The Economics and Physics of Co-Location

Co-location is the simple idea of placing a large load physically next to a generator and connecting them behind the utility's meter, so the electrons flow from plant to servers without traversing the public transmission system. The appeal is obvious once you have watched a few interconnection queues crawl. If the power never touches the shared network, the reasoning goes, then the years of transmission studies and the tens of millions in network upgrades that gate a conventional connection should not apply. You have a willing generator, a hungry load, and a short private wire between them. For a developer racing a construction schedule against a hyperscaler's deployment plan, that speed is worth an enormous amount.

The economics reinforce the appeal. A generator that sells into the wholesale market lives and dies by volatile prices and, for an inflexible baseload plant like a nuclear unit, by the pain of ramping down when prices fall. A co-located data center offers that generator something precious: a steady, high-load-factor buyer willing to sign a long contract at a negotiated price, absorbing the plant's output around the clock. For the load, the deal can mean power at a price below the delivered retail tariff, because it avoids the transmission and distribution charges bundled into that tariff, and it means speed, because it appears to bypass the queue. On paper it looks like a rare arrangement where both parties are better off. That is precisely why so much capital rushed toward it.

But the physics does not cooperate with the tidy story, and this is the part that co-location enthusiasts underweighted. A data center at a hyperscaler's scale cannot actually tolerate the availability of a single plant. Even a nuclear unit with an excellent capacity factor goes offline for refueling roughly every eighteen to twenty-four months, and it can trip unexpectedly. When the plant is down, the load still needs power, and the only place that power can come from is the shared transmission network. So the co-located load almost always keeps a connection to the grid for backup, and the moment it does, it is drawing on the network's capacity, its reserves, and its stability services, precisely the things a conventional interconnection study exists to account for and charge for.

There is a subtler physical dependence too. Large synchronous generators do more than supply energy; they provide inertia, voltage support, and frequency regulation that hold the network stable. A load bolted to one plant behind the meter still sits inside the regional grid's electrical fabric, and its behavior, its sudden swings, its fault contribution, propagates outward regardless of where the meter sits. The fence between behind-the-meter and on-grid is a commercial and legal boundary. It is not a physical firewall. The wider system still feels the load, and still has to be planned around it.

This is the tension that makes co-location genuinely hard rather than merely clever. The arrangement can deliver real speed and real savings, and in some configurations, a truly islanded load that never leans on the grid, it raises few objections. But the configurations hyperscalers actually want, full availability, seamless backup, the ability to grow, all pull the load back into reliance on the shared network. And once a load relies on the network, the question of whether it is paying its fair share of that network becomes unavoidable. That question is not an accounting quibble. It is the fault line along which the entire nuclear co-location controversy broke open, and understanding it is the difference between treating co-location as a magic bypass and treating it as the carefully bounded arrangement it actually has to be.