The AI Bottleneck Stopped Being Compute. It's the Power Grid Now — and Nobody Has a Roadmap for That.

For two years the story of AI capacity has been about chips. Who has H100s, who's getting B200 allocations, who's stuck waiting on TSMC. NVIDIA's stock chart turned into a graph of compute scarcity. Every roadmap meeting started with "how many GPUs do we have."

That conversation just got obsolete on a Tuesday. PJM — the electricity grid that runs from Chicago to D.C. — auctioned 2026-2027 capacity at $329.17 per megawatt-day. The 2024-2025 number was $28.92. That's an eleven-times jump in two years. Microsoft is restarting Three Mile Island as a dedicated private power source for its data centers. Meta signed a 20-year contract for 1.1 GW of nuclear from Constellation. Some hyperscalers are installing literal natural gas reciprocating generators on-site because they can't wait for grid interconnection.

The bottleneck stopped being silicon. It's electricity now. And electricity does not scale on Moore's Law.

You Can Buy More Chips. You Have to Make More Electrons.

There's an asymmetry between the two halves of the AI buildout that nobody is pricing in. The chip half is hard but predictable. TSMC adds capacity on a known timeline. NVIDIA architects a successor every 18-24 months. It's a manufacturing-yield problem, and the industry has been solving manufacturing-yield problems for fifty years.

The power half is hard and not predictable. Building a new transmission line in the U.S. takes ten years. Permitting a substation takes three. A nuclear plant from a green field takes fifteen to twenty. Even the "fast" options — combined-cycle gas, grid-scale batteries, restarting mothballed plants — have lead times measured in years and lean on an entirely different industrial base than the one that builds GPUs. You can buy more chips. You cannot buy more electrons. You have to make more electrons, and the country that taught itself to make them quickly is China, not the U.S.

So the AI capex story has two phases. Phase one (2023-2025) was "spend whatever it takes to corner the GPU supply." Phase two (2026 onward) is "spend whatever it takes to corner the power supply, because the GPUs are useless without it." The ~$690B annual hyperscaler capex is increasingly flowing into power purchase agreements, nuclear restarts, and turbines — not more racks. The bottleneck migrated.

What Developers Are About to Feel

Inference costs are about to stop falling. The 70-90%-a-year price-per-token decay we've taken for granted ran on software improvements plus cheaper hardware plus electricity being a rounding error. That last assumption is breaking. Once the marginal cost of an inference is dominated by electrons rather than chips, the decay curve flattens. Plan your unit economics like inference might cost the same in 2028 as it does today.

Capacity will become regional, not infinite. Your cloud bill assumes you can spin up however many GPUs you want, wherever, whenever. That breaks first where the grid is most constrained — Northern Virginia, Texas, Phoenix, the upper Midwest. Expect capacity reservations, multi-month waits, and prices that vary regionally by 3-5x. Multi-region architecture stops being a high-availability concern and becomes a capacity availability concern.

"Bring your own power" becomes a strategy. Hyperscalers are doing it at nation-state scale. Mid-tier customers will do it at colo scale — leases tied to specific plants, bilateral PPAs, on-prem in places with surplus power. We're going back to a world that looks weirdly like the 1990s, where where-you-host depends on where-the-power-is.

The agent era makes it worse. A coding agent running forty minutes in the background consumes orders of magnitude more inference than a chatbot answering a quick question. Every "let it cook overnight" run is a power load the grid wasn't sized for.

Who Controls the Electrons That Run the Model

The dominant 2026 framing is about who controls the model. I think that's the wrong axis. The axis that matters for the next decade is who controls the electrons that run it.

Look at the moves. Microsoft, Google, Amazon, and Meta have collectively become some of the largest private buyers of electricity in the United States. Nuclear plants slated for shutdown are staying open because Microsoft signed twenty-year contracts. SMR companies with no real customers eighteen months ago now hold multi-billion-dollar commitments. The companies you think of as software companies are quietly turning into energy companies that also do software — because that's the only way to guarantee they can run the software in 2030.

Two implications nobody's articulating. First, the AI policy debate is in the wrong building. Everyone's in Washington arguing model licensing and export controls. The fight that decides the next decade is at state public utility commissions and FERC — transmission rights, capacity auctions, interconnection priority. Microsoft, Meta, and Google staffed up energy regulatory teams two years ago. The application-layer startups that didn't are about to learn their cloud bills depend on lobbying outcomes they can't influence.

Second, edge AI gets more attractive, not less — every inference you don't send to a hyperscaler is one less competing for grid-constrained centralized capacity. The next two years will quietly reroute a lot of production workloads to small local models on commodity hardware, not because they're better, but because the electrons are cheaper at the edge.

The maker in me wants to leave you with this: we built an industry that scales with electricity, in a country that stopped scaling its electricity supply forty years ago. Every other 2026 AI headline — the capex bet, the export controls, the agent rollouts — is downstream of whether the grid can carry the load. So far the honest answer is: not really, and not fast enough. The companies that win the next phase are the ones treating kilowatts as a strategic resource on the same level as model weights and headcount. Nobody's saying it like that yet. They will be soon.

Sources: AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment | Belfer Center · Data Centres, AI and Cryptocurrencies Eye Advanced Nuclear | IAEA · Global energy demands within the AI regulatory landscape | Brookings · Energy supply for AI | IEA