US data centers projected to consume 18 bcf/day natural gas by 2035

Full analysis

BloombergNEF says U.S. data centers could burn 18 billion cubic feet of natural gas a day by 2035, more than Germany and Japan combined, and nearly double what the same shop forecast nine months ago. For anyone who buys or runs AI, the useful question is not "will AI eat the grid." It's whether the cost of running a model is about to become an energy-contract problem you can't code your way out of, and whether the price of that gas lands on your inference bill.

How hard is this to undo? For the reader, easy. You don't own turbines. But your cloud provider is making choices right now, at 18-month turbine lead times, that are very hard to undo, and you inherit the pricing.

What's actually being decided: not "is AI bad for the climate." It's whether compute cost stops falling on schedule because energy, not chips, becomes the binding constraint, and which vendors get insulated from that.

Deadline: none sharp. This is a 2030–2035 curve. But PPAs and gas turbine orders being signed in 2026 set the floor you'll pay on in 2028.

The Skeptic. A 100% upward revision in nine months is not a reason to trust the forecast. It's a reason to trust it less. This is a straight-line extrapolation of capex press releases, and press releases are not deployed megawatts. These same four hyperscalers have a long history of announcing, overbuilding, underutilizing, and writing down. Nobody has priced inference getting cheaper per token: smaller models, distillation, custom silicon like Google's TPUs and Amazon's Trainium all cut watts per useful answer. The last "AI eats the grid" story was the 2000 fiber boom, which ended in stranded assets and a decade of flat demand. 18 bcf/day assumes the scaling curve never bends. It always bends.

The Compute Pragmatist. Energy just became a first-order cost, not an ops footnote. If gas prices climb because data centers and LNG exports both pull on the same pipe, the marginal cost of a training run moves, and it moves unevenly. Labs with locked-in power contracts or their own generation get a cost floor competitors buying spot power can't match. That's what Meta, Microsoft, Google, and Amazon are really buying with onsite gas plants: not energy security, control of the cost floor. The grid-connected demand is five times the size of the owned plants, which means mid-tier clouds and colocation shops face a disadvantage they cannot capital-efficiency their way out of.

The Enterprise Buyer. For a CTO signing a multi-year inference commit, this is the argument for locking price now. If your provider's power cost is heading up, month-to-month token pricing is exposure you don't want. The uncomfortable part: the same integration that gives hyperscalers a cost floor gives them pricing power over you. The off-grid plant strategy also carries a reputational tail. Noreva warns utility ratepayers eat higher bills even when hyperscalers absorb their own costs. If your company has public climate commitments, "our AI runs on a private gas plant that bypasses the grid" is a line you may have to answer for to a board or a regulator.

The Safety Lens. A million metric tons of greenhouse gas a day from data centers, roughly 12% of current U.S. emissions, and the off-grid plant strategy is built to route around the visibility that grid connection provides. Onsite generation at this scale sits in a gap where FERC, state utility commissions, and EPA all have thin jurisdiction. The companies best positioned to accelerate AI are also the ones most insulated from the bill for it. Regulators fixated on model alignment and misuse are missing a physical harm that's already being poured in concrete. This one is genuinely hard to undo. A turbine ordered in 2026 runs for 30 years.

Where they split. The Skeptic and the Compute Pragmatist are the real disagreement. The Skeptic says the demand curve bends because efficiency always shows up and capex announcements evaporate. The Pragmatist says it doesn't matter, because even if per-token cost falls, total demand is set by how much compute the labs choose to buy, and they keep choosing more. Both can be right: watts per answer drop while total watts climb, because cheaper inference gets used more. That's the pattern every prior computing wave followed.

The second split is Safety versus Enterprise Buyer. The Buyer sees vertical integration as a cost win to lock in. The Safety Lens sees the same move as an accountability gap that lands on ratepayers and the climate. The off-grid plant is a good deal for whoever owns it and a cost pushed onto everyone else on the pipe.

What it hinges on. One belief: does total AI compute demand keep outrunning efficiency gains, or does it plateau? If it plateaus, 18 bcf/day is another stranded-fiber story and this forecast gets revised back down. If it doesn't, energy becomes the thing that decides who can afford to train frontier models, and the four companies buying their own generation lock in an advantage nobody else can match with clever engineering.

The council leans toward the Pragmatist on direction and the Skeptic on the specific number. Demand growth is real and structural. 18 bcf/day is a wet finger in the wind that will get revised again, probably more than once, and not always up.

What to verify before you treat this as gospel: watch whether BloombergNEF's next revision holds, and watch whether any hyperscaler's actual deployed power capacity tracks its announcements. Announced plants and energized plants are different numbers. Underwrite the second one.

Prediction: BloombergNEF will revise its U.S. data-center gas-demand forecast again by more than 15% in either direction before its next annual outlook lands (by end of 2026), because the underlying compute-demand curve is moving faster than the model can settle.

Confidence: Medium — the doubling in nine months proves the model isn't stable, but direction is genuinely uncertain.

Why: BloombergNEF already doubled this number in nine months, which means their demand model has no stable anchor and is chasing capex announcements that keep changing. When a forecast moves 100% in three quarters, the next revision is far more likely to be another big move than to hold steady, because nothing in the underlying inputs has stabilized: turbine orders, model efficiency, and hyperscaler buildout plans are all still in flux. The reason I won't call the direction is exactly the Skeptic-versus-Pragmatist split: efficiency gains could pull the number down while raw compute demand pushes it up, and either could dominate the next print. A stable forecast is the least likely outcome here.

Revisit by 2026-12-31: We're right if BloombergNEF's next data-center natural-gas demand projection moves more than 15% up or down from 18 bcf/day. We're wrong if the next projection lands within 15% of 18 bcf/day.

The value here isn't the number. It's that anyone underwriting power contracts or inference pricing off a single BloombergNEF figure is building on sand that's still shifting under them.