Two Masters, One Molecule

A war-driven LNG shock is colliding with the surge in gas-powered AI infrastructure.

Editor’s Note: This is the third article in a series analyzing the impact of Operation Epic Fury on global energy markets and the wider economic consequences that may follow. The first article can be read here, and the second here.

While the world watched oil cross $100 a barrel earlier this week, something quieter happened in the liquefied natural gas trade. American LNG exporters began rerouting cargoes toward Europe and Asia, where gas prices have surged 50% to 67% above year-ago levels. Reuters reported that U.S. exporters face domestic gas costs of roughly $3.63 per million BTUs against overseas prices above $12. At that spread, every available tanker load of American gas heads for the highest bidder.

The panic has a specific origin. On March 2, Iranian drone strikes hit QatarEnergy’s Ras Laffan complex—the largest LNG plant on earth—and the company halted all production. The International Energy Agency estimates that the Strait of Hormuz closure removed a fifth of global LNG supply from seaborne trade. Two days later, seven of the world’s largest technology companies signed a White House pledge to build their own gas-burning power plants to feed data centers, with OpenAI’s Stargate project ordering 29 gas turbines for a single campus in Abilene, Texas. Natural gas now serves two masters that have never competed for it at the same time: the countries that need it to survive and the companies that need it to build artificial intelligence.

For now, the U.S. looks insulated. The Wall Street Journal reported that domestic gas prices climbed just 11% last week, compared with 67% in Europe, thanks to ample storage, record production, and export capacity locked into long-term contracts. RBC’s Christopher Louney told the Journal there is only so far headline risks can pull U.S. prices. That assessment may be accurate this week. It is less accurate as a description of 2028.

U.S. LNG export capacity stood at 15.4 billion cubic feet per day at the end of 2025. The Energy Information Administration projects it will nearly double to roughly 30 billion cubic feet per day by 2030, with seven new terminals under construction. Each one is locked into long-term deals that guarantee those molecules leave the country. The contracts ensure the gas is shipped out, but not that enough remains at home.

Among the new capacity, Venture Global, which accounted for nearly two-thirds of global LNG capacity growth last year, told investors that more than 30% of its expected 2026 output is available for spot sales at whatever price the global market will bear. Domestic production, meanwhile, is growing at just 2% a year. The cushion that held last week depends on export capacity staying near its ceiling, and that ceiling is about to be raised.

The other claim on domestic gas barely existed two years ago. The Electric Power Research Institute assessed in February that data centers could consume up to 17% of U.S. electricity by 2030, up from roughly 5% today. PJM, the grid operator covering 13 states from New Jersey to Illinois, attributes nearly all of its projected load growth through 2030 to data centers. J.P. Morgan’s Michael Cembalest projects in his March energy paper that data centers will add 3 billion to 4 billion cubic feet per day of gas demand, and that three-quarters of operators planning to generate their own power intend to burn gas. The turbines to serve that demand are sold out through the end of the decade, with lead times stretching to seven or eight years. Solar cannot substitute at this scale: Cembalest calculates that a single large campus would need 40,000 acres of panels.

Export capacity doubling from one direction, data center demand arriving from the other, and the global market they connect to was already at its limit. Morgan Stanley estimated before the war that the entire global LNG market, roughly 420 million tons a year, carried a surplus of just 6 million tons. That is a cushion of 1.4%, and one facility outage erased it. Qatar’s energy minister told the Financial Times that restarting deliveries could take weeks or months. Morgan Stanley has pushed its forecast for Qatar’s next major expansion from 2026 to early 2027; any extension of the outage beyond a month, the bank’s analysts wrote, “quickly brings a deficit.”

The countries most exposed are the ones that matter most to the global technology supply chain. Taiwan imports 97% of its energy and holds just 11 days of LNG reserves; TSMC, which fabricates over 90% of the world’s most advanced semiconductors, consumes an estimated 5% to 10% of the island’s total energy. Japan has requested permission to draw on national oil reserves, and South Korea activated an emergency response team on March 3. Both countries hold only two to four weeks of LNG supply.

The war could end this month or grind on through the summer. The investment in gas-dependent AI infrastructure is a ten-year bet. Even if the Strait reopens next week, the turbine backlog does not shorten, the export terminals under construction do not stop, and the North Field expansion stays delayed until 2027. The relevant question is not how long this crisis lasts. It is how many more the system can absorb before the next one, given that the global surplus was 1.4% before the first bomb fell.

Most financing behind gas-dependent data center projects was arranged when gas cost $3.50 to $4.00 per million BTUs. A sustained increase of even $1 to $2 changes the return assumptions on billions in committed capital. Meta’s planned Hyperion facility in Louisiana calls for three on-site gas power plants at a reported $3 billion. OpenAI’s Stargate targets 10 gigawatts by 2029.

As The New Record covered on Sunday, the Iran war has already upended the administration’s macroeconomic framework by breaking its cheap-energy assumption. The deeper problem is what comes after. The world’s most expensive technology buildout is being constructed on a fuel source whose global margin of safety was thinner than a rounding error. AI timelines may not slip because of chips, or models, or regulation. They may slip because the gas ran short before the servers were ready.