Meta, Microsoft, Google, and Amazon have collectively locked in a wave of natural gas turbine orders and site-specific power deals over the past two years, betting that generation capacity secured today will still make economic sense a decade or more from now. That bet is now facing serious scrutiny.
According to TechCrunch, a new industry forecast suggests the rush into gas could turn into one of the more expensive missteps of the AI infrastructure buildout — not because gas can't power data centers, but because the economics underpinning decades-long contracts may not hold for the full life of those deals.
The stakes are unusual for a fuel decision. These aren't short power-purchase agreements that can be renegotiated in a few years. Gas turbines take years to build and are typically financed over 20-to-30-year horizons, meaning hyperscalers have effectively become anchor customers for an industry that, until the AI boom, was shrinking in most developed markets.
Why gas became the default choice
The logic behind the gas rush is straightforward, even if the long-term math is now in question. Grid interconnection queues in major U.S. markets routinely stretch past three to five years, far slower than the timelines hyperscalers need to bring new AI campuses online. Turbine manufacturers, led by GE Vernova and Siemens Energy, have order backlogs extending toward the end of the decade, which pushed several hyperscalers to lock in capacity early just to secure a place in line.
- On-site or co-located gas generation lets a hyperscaler control its own build timeline instead of waiting on a utility or regional grid operator.
- Long-term turbine orders guarantee a manufacturing slot in a supply-constrained market.
- Gas plants can be sited and permitted faster in many U.S. states than large-scale wind, solar-plus-storage, or nuclear projects of comparable output.
For companies racing to bring gigawatt-scale AI capacity online, that speed premium looked worth paying. Some of these deals have already produced dedicated gas-fired campuses tied directly to a single hyperscaler's data center buildout, rather than power drawn from the general grid.
Where the new forecast complicates the bet
The risk TechCrunch flags is less about gas prices spiking and more about the shape of the cost curve over the life of a decades-long contract. If the price of alternative generation — solar-plus-storage, advanced grid batteries, or nuclear restarts — falls faster than expected over the next five to ten years, hyperscalers could end up locked into gas contracts priced for a market that no longer exists, while competitors with more flexible power portfolios pay less per megawatt-hour for the same AI workloads.
In our estimation, the bigger unknown compounding this is demand: these gas commitments assume AI compute growth continues on something close to its current trajectory for the full contract term, and any meaningful slowdown would leave hyperscalers paying for capacity they don't need on top of paying above-market rates for it.
What this means for AI builders
Most AI teams don't sign power deals directly, but the economics of hyperscaler energy sourcing eventually show up in compute pricing, capacity availability, and platform reliability. A few practical takeaways:
- Long-term compute cost assumptions should account for the possibility that energy costs, not chip costs, become the dominant swing factor in cloud pricing over the next decade.
- Capacity commitments tied to a specific hyperscaler's data center region are only as reliable as that provider's underlying power buildout — worth a direct question in any multi-year infrastructure negotiation.
- Providers with more diversified power portfolios — mixing gas with renewables, storage, or nuclear — may be better insulated from the scenario TechCrunch describes, which is a reasonable factor to weigh when comparing cloud vendors for large, sustained AI workloads.
AiiN's takeaway
The gas rush made sense as a speed play: it was the fastest way to get power flowing to AI campuses while grid queues and turbine backlogs made every other option slower. But speed and cost-efficiency are different bets, and hyperscalers made this one on a timeline measured in decades, not product cycles. If the forecast TechCrunch cites proves out, the companies that diversified their power sourcing — rather than defaulting to gas because it was available first — will have the more resilient cost base for the next phase of the AI buildout.