The Era of Cheap, Abundant Compute Is Over, and Power Is Now the Ceiling on AI

What Happened

On June 18, 2026, the US Federal Energy Regulatory Commission (FERC) issued orders to force faster grid connection for AI data centers. Acting under Section 206 of the Federal Power Act, it directed each regional grid operator with tailored orders and set a target of processing large-load interconnection requests within 90 days, a process that has been taking years. Grid operators have 30 days to report how much spare generating capacity, if any, they hold. In exchange, data centers face conditions: they pay their own interconnection costs, and when the grid is under stress they must bring their own power or curtail demand. The action follows Energy Secretary Chris Wright’s October 2025 directive to reform large-load interconnection.

The numbers behind it are stark. US data center power demand is projected to nearly double from 80 GW in 2025 to 150 GW by 2028. About 2.2 TW of generation and storage projects sit in interconnection queues, nearly double all the capacity installed on the grid today. Ratepayer anger is rising too: one Virginia resident’s January bill hit $281, up roughly three times from about $100 the month before. More than 30 states filed over 300 data-center bills this year, and Maine is set to pass the first statewide moratorium, pausing new data centers until late 2027. Meanwhile hyperscalers are locking in their own supply: Microsoft secured 835 MW from the Three Mile Island restart on a 20-year, $16 billion deal, and Amazon and Meta signed gigawatt-scale nuclear and gas agreements.

What It Means for Founders

For a decade, founders treated compute like tap water. You turned it on in the cloud when you needed it, and the price fell every year by default. This FERC order exposes the assumption underneath that default. The real limit on compute is not the chip but the electricity to run it, and power has become a scarce resource gated by multi-year queues and political backlash. The bottleneck is no longer “just buy GPUs” but “where do you find the megawatts to feed them.”

That shift forces a recalculation on any business built on a falling compute-cost curve. If you run an inference-heavy SaaS, the assumption baked under your pricing, that inference gets cheaper every year, starts to wobble, because rising power costs feed back into per-hour cloud GPU rates. Hardware, edge, and on-premise teams find themselves outbid for capacity by hyperscalers. But the same bottleneck opens new markets. Anything that saves, shifts, or resells power grows: efficiency tools that deliver the same output on fewer watts, scheduling that moves demand into slack hours, small and quantized models, brokerage for idle power or idle GPUs, and software that optimizes data-center power and cooling. Scarcity always puts a price on thrift.

What You Can Do Now

If your team burns a lot of compute, measure your true per-unit inference and power cost now, then stress-test whether your pricing survives that cost doubling. When the risk is securing capacity rather than the sticker price, locking volume through long-term contracts or reserved instances beats buying on the spot market later. If you are hunting for a new business, look at where the bottleneck sits. Power allocation, demand shifting, and efficiency are getting priced for the next several years. The queues will not clear before 2027 at the earliest, so the team that designs to profit on expensive power wins the next round.