BNEF: US Data Centers Could Burn More Gas Than Germany and Japan Combined by 2035

BloombergNEF's New Energy Outlook 2026, released September 14, 2026, reclassifies US AI data centers from 'large power customer' to 'sovereign-scale fuel buyer.' The updated forecast calls for grid-connected US data centers to add 15 billion cubic feet per day of natural-gas demand by 2035 — more than the entire current gas consumption of every country on earth except China, Russia, Iran and the United States itself.

Add the 2.9–3.4 Bcf/d of behind-the-meter gas plants that hyperscalers are now siting directly next to their server halls, and US data centers will burn roughly 18 Bcf/d by 2035 — enough to rank them as the world's fifth-largest gas consumer, ahead of every G7 nation other than the US, Canada and the UK. That single number, more than any ESG disclosure, is what rewrites the AI-power narrative.

The data-center gas surge is not a solar-versus-gas argument lost on price. It is a grid-mechanics and lead-time argument won on speed. Three structural forces lock gas in as the marginal fuel for AI compute through at least 2032.

The most under-appreciated data point is the structural mix shift in planned US generation. Per EIA planned-projects data cited by American Action Forum, the gas share of announced US generation capacity rose from 11.1% in 2024 to 18.1% in 2026, while solar's share fell from 51% to 45.6% and wind's share halved from 21.4% to 10.3%. That is not a one-year wobble; it is the mix pivoting away from clean.

The 15 Bcf/d increment does not flow equally. It splits three ways: turbine OEMs capture the equipment rent, midstream pipelines capture the molecule-transport rent, and IPPs capture the generation margin. The supply chain is short enough that pricing power is concentrating fast.

GE Vernova is the cleanest pure-play on the thesis. Its Q2 2026 press release shows Power-segment orders up 134% year-on-year and equipment backlog up $16.6 billion (69% YoY) to $40.6 billion. The 116 GW of gas-turbine backlog and slot reservations implies roughly five years of revenue visibility at current run-rate, and management is guiding toward $200 billion of total backlog by 2028 (Q1 2026 earnings call). At a trailing P/E of 25x and forward P/E near 39x, the multiple already discounts much of the order book — but free cash flow conversion is what matters here, not earnings.

Siemens Energy is the parallel play in Europe and the Middle East. Q3 FY2026 orders hit a record €17.9 billion, with Gas Services alone booking €10.0 billion (+61.9% YoY) on €3.76 billion of revenue — a 2.65x book-to-bill ratio that locks in years of growth. The €73 billion Gas Services backlog now exceeds total company revenue at peak. Its Siemens-brand US service network also captures the aftermarket annuity on every turbine GE Vernova ships.

On the molecule side, Williams Companies is investing $5.1 billion in modular gas-fired and hybrid power plants co-located with data centers, alongside Transco pipeline expansion. EQT has signed multi-decade supply deals to feed marquee gas-fired data-center projects in Appalachia. Both are leveraged to volume, not price — which matters because the BNEF base case assumes wellhead gas stays cheap enough to make coal-to-gas switching uneconomic in reverse.

Independent power producers split cleanly. Vistra is the gas-and-nuclear hybrid: 38.7 GW of generation capacity plus 2,600 MW of 20-year zero-carbon PPAs it signed with Meta in January 2026. Talen owns the Susquehanna nuclear plant at the heart of the Amazon Web Services co-location fight (a $18 billion, 17-year, 1,920 MW PPA signed June 2025), but its leveraged balance sheet (debt-to-capital 86%) makes it the volatile name on the list. Constellation Energy sits furthest from the gas story and therefore captures the nuclear-purity premium — 32.4 GW of mostly carbon-free capacity — but pays for that purity with the highest P/E in the group on forward earnings (forward P/E 21x).

Through 2024, the dominant equity narrative held that AI compute would be net-positive for the climate because hyperscalers had signed 24/7 carbon-free-energy (CFE) pledges by 2030, and because solar PPAs were dropping below $30/MWh. BloombergNEF's September 14, 2026 update breaks both legs of that thesis at once. Carbon-free supply at gigawatt scale is not arriving fast enough, and hyperscalers themselves are now signing the gas PPAs their 2030 pledges were supposed to preclude.

The escape hatch the four largest US hyperscalers are reaching for is carbon capture and storage (CCS). A July 2026 Rice University study argued that CCS is now the only credible bridge between the 15 Bcf/d demand surge and the 2030 pledges. But BloombergNEF's NEO 2026 implies roughly 1 million metric tons per day of additional greenhouse-gas pollution from the grid-connected increment alone (using IEA emission factors). At ~365 million tons/year, that is more than the entire annual emissions of Argentina. No CCS project pipeline in the world is sized to absorb that.

Over the next two quarters, the catalysts are concrete and dated. Long-term, the risks are structural but softer.