Cipher Digital doesn't build power plants. Or it didn't, until last week. The Nasdaq-listed data center operator announced it is laying lateral pipelines to feed up to 2.5 gigawatts of new gas-fired generation at its own sites, enough on-site power to keep pace with the tenants lining up for its high-performance computing capacity. Cipher has begun development of natural-gas lateral pipelines at several locations, targeting enough fuel supply to support up to 2.5 GW of new on-site generation, and expects the projects to make electricity available before the end of 2027. Nobody at Cipher mentioned a GPU shortage. They didn't need to.
That silence is the story. For two years, the AI hardware race was a story about silicon: who could get Nvidia's Blackwell chips first, whose orders TSMC would fill, whether SK Hynix could stack enough high-bandwidth memory to keep the whole machine fed. That scarcity is easing. Nvidia's high-end AI GPU shipment mix is expected to shift significantly in 2026, with the Blackwell architecture forecast to account for more than 70 percent of shipments, according to TrendForce. The new scarcity sits somewhere the chip analysts don't usually look: substations, transformers, and gas laterals. Electricity infrastructure — not chips, capital, or software — is emerging as the dominant constraint on AI expansion.
The demand numbers explain why. Goldman Sachs Research forecasts U.S. data center power demand will jump from 31 GW in 2025 to 66 GW by 2027 — a doubling in roughly two years, on a grid built for decades of gradual load growth, not a compute boom. Gartner puts a number on the mismatch that ought to worry every hyperscaler's site-selection team: Gartner projects that 40% of AI data centers will be power-constrained by 2027. Individual campuses aren't asking for a rounding error, either. Individual AI data center sites are requesting 100 to 750 MW of power each, loads that many regional grids simply weren't built to deliver on short notice. Meanwhile the queue of everything else waiting to connect keeps growing behind them: nearly 2,300 gigawatts of generation and storage capacity are currently sitting in U.S. interconnection queues, with wait times in many regions now stretching past five years.
Faced with that math, developers are doing what Cipher just did: skip the queue entirely. The shortage, along with long wait times for grid connection, has driven the rise of the BYOP (bring your own power) model of onsite power generation, and while many developers intend to use it only until the grid catches up, a few plan to bypass the grid indefinitely. Meta has gone furthest: one project opting to forgo a grid connection is a Meta Platforms datacentre campus near Columbus, Ohio. Cipher's chief executive, Tyler Page, framed the logic bluntly: "Securing and delivering power is the foundation of everything we do and bringing our own electricity generation allows us to unlock significant new capacity quickly." Chips can wait in a warehouse. Megawatts cannot be stockpiled.
Politicians have noticed, and they are not thrilled. Pennsylvania Governor Josh Shapiro, whose state sits atop the Marcellus Shale that could fuel much of this build-out, just signed an executive order applying what he called the country's toughest guardrails on data center growth, telling developers "we will not be bullied by these developers. We will not be bulldozed by the lawyers for these big tech companies." The order requires data centers to disclose their energy and water use and directs Pennsylvania's utility regulator to keep costs from shifting onto households, a real concern in a state where average household electricity rates jumped nearly 14 percent in the last year. The gas industry, which had expected to be the boom's biggest beneficiary, was blindsided; a Marcellus trade group called the pivot disappointing, per Natural Gas Intelligence. Virginia has gone further still, imposing a new consumption tax — a new consumption tax of $0.011 per kilowatt-hour on all data center electricity use starting July 1, 2026, a levy expected to generate around $600 million a year — turning grid strain directly into state revenue.
The obvious counter is that chips still matter enormously. SK Hynix supplies the majority of HBM3e stacked memory required for Blackwell chips, while TSMC's CoWoS packaging capacity is fully allocated through at least mid-2027. That's a real constraint, and it isn't vanishing. But it's a constraint with a known clock. Foundries can be expanded, memory lines retooled, packaging capacity doubled — TSMC has done exactly that before. High-voltage substations cannot. As DNV's energy chief Ditlev Engel put it, "Grid connectivity is increasingly becoming the factor that determines which AI projects move forward and which remain on paper." A GPU shortage delays a product launch. A power shortage cancels the site.
That asymmetry is why the smart money in AI infrastructure is quietly shifting from foundry equity to gas turbines, transformers, and utility-scale interconnects. Energy stocks have already caught the drift: the sector-wide XLE has traded firmly higher through the year, closing at $64.06 and sitting well above its 50-day average, per market data. Natural gas, the fuel Cipher is now piping directly to its own sites, last settled at $2.90/MMBtu per the latest EIA data.
The next phase of the AI build-out won't be won by whoever secures the most Blackwell allocation. It will be won by whoever owns the transformer.
