What 15,000 Looks Like

Picture a two-megawatt unit humming in the mechanical room of Santa Fe Community College. Not a building. A closed box the size of a delivery truck, tucked next to the boilers, running inference and video processing for the college's own systems and a handful of regional partners. Its cooling loop comes out the other side at fifty degrees Celsius, piped straight into the campus heating system. In August, the surplus heat runs a greenhouse instead. Net water consumption: close to zero. Nobody's electric bill in Santa Fe went up $37 a month to build it.

That is one node. The last piece put a number on how many more like it the country needs: fifteen thousand.

Here is where that number comes from, and why it is the good half of this argument.

Of the roughly 100 gigawatts of data center capacity the United States will need by 2030, about 30 belong in hyperscale campuses — the frontier-training clusters that genuinely cannot be split apart. I have already argued where those thirty gigawatts should and should not go, and that the upper Rio Grande is not on the list. The remaining seventy gigawatts is a different animal entirely: inference, storage, video transcoding, enterprise computing, scientific workloads, archival. None of it requires the internal bandwidth that forces hyperscale clustering. All of it can run in facilities sized between one and ten megawatts, built into or beside buildings that already exist. At an average of five megawatts a node, seventy gigawatts of distributed compute comes out to roughly fifteen thousand mid-scale facilities, nationally.

Fifteen thousand is not a strange number. The country already runs about sixteen thousand wastewater treatment plants, seventy-four hundred hospitals, and a hundred thirty thousand K-12 schools. We have built infrastructure at this scale before, more than once, without treating it as a moonshot. Spread evenly by population, that's on the order of three hundred nodes per state. Weighted for New Mexico, something like sixty to a hundred. Weighted down to Rio Arriba County, one or two.

One or two, not zero. That is the whole point of this piece.

Every one of those numbers comes from the same physics. A heat pump running at a coefficient of performance around 5.0 takes electricity costing ten cents a kilowatt-hour and delivers usable heat at something closer to eight-tenths of a cent. High-temperature heat pumps can lift a data center's thirty-two-degree cooling water to eighty-five degrees or better — hot enough to run a district heating loop, a greenhouse, an aquaculture tank, a grain dryer. This is not speculative technology. Stockholm has been doing it for a decade. Helsinki runs district heat off data center waste heat as a matter of municipal policy. Hohhot, in Inner Mongolia, does it at a scale China treats as unremarkable.

What that means on the ground, here, looks like this. A five-megawatt unit at a regional agricultural processing facility, running cold storage in summer and grain drying in fall, its waste heat warming greenhouses through a Rio Arriba winter that currently loses most of its growing season to frost. A half-megawatt unit sized to a tribal community building — not a campus, a building — heating the structure itself and a year-round greenhouse next to it, and running the regional broadband link on the side. A two-megawatt unit at Northern New Mexico College doing exactly what the one in Santa Fe does: inference and edge computing for the institution, heat for the institution, water consumption approaching nothing.

None of this is charity infrastructure. The capital cost per megawatt for mid-scale modular facilities with heat reuse runs eight to fifteen million dollars — the same range as hyperscale, sometimes cheaper. The difference was never the price. It's who captures the benefit. A three-hundred-megawatt hyperscale campus concentrates its property tax, its handful of jobs, and its water and grid strain in one county, usually a county that didn't get much say in the siting. Fifteen thousand nodes spread that same seventy gigawatts of value across every hospital, every ag co-op, every community college, every tribal building with the load to support it — and the heat stays where the cold is, instead of getting shipped nowhere at all in a cooling tower's plume.

There's a version of this argument that stays purely technical, heat pumps and coefficients of performance, and it would still be worth making. But I keep coming back to something closer to home. An acequia works because the water is governed by the people who use it — a mayordomo elected by the parciantes on his own ditch, a repartimiento negotiated by neighbors in a drought year, nobody upstream extracting value while somebody downstream absorbs the cost. Fifteen thousand distributed nodes are the acequia model applied to compute. The municipality that hosts the node keeps the heat, keeps the jobs that exist, keeps some say over what runs on it and who it serves. Nobody in Española is signing an NDA with a hyperscale developer who breaks ground, builds a fence, and answers to shareholders in another state. The community that generates the data keeps the infrastructure that processes it, sized to what that community actually needs.

That is not a small difference dressed up as a big one. It is the same inversion this whole series has been arguing for, applied to the seventy percent of the problem that was never supposed to be a fight in the first place: the aquifer gets first vote, the community gets second, capital gets third — and at mid-scale, all three can actually sit in the same room.

None of it happens by default. The current federal incentive structure subsidizes the hyperscale build-out and treats mid-scale distributed compute as an afterthought, if it treats it as anything. That has to flip: mandatory heat reuse as a condition of favorable tax treatment, graduated cost recovery pushing capital toward the model that spreads the benefit instead of concentrating it. Congress has not been asked to make that trade because almost nobody has laid out what the alternative actually looks like, node by node, county by county.

So here it is, for one county. Rio Arriba doesn't need to be on anyone's no-build list to also not need a three-hundred-megawatt campus. It needs one or two nodes, sized to a hospital or a community college or a tribal building that already exists, running the heat into a greenhouse that already needs it, governed by the people who turn it on every morning. That's not a compromise position between "no data centers" and "whatever Washington sites." It's the version of this fight actually worth winning.

heat is wasted only when nothing nearby is cold — site the compute where the cold lives.