Data Centers & Power

Headline capacity figures (gigawatts) tell you how much compute a country can power; counting facilities tells you how concentrated the footprint is. A 2026 mapping of data centers by country made the concentration vivid: the United States led with roughly 3,960 data centers — more than the next 14 countries put together — the single largest footprint in the dataset by a wide margin.

Two cautions make the count more useful, not less. First, a "data center" is not a fixed unit: the U.S. total is inflated by many smaller and legacy facilities, so a raw count overstates the lead relative to the gigawatt-capacity picture, where China is a much closer second. Second, footprint and capability aren't the same — a few hyperscale campuses can out-power hundreds of small colocation sites. So the right reading is layered: by facility count the U.S. is overwhelmingly dominant; by powered capacity it leads but China is the credible #2; by rate of new buildout (see the power-race item) the U.S. is behind. Three lenses, three different stories — which is exactly why a single ranking misleads.

We log this as reference, not alarm: it's the baseline map the rest of this beat updates against.

If the global compute map has a single capital, it is Northern Virginia — "Data Center Alley," centered on Loudoun County. The concentration is hard to overstate: the region carries on the order of 3 GW of data-center capacity, more than entire countries like the UK or Singapore, the legacy of decades of fiber routes, tax incentives, and proximity to internet exchange points compounding into self-reinforcing density.

That very density is now the problem. So much load has been added that the binding constraint has flipped from "is there land and fiber?" to "can the utility actually deliver the electricity?" — local transmission and generation planning, not demand, set the pace. The result is twofold: new projects face longer power-availability queues in the core, and growth spills outward to secondary markets (Phoenix, Atlanta, Columbus, and beyond) chasing grid headroom the way an overflowing reservoir finds new channels.

Northern Virginia is the concrete, county-level version of the macro story this beat tracks. It shows that "capacity" isn't an abstraction — it's substations, transmission lines, and a utility's build schedule — and that even the world's most successful hub eventually runs into the physics of power delivery.

Renewables are cheap but intermittent; a frontier training cluster needs power that is both clean and constant, twenty-four hours a day. That combination — carbon-free *baseload* — points squarely at nuclear, and the hyperscalers have moved there with conviction. The headline moves include deals to restart a mothballed reactor specifically to feed a data-center fleet, and a wave of agreements to buy power from small modular reactors (SMRs) that don't fully exist yet.

The appeal is straightforward: a reactor delivers a steady, predictable block of carbon-free megawatts, which is exactly what a power-constrained, sustainability-pledged AI operator wants and what an intermittent solar farm can't promise alone. The catch is time. Restarting an idled plant takes years of licensing and refurbishment; SMRs are a technology still being commercialized, with first deliveries pushed out toward the back half of the decade. So nuclear is best read as a *bet on the structural power shortage persisting* — a signal that the operators expect electricity, not chips, to be the long-run constraint, and are willing to underwrite a slow, capital-heavy supply to secure it.

For a measurement-first view, the lesson is the timeframe. Chip shortages clear in quarters; power constraints clear in years or decades. When the companies with the best demand visibility start contracting for reactors, they are telling you which bottleneck they think is durable.

A revealing fight emerged around xAI's data-center power: the U.S. Department of Justice raised the matter of unpermitted gas turbines being used to supply electricity to the facility. Strip away the specifics and it's a preview of a structural tension the whole industry is walking into.

The logic is simple and hard to escape. Grid interconnection for a large new load can take years; a frontier AI buildout wants power in months. When the public grid can't deliver on the AI timeline, the fastest path is on-site generation — gas turbines, in practice — sited at the data center itself. That solves the schedule problem and creates two new ones: emissions and air-quality permitting that the grid would otherwise have handled, and a regulatory process that simply wasn't built for "a single private buildout adds a power plant's worth of generation on a startup timeline."

This is the local, physical face of the same constraint the U.S.–China power-race item describes from altitude. Whether the binding limit on AI ends up being chips, capital, or megawatts, the megawatts increasingly arrive with a permitting fight attached — and where (and whether) an operator can stand up its own generation becomes part of the siting calculus.

For most of the AI era the bottleneck people argued about was silicon — who has the best chips, who is cut off from them. A quieter, harder constraint has moved to the front: electricity. A frontier training run is, physically, a very large electrical load that has to be delivered to one place, continuously, for weeks. That makes the rate at which a country can add generation and transmission a direct input to how much AI it can build.

On that axis the gap is stark. Anthropic publicly flagged that China added roughly 400 gigawatts of power capacity in a recent year, while the United States added "several dozen" — on the order of one-tenth as much. The U.S. still leads on installed data-center capacity today (it hosts roughly 40% of the world's), but leading on the existing base and leading on the rate of new buildout are different things, and it is the second that determines who can stand up the next generation of training clusters.

The constraint shows up differently in each market. In the U.S. the problem is rarely "not enough power in the country" and usually "not enough power to that substation" — local transmission jams that strand demand in saturated hubs like Northern Virginia and push new builds toward Phoenix, Atlanta and Columbus. In China, a national plan ("East Data, West Computing") deliberately routes heavy training workloads to western provinces where renewable power is abundant and cheap, treating geography as a scheduling problem. The result is the same question from two directions: can the grid keep up with the compute?

A recurring theme in opinion writing is to map AI onto a small set of geopolitical "superpowers" — typically the United States and China in a tier of their own, with a third seat variously assigned (one widely-read column nominated Russia). The framing is seductive because it captures something real: data-center capacity and the power to feed it are concentrated in very few states, and that concentration has national-security weight.

We log this as a lens, not a fact, and the distinction matters. An opinion column is a *lead* — a prompt to go check the measurable substrate underneath it — not a source to quote as settled. On the measurable axes this beat actually tracks (installed capacity, rate of new power buildout, chip access), the U.S.-and-China top tier is well supported; the identity of any "third superpower," by contrast, depends heavily on which metric you privilege — installed capacity favors European leaders like Germany, while raw energy buildout or geopolitical posture points elsewhere. The honest read is that "superpower" is a narrative wrapper around a handful of hard numbers, and the numbers are what we report.

So treat the superpower framing as a way to think about stakes and sovereignty — not as a leaderboard. Where it makes a testable claim ("country X rivals the top tier"), it becomes exactly the kind of thing we want to check against capacity and power data rather than repeat.