week, and it answers a question the industry has been dodging for years: how much natural gas will AI data centers actually use by 2035? A lot more than anyone assumed even nine months ago. BloombergNEF's own December forecast put that number at 6.9 billion cubic feet a day. Nine months later, it more than doubled. The revision, reported by Bloomberg News on September 14, reflects how fast AI compute demand has outrun every previous estimate of how much power it would need. Put the number in context. It gets uncomfortable fast. Fifteen billion cubic feet a day is more gas than Germany's entire economy burns in a year, roughly 9 billion cubic feet a day across every furnace, factory and power plant in the country. It's more than the total gas consumption of every nation on Earth except three: China, Russia and Iran. Only the US itself burns more. One industry, in one country, is about to out-consume entire nations. That industry is AI. BloombergNEF just revised its US data center power forecast 83% higher in seven months https://startupfortune.com/bloombergnef-just-revised-its-us-data-center-power-forecast-83-higher-in-seven-months/ BloombergNEF's July 21 report projects US data centers will consume 194 gigawatts of electricity by 2035, one-fifth of all US power, revising its December forecast upward by 83% in seven months. Nearly half of that capacity will be devoted to AI, reshaping the investment case for nuclear energy, grid infrastructure, and compute-adjacent startups. - US data center power consumption forecast https://startupfortune.com/bloombergnef-just-revised-its-us-data-center-power-forecast-83-higher-in-seven-months/ - AI energy demand grid infrastructure impact https://startupfortune.com/bloombergnef-just-revised-its-us-data-center-power-forecast-83-higher-in-seven-months/ BloombergNEF expects gas to supply 69% of the power for newly built, grid-connected AI data centers through 2035. The reason is simple economics, not ideology. Gas is cheap in the US and plentiful after a decade of shale drilling. And turbines can ramp up and down within minutes. Solar and wind can't do that on demand. New nuclear plants take a decade or more to permit and build. A data center running AI training and inference around the clock needs power that shows up exactly when it's needed. Not when the sun happens to be out. The power sector is now the second largest driver of US gas demand through 2035, according to BloombergNEF. Only the new liquefied natural gas export terminals coming online on the Gulf Coast rank higher. Total power-sector gas consumption is projected to hit 54 billion cubic feet a day by 2035. That's up 18 billion cubic feet a day from 2025. A separate BloombergNEF analysis published in August found 99 gas plants in development specifically to power data centers, spread across 22 states from Alaska to Georgia. More than a third sit in Texas. Oil and gas infrastructure is already in place there, and permitting moves fast. The companies backing these projects range from Amazon and Microsoft to OpenAI and Anthropic. Run at typical industry rates, those 99 plants alone would emit about 318 million metric tons of carbon dioxide a year: roughly 20% of what the entire US power sector emitted last year. Run flat out to meet AI demand around the clock, and that share climbs toward a third. Many of these plants are what the industry calls behind the meter. They sit on or next to a data center campus and generate power directly for it. That means a company can start construction without waiting on a utility or a grid operator like PJM to approve an interconnection. Speed is the whole point. The climate pledges nobody wants to revisit Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed what the White House is calling the Ratepayer Protection Pledge earlier this year, promising to build and buy all the new power their data centers need - or fund it themselves - so household electric bills don't spike. Nothing in that pledge requires the power to be clean. As Forbes reported in July, Meta has already pulled back from a clean energy agreement it once championed and leaned into natural gas capacity instead. Amazon trucked in semitrailer-mounted gas generators to power a hybrid data center and turbine facility in Memphis - and in West Texas, gas turbines are already running Stargate, the OpenAI-backed compute buildout, ahead of anything the regional grid can deliver. Every major AI company still lists a net-zero target on its website, most aimed at 2030 or 2040. BloombergNEF's own numbers point the other way. And through at least 2035, gas wins. By a wide margin. AI Agents Lied, Stole and Killed Each Other in a Weeks-Long Simulation https://startupfortune.com/ai-agents-lied-stole-and-killed-each-other-in-a-weeks-long-simulation/ Emergence AI ran autonomous agents inside simulated worlds for up to sixteen days and watched them lie, steal, and kill off rivals, with one mixed-model world logging 352 crimes and seven deaths. A separate DeepMind experiment saw 100 math agents split into cheaters, converts, and whistleblowers within half an hour. - ai agents behavior emergent moral ethics simulation https://startupfortune.com/ai-agents-lied-stole-and-killed-each-other-in-a-weeks-long-simulation/ - long term ai agent interaction virtual worlds study https://startupfortune.com/ai-agents-lied-stole-and-killed-each-other-in-a-weeks-long-simulation/ Also read: AI Agents Lied, Stole and Killed Each Other in a Weeks-Long Simulation https://startupfortune.com/ai-agents-lied-stole-and-killed-each-other-in-a-weeks-long-simulation/ • Kimi K3 Closes the AI Capability Gap to Four Months at a Fifth of the Price https://startupfortune.com/kimi-k3-closes-the-ai-capability-gap-to-four-months-at-a-fifth-of-the-price/ • AIUC Wants To Insure Your AI Agents Before They Go Rogue https://startupfortune.com/aiuc-wants-to-insure-your-ai-agents-before-they-go-rogue/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.