Everyone Is Building the Same Data Center

Fly into any of the handful of counties that have become synonymous with the AI buildout — parts of northern Virginia, central Ohio, the high desert outside Phoenix, the sagebrush around Cheyenne — and the buildings look almost embarrassingly alike. Long, low, windowless boxes arranged in twos and fours, ringed by chain-link and substations, with cooling towers or dry coolers stacked along one flank like ribs. Microsoft’s version, Google’s version, Meta’s version, the version some sovereign wealth fund is financing on spec for a colocation tenant it hasn’t signed yet — they are converging, structurally and mechanically, on a single design. That convergence is not an accident of taste. It is what happens when the actual constraint on the industry stops being imagination, or even capital, and becomes something far more prosaic: amperage.

For most of the cloud era, data centers competed on things you could see in a spec sheet — chip generation, network fabric, PUE, proprietary cooling. Those differences still exist, but they have stopped being the deciding factor in where and how fast a company can build. The deciding factor now is whether a given plot of land can be fed enough electricity, on a timeline measured in single-digit years rather than one, to run tens of thousands of accelerators at full tilt. Silicon differentiation has not disappeared. It has simply been demoted, because the binding constraint sits several links upstream of the server rack, in the interconnection queue of whichever regional grid operator happens to run that patch of ground.

The gigawatt is the new SKU

It is worth being precise about scale, because the numbers involved have quietly moved from “large industrial facility” to “small city.” A single frontier-scale training campus now under discussion in the industry runs on the order of a gigawatt or more of continuous draw — roughly the output of a full-size nuclear reactor, dedicated to one company’s compute. Multiply that across the four or five hyperscalers simultaneously trying to secure capacity, and you get an aggregate demand curve that, by most utility planning estimates I’ve seen referenced informally, is growing faster than grid operators have modeled for in decades. This is the part that doesn’t show up in earnings-call slides about “AI infrastructure investment”: the bottleneck isn’t fabrication capacity at TSMC, or even the availability of high-bandwidth memory, both of which are real constraints but ones the industry has priced and hedged against for years. The bottleneck is a grid interconnection process built for a world where large new loads arrived one at a time, a few times a decade, not five times at once from parties each capable of writing a check for the entire annual capital budget of the utility they’re asking to serve them.

That shift explains the design convergence directly. When the limiting resource is power rather than land or capital, the optimization problem collapses to a narrow set of variables: maximize compute density per megawatt, minimize time-to-energization, and site wherever interconnection is fastest rather than wherever labor, climate, or fiber happen to be cheapest. Everyone solving the same equation arrives at similar answers. Hence the modular, prefabricated halls that can be dropped onto a pad in months rather than poured in place over years; hence the near-universal shift toward higher rack densities and liquid cooling, which reduce the footprint (and therefore the civil and electrical buildout) needed per unit of compute; hence the fact that the “signature” architectural choices companies used to tout — custom air-handling designs, novel building envelopes — have given way to whatever off-the-shelf modular vendor can deliver switchgear on the shortest lead time.

Behind the meter, ahead of the queue

The more interesting story is not the buildings but what companies are doing to get around the queue rather than through it. Several developers have begun negotiating to co-locate new data centers directly against existing or planned generation — natural gas plants, nuclear uprates, and in a few well-publicized cases, previously shuttered nuclear units being brought back for a single anchor tenant — specifically to avoid the years-long wait for a conventional interconnection study and upgrade cycle. This “behind the meter” approach is less a technology strategy than a permitting strategy: if you own or contract directly with the generation source and never actually cross onto the public transmission grid at meaningful scale, you can sidestep much of the queue entirely. It is, in effect, the same logic that led big retailers to build private logistics networks rather than wait on a public freight system that wasn’t going to move fast enough for them — except the freight here is electrons, and the private network is a power purchase agreement plus a fence line.

That has produced a curious inversion. Companies that a decade ago wanted nothing to do with owning physical energy infrastructure are now co-investing in gas turbines, small modular reactor startups, and geothermal pilot projects, not because they’ve developed a strategic interest in power generation but because generation has become the only lever left that determines whether a training run happens on schedule. One infrastructure lead I spoke with, who works across several of these campus buildout teams, put it more bluntly than any public filing would: the chip roadmap is now the easy part of the plan.

The chip roadmap is now the easy part of the plan.

Why the queue itself won’t clear soon

It is tempting to assume this is a temporary bottleneck that new transmission investment or faster permitting will resolve within a normal planning cycle. There’s reason for skepticism. Transmission buildout in the US has historically taken on the order of a decade from proposal to energization once siting disputes, environmental review, and cost-allocation fights among neighboring utilities are factored in — a timeline set by a regulatory and legal architecture that predates the current demand spike by fifty years and shows no sign of being rewritten wholesale. Meanwhile the interconnection queues themselves, at least by the pattern visible in public regional-operator filings, have become clogged partly with speculative requests — projects filed to hold a place in line rather than because the developer has committed capital, which further slows the ones that are real. Every fix on the table — queue reform, cost-allocation changes, co-location rules, faster environmental review for transmission — is a multi-year regulatory project running up against demand that is compounding on a multi-quarter cycle.

Differentiation moves downstream, then disappears

The natural endpoint of this dynamic is that the companies best positioned to keep building are not necessarily the ones with the best model architecture or the most efficient chip, but the ones with the best government-affairs team, the deepest relationships with utility commissioners, and the balance sheet to underwrite a gas plant or a reactor restart on a bet that they’ll need the power in three years rather than five. That is a different kind of moat than the industry has been used to competing on, and it rewards a different kind of company — one that looks, increasingly, like a utility holding a compute business, rather than a compute company that happens to need power. The data centers themselves will keep looking the same, because the physics and the queue leave little room for them not to. The real competitive terrain has simply moved somewhere the buildings can’t show you: into substation interconnection agreements, utility commission dockets, and the fine print of who gets priority when the grid operator finally says yes.