When to colocate data centers with generation

Start with the clarification that reframes everything. Nearly every cloud data center already has generation on site: diesel backup gensets, sized for rare transmission-level outages and limited by air permits that cap runtime and eat into the allowance needed just to test them. What is new is generation meant to be prime power, running close to around the clock alongside a grid connection. The case for it has two parts. First, speed: if the utility says five to seven years for a connection at the scale you want, building your own might be faster. Second, that a 24/7 load needs a 24/7 generation source to match it, an argument Janous hears from the current administration as a case for more baseload plants and calls flatly false.

He attacks the second part first, because it decides what problem you are solving. A utility does not study whether it can supply you across all 8,760 hours of the year. It asks whether adding you pushes the system past what it can serve on the hottest summer afternoon and the coldest winter morning. That makes this a capacity problem rather than an energy problem, and capacity problems have far more solutions than building a machine that runs constantly. Janous describes the grid as moving power through space and time, so its pieces substitute for each other: a transmission line substitutes for baseload generation, since moving power across congestion reduces what has to be generated locally, and long-duration storage substitutes in turn for transmission. Add grid-enhancing technologies, advanced conductors, storage of varying durations and virtual power plants, and you can replicate a 24/7 output with a dozen things instead of one, faster because much of it already exists and cheaper because you build less.

The economics are unattractive, though not disqualifying. Work the arithmetic on a 100-megawatt data center: a power usage effectiveness of 1.2 puts you at 120 megawatts of generation, redundancy adds 20 or 30 more, so you build 150 and run it at perhaps 90 on average, because data centers chronically underuse nameplate. At roughly $2,700 per kilowatt with an overbuild approaching 2x, the cost per kilowatt-hour is extraordinarily high: in Texas something like $150 to $200 per megawatt-hour around the clock, while the data center next door buys grid power near $20 and the old reason to own a baseload machine there, occasional scarcity prices of $5,000 to $9,000, has been decimated by solar and storage. But Kann turns Janous’s own bit-watt spread back on him: electricity is small against what a data center earns, so a developer might pay the premium for speed anyway. Janous concedes it, probably in a lot of cases. Overpaying is not a deal killer, just no economic benefit and worse margins than a competitor who got grid access.

Which leaves the version both accept: not going off-grid, but shrinking your interconnection footprint, siting a 500-megawatt data center where the utility can deliver 300 and holding your grid draw under that ceiling. Janous agrees, then reframes it as orchestration, because behind-the-meter generation is only one way to fill the gap. A long-duration battery on the utility’s side of the meter might do the same job, and the fastest option builds nothing: aggregate enough sheddable load inside the deliverable zone that the utility credits it as capacity. Nobody has done that at real scale, and vertically integrated utilities object that they want rate base rather than a program, to which Janous answers that a virtual power plant is a bridge letting them connect the load sooner and build against it later. The limit he volunteers is that this is where his argument may fall apart: orchestrating across roughly 3,000 US utilities and several regional markets with different accreditation rules is hard, and the tool that would price a stack of capacity options for any point on the grid, trusted by utility and developer alike, does not exist. On-site nuclear, meanwhile, fails every test at once. It is not faster, not cheaper early, you will already hold your full interconnection by the time it arrives, and no one plugs a $20 billion facility into a novel reactor without a decade of operating data.