The electricity demand of AI data centres has pushed large technology companies into building their own power plants, and the fuel they have chosen is natural gas. Energy research firm Noreva says that choice carries a serious price risk.

The forecast

Noreva chief executive Peter Gardett told TechCrunch that natural gas prices in parts of the United States could triple in the coming years. His reasoning is that three factors are arriving at once: rising hyperscaler demand, slowing growth in supply, and increasing liquefied natural gas exports.

In numbers: prices currently run between $2 and $4.50 per million BTU, trading just under $3 at the Henry Hub in Louisiana. Noreva forecasts that in some regions the figure could exceed $10.

Who has committed how much

Cheap gas prices have led these companies to lock in a substantial share of the market on long-term terms:

  • Meta — a 7.5 gigawatt plant for the Hyperion data centre in Louisiana.
  • Amazon — a 7.6 gigawatt plant in Texas.
  • Microsoft — its own gigawatt-scale plant in Texas.
  • Google — a gigawatt-scale plant, also in Texas.

Facilities on this scale are built to run for decades. The choice of fuel is not a one-year procurement decision but a commitment lasting the life of the plant.

Why it reaches token prices

At a large gas plant, fuel accounts for roughly half the cost of the electricity produced. Tripling the gas price means roughly doubling the cost of that electricity.

And electricity is the largest variable cost in AI inference. A model's per-token price is directly tied to the energy consumed by the hardware producing those tokens. A data centre operating under a long-term gas commitment has to absorb a rise in electricity costs either in its margin or in its pricing.

That amounts to pressure in the opposite direction at a time when model prices have been falling for a year.

The side that shows up on the bill

There is a public dimension as well. As TechCrunch reports, 80% of consumers are already concerned about the effect of data centres on their electricity bills. A jump in gas prices affects not only data centres but everyone drawing from the same grid.

This is the point at which AI infrastructure stops being a technical subject and becomes a local political one. If the forecast holds, the argument will be conducted through energy bills — and in that argument data centres stand on the exposed side.

The limits of the forecast

It is worth noting that this is a forecast, and the track record of energy price predictions is not strong. Supply could grow faster than expected, the rise in LNG exports could slow, or data centre demand could fall short of projections.

The asymmetry is real, though: if gas stays cheap the companies will have made the right call, and if it does not they are left with a commitment lasting decades. The costs of that bet are not evenly distributed.

Most of these companies are trying to hedge the risk through diversification — nuclear, solar and grid agreements are all being pursued in parallel. Even so, natural gas remains the most-chosen option for now, because it can be built quickly and runs without interruption.