Meta, Microsoft, and Google Are Building Massive Natural Gas Plants for AI. It May Not End Well.

Three of the world's most powerful technology companies — Meta, Microsoft, and Google — are making large bets on new natural gas power plants to fuel their AI data center expansion. It is a pragmatic answer to an immediate problem: AI training and inference workloads require far more consistent, high-density power than renewable sources can currently provide at scale. But the long-term consequences of that bet deserve harder scrutiny than they're currently receiving.

The Immediate Problem Is Real

The power demands of frontier AI are not a future concern — they're a present bottleneck. Training a single large language model can consume megawatt-hours of electricity. Running inference across millions of simultaneous users compounds that demand exponentially. The US electrical grid, in its current state, cannot reliably deliver the consistent high-density power that AI compute requires. Natural gas plants, which can be sited near data centers and dispatched on demand, are a pragmatic solution to a genuine engineering constraint.

This is why the announcements are not surprising. What's worth examining is the assumption embedded in the investment: that natural gas will remain the right answer across the 15-to-25-year operational lifetime of these plants.

The Stranded Asset Problem

Long-duration battery storage costs have been falling faster than most models predicted five years ago. Next-generation nuclear (small modular reactors) is moving from theoretical to permitted at multiple sites. Grid-scale renewable buildout is accelerating. By the mid-2030s, the combination of these factors will likely make on-site natural gas generation significantly more expensive than grid power from clean sources — assuming the regulatory environment doesn't accelerate that timeline further.

A natural gas plant built today to serve a 2026 data center is a 20-year capital commitment. The companies making these decisions are betting that the economics won't shift fast enough to strand those assets. That bet may be right. But the history of energy infrastructure is littered with stranded assets built on the assumption that current conditions would persist.

The Regulatory Wildcard

Carbon regulation in the United States has historically been inconsistent. The current administration's posture is favorable to fossil fuel investment. A future administration could change that calculus substantially, particularly if climate pressures intensify and AI's energy footprint becomes a public and political issue — which, given the trajectory of AI infrastructure investment, it almost certainly will.

For companies that have made net-zero commitments — and all three of Meta, Microsoft, and Google have — building new fossil fuel capacity creates an obvious tension that will require increasingly elaborate accounting to manage.