Everyone keeps blaming the substation. The real constraint on the next wave of data center buildout is thermodynamics on the back side of the rack. Chips can be fed. The heat they throw off has to go somewhere, and in a lot of markets there is nowhere left to put it that isn’t already fought over. Air gets hotter, water gets scarcer, neighbors get louder, and the sidewalk out front keeps filling up with press releases about megawatts.
That mismatch is reshaping how operators choose between two broad approaches: reject the heat, or sell it. Neither one is winning cleanly. Each has rooms it owns and rooms where it falls apart.
The Old Model Rejects the Heat. The New Model Tries to Sell It
Legacy cooling design followed a simple brief: pull heat off the chips, move it outside, make it the atmosphere’s problem. Chillers, cooling towers, dry coolers, evaporative pads, and the specialized heat exchangers that sit between the server loop and the outdoor loop all exist to do that one job cleanly. It worked when a rack drew five kilowatts and the site had land, water, and a tolerant grid.
The newer model treats the same heat as inventory. Pipe it to a neighboring hospital, a greenhouse, a university loop, or a municipal network, and let someone else burn less gas because you ran your GPUs. On paper, two problems solved with one capital stack. In the field, it collides with physics, contracts, and geography in ways that take real work to unwind.
Density Is the Reason Both Camps Are Rethinking Everything
Rack density is the forcing function behind the whole debate. Industry reporting puts typical racks at 30 to 50 kW today versus 2 to 5 kW a decade back, with AI-tuned designs headed past 100 kW. At those loads, blowing cold air at a cabinet stops being a serious plan. Direct-to-chip liquid loops and rear-door heat exchangers show up because nothing else keeps the silicon alive.
That shift cuts both ways in the reject-vs-sell argument. Liquid loops produce hotter, more concentrated return water than legacy air systems ever did, which happens to be exactly what a district heating customer wants to buy. The same density also means a single site now generates thermal output on the scale of a small industrial plant, so the “just vent it” answer gets more expensive and more visible every quarter. Federal cooling research programs have started chasing the same problem.
Reuse Wins When the Buyer Is Already Next Door
Heat reuse stops being a slideshow and starts being a project when a serious thermal customer sits close enough to pipe to without rebuilding a neighborhood. The technology to move low-grade heat to buildings exists, has existed, and is routine in parts of Northern Europe. What most markets lack is co-location and a contract.
Where that lines up, the economics flip. The data center stops paying to reject every joule and starts getting paid, or at least credited, for a slice of them. The equipment is nothing special. What the deal needs is zoning, piping, and a counterparty willing to sign a twenty-year agreement.
The Hard Parts Nobody Puts in the Press Release
Reuse is not free energy with a bow on it. The waste stream coming off a server hall is low-grade heat, usually well under the temperature a traditional district system wants at the inlet. Raising it to a usable inlet temperature takes heat pumps, which take electricity, which eats into the carbon story the project was sold on.
Then there’s seasonality. Heating demand peaks in winter while a hyperscale campus throws off the most surplus heat in summer, when nobody wants it. Operators who build around reuse still need a full rejection system for the months the buyer isn’t buying. Two cooling architectures, not one, and the capital math has to clear on both.
Pick the Fight That Matches the Site
For most operators, this is a site-by-site decision made against local conditions, not a single corporate choice. A remote campus with cheap land and dry air should probably optimize a rejection system and stop pretending a district loop is coming. An urban or suburban site next to a university, a hospital complex, or a dense residential development should price out reuse seriously, including the heat pump penalty and the backup rejection it still needs.
The sites that get into trouble are the ones that pick an architecture to match a corporate narrative instead of the ground under them. Chasing a reuse headline where there is no buyer within pipeline distance wastes capital. Defaulting to evaporative rejection in a basin already fighting over an aquifer invites the permit denial that stops the project cold.
The data center next door is going to be built. The open question is whether the heat it generates ends up warming a building, warming a river, or warming an argument at the next planning board meeting.
