US data centers could consume more natural gas than Germany and Japan combined by 2035

BloombergNEF projects U.S. data centers will consume about 18 Bcf/d of natural gas by 2035—nearly double its forecast from nine months ago and more than Germany and Japan combined. Grid-connected facilities are expected to add 15 Bcf/d of power-sector gas demand, five times the growth projected from all other grid-connected sectors, while on-site generation could add another 2.9-3.4 Bcf/d. The AI-driven demand surge, alongside expanding LNG exports, could push gas prices materially higher and raise utility costs; associated incremental emissions are estimated at 1 million metric tons of greenhouse gases per day, roughly 12% of current U.S. emissions.

Analysis

The underappreciated transmission mechanism is not merely higher Henry Hub pricing; it is the widening basis and capacity value in power-constrained load pockets. Incremental AI load is geographically concentrated, so Appalachian producers with takeaway access (EQT, RRC, AR) and Gulf Coast LNG-linked gas exposure should outperform broad gas beta if local generation demand tightens. Midstream operators with contracted gathering, processing and pipeline assets—WMB, KMI and TRGP—offer a lower-volatility way to monetize volume growth, while merchant generators with gas-heavy fleets near data-center clusters (VST, CEG, NRG) gain scarcity pricing but face regulatory scrutiny over retail-rate pass-through.

For hyperscalers, energy cost is unlikely to impair near-term AI capex, but it can change the economics at the margin: sustained power and gas inflation raises inference cost, delays data-center utilization ramps, and makes long-duration AI revenue assumptions less valuable. MSFT and AMZN have the greatest cloud-margin sensitivity because they sell compute externally; META and GOOG can absorb more cost internally but face a more direct emissions and permitting narrative risk. The 1-3 month catalyst is utility interconnection and generation announcements; the 6-18 month catalyst is whether LNG project commissioning and power demand tighten the gas balance faster than associated-gas supply responds.

Consensus may be extrapolating announced load into a linear gas-demand outcome. The key falsifier is not project announcements but executed power contracts, turbine delivery, transmission upgrades and actual server deployment. A recessionary slowdown in AI monetization, faster renewable-plus-storage buildout, nuclear restarts, or weaker LNG utilization would reduce the call on gas; conversely, Henry Hub above roughly $5/MMBtu would begin to pressure data-center economics and invite political intervention on utility bills and LNG exports.

Key Decisions for Investors

• Build a 6-12 month long EQT / short MSFT relative-value position: EQT offers direct upside to a structurally tighter U.S. gas balance, while MSFT is the hyperscaler most exposed to externally sold compute-margin compression. Size for a 15-20% gas-price move; exit if Henry Hub remains below $3.25/MMBtu through winter or Azure growth accelerates without margin pressure.
• Prefer long WMB and TRGP over an outright UNG position for 12-18 months. Contracted infrastructure cash flows monetize higher throughput with less commodity volatility; the principal risk is LNG delays or data-center projects failing to reach financial close, which would cap volume-growth expectations.
• Maintain a watchlist rather than initiate a broad short in AMZN, GOOG, META and MSFT. Trigger a tactical underweight only after evidence of power-cost pass-through, cloud gross-margin guide-down, or major permitting opposition; absent that evidence, their balance sheets can absorb energy inflation and AI demand remains the dominant earnings driver.
• For power scarcity exposure, favor VST over CEG on pullbacks over the next 3-9 months, but hedge regulatory risk with a modest XLU long. VST has greater upside if gas-fired marginal power prices reset higher; reduce exposure if state regulators cap large-load cost recovery or if forward power spreads fail to widen.