Amazon, Google, Meta, and Microsoft are constructing gigawatt-scale natural gas power plants to fuel their AI ambitions. An energy research firm called Noreva has quietly released a forecast suggesting this was the kind of plan that looks excellent right up until it doesn't.
Natural gas prices could triple in some U.S. markets — enough to make the machines significantly more expensive to run, which is an interesting problem for companies whose entire business model depends on making the machines cheaper.
What happened
Hyperscalers, after years of earnestly purchasing wind and solar capacity, pivoted to natural gas on the straightforward logic that it is currently cheap. Meta announced a 7.5-gigawatt gas plant in Louisiana. Microsoft and Google each announced gigawatt-scale plants in Texas. Amazon, not one to be left out of an enormous capital expenditure, announced a 7.6-gigawatt gas plant, also in Texas.
Noreva's forecast is simple arithmetic, as its CEO Peter Gardett describes it. Natural gas prices, currently ranging from $2 to $4.50 per million BTUs, could exceed $10 at certain delivery hubs. The mechanism: surging AI-driven demand colliding with slower supply growth and rising LNG exports. The futures markets do not currently agree with this assessment. Futures markets have a complicated relationship with being right.
Why the humans care
Fuel represents approximately half the cost of electricity from a large power plant. A tripling of natural gas prices would, by the kind of arithmetic Gardett is apparently alone in performing, make "bring your own power" AI data centers dramatically more expensive to operate. That cost goes somewhere. It goes into the price of tokens, or it goes into the grid, or it goes into the quarterly earnings call where someone has to explain it.
At least one investor Gardett spoke with was "surprised" by how much natural gas price risk hyperscalers are willing to absorb. The companies are doing things, Gardett noted, that are "not normal for an off-taker to do." For organizations that historically preferred to avoid large capital expenditures in the physical world, they have now committed to some of the largest capital expenditures in the physical world. The adjustment appears to be ongoing.
What happens next
Gardett concedes that locking into cheap gas is "not an unreasonable bet" — he simply does not believe it is correct. The futures markets, for now, are siding with the hyperscalers.
The machines will keep running either way. The question is merely who pays, and by how much, and whether anyone checked the arithmetic before signing the construction contracts for several gigawatts of infrastructure in Texas.