Waste heat doesn't sell.

There is a sentence in almost every datacenter deck of the last three years: we will valorise our waste heat with the local district heating network. It sits near the end, between decarbonisation and community impact, and nobody puts a number on it.

We did. Here is what came out.

Every figure on this page is recomputed by a standalone simulation, released with the dossier.

The revenue nobody checks

A 100 MW site rejects 834 GWh of thermal energy a year. Sold at the best public tariff we could find — return-line injection into a Nordic network, €24.8/MWh — that is €11.3 to €13.1 million net per year, depending on how much of the rack is liquid-cooled.

Against the compute revenue of the same capacity, measured in a listed operator's annual report: 0.3 to 2.0%.

That is below the rounding error of a financing plan. No business model stands on it, and any deck presenting waste heat as a meaningful revenue line will be taken apart by the first analyst with a calculator.

So why do the live sites do it anyway?

Because they don't sell it. They give it away.

The three live connections we could measure all work the same way. At Hamina, the operator gives its heat to the city, which itself carries the 7.5 MW heat pump and the 1.3 km of pipe. At Odense, the heat goes out against reimbursement of the equipment, nothing more. At Deep Green, it heats municipal swimming pools, free.

None of them invoices it. All of them get something more expensive in return.

In a sector where the grid connection queue runs into years, where a permit is won or lost on neighbourhood acceptance, and where a local council can stall a €500M project with one vote, heat is not a product. It is an access asset. It buys the permit, it buys acceptability, sometimes it buys priority on a substation. Those are the genuinely scarce resources here — not capital, not land, not machines.

Three things we found written nowhere

1. Temperature doesn't rise with power. It falls.

This is the opposite of everyone's intuition, including ours at the start. A more powerful chip runs hotter, so the loop should be hotter. Wrong: to move more watts through the same thermal resistance without exceeding the junction limit, you have to cool the inlet water. The loop comes down.

We put a number on it. A manufacturer's showcase measures 50.5 K between water and silicon for a 1,893 W chip — 26.7 kelvin per kilowatt of chip power. That is an exchange rate, and it turns every choice of heat customer into a bet on the next hardware generation:

A current-generation processor draws around 1,200 W. The showcase shows 1,893. The heat network dies between the two.

2. The only customer that survives is the one who wants cold

If deliverable temperature falls with every generation, the only durable outlet is one whose target sits below the rejection temperature. Where a network requires lifting 7 to 21 K and industry 47 to 91 K, there is nothing to lift — the exchange runs the right way, and it still will when the chip draws three kilowatts.

That customer exists and is in service: an Arctic trout farm fed from 800 metres away by a Norwegian datacenter since autumn 2025. 1.75 MW, 9,000 tonnes of fish a year, and the water goes back to cool the datacenter.

We state its limit too, because it is structural: a 100 MW site would need 12 to 55 farms of that size. The best branch per megawatt is the worst by volume, and no commercial arrangement lifts that.

3. The volume absorbed is never the volume announced

A datacenter is a flat source: same power 8,760 hours a year. A heat network is a seasonal demand. The gap between the two decides what is actually valorised, and it has nothing to do with the negotiated price.

Our hourly model gives 55 to 70% absorption for a flat 5 MW source against an ordinary network. The only instrumented hyperscale site we found delivers 61%.

One operator advertises 96% heat recovery at its showcase site. Delivery data from the same site, published by the same source, gives 57%.

The contractual bomb nobody has defused

Put the three findings together and you get a legal problem we have not seen addressed in any public contract.

A network signing today on 60 °C at the delivery point finances its pipe over its depreciation schedule. It loses its source beyond 1,706 W per processor — possibly at the next hardware generation. The datacenter, meanwhile, will have replaced its fleet for reasons that have nothing to do with heat.

Two parties, two horizons, one contract. And none of the texts we read says who carries the gap.

Three clause structures exist, in increasing cost to the site operator: a minimum temperature guarantee with penalties; an energy-only commitment with the customer adapting; or a revision clause indexed on installed chip power — the only one that addresses the actual cause, and the only one no public contract contains.

All three need the same thing to be enforceable: a meter at the delivery point, an agreed recording interval, and a history both parties can read. We already measured what its absence costs: 96% announced, 57% delivered. The same gap, inside a twenty-year heat supply contract, gets settled in front of a judge or at the meter.

What it's worth for the climate

The 624 GWh a large network absorbs, replacing a gas boiler, avoid 140,000 tonnes of CO₂ a year. The site consumes 876,000 MWh. The threshold where one exactly offsets the other is the tipping point:

Tipping point in gCO₂ per kWh of the electricity feeding the site.

Two operational conclusions. Materials barely count: the embodied carbon of the pipe — measured on a normalised environmental declaration — is 0.1 to 1.7% of one year of avoided emissions. The pipe pays itself back in days. Architecture counts enormously: lifting the temperature instead of injecting into the return line burns 44 to 278 GWh of electricity a year, drawn from the same grid as the site, adding zero delivered kilowatt-hours. That single decision is worth up to 38 g of tipping point — forty times the weight of every material in the network.

Which gives regulators a direct consequence: mandating a delivery temperature means mandating a pump. Mandate delivered energy, never temperature.

The method, and the proof it works

This dossier isn't worth much for its conclusions. It is worth something for what happened to it while it was being written.

Seven documents, 380,548 characters, 138 tables, 213 sources opened one by one. Not research summaries: manufacturer datasheets, published tariffs with their archives, annual reports, normalised environmental declarations, national weather data. Every page was opened.

Each document was taken apart after being written, at least twice: a claim, a measurement against it, a verdict. The corrections are kept in the dossier, dated, rather than erased.

At the end we wrote a simulation that recomputes every result through an independent path. It found four divergences, and they went both ways:

The fourth is the instructive one. It moved the dossier's central climate result, it was introduced while correcting something else, and three re-readings missed it. Only an independent recomputation brought it down.

In this market, numbers circulate without instruments. An operator announces 96% and delivers 57. A 65 °C return temperature is quoted everywhere and appears on no manufacturer page — at the specification's real flow rate it is 56 to 59. A university dissertation gives seventeen times the normalised value for pipe carbon, and its own results table repeats identical numbers for two different indicators.

We discarded those sources, and we wrote why in the dossier. What this work offers is not an opinion on waste heat. It is a set of numbers a challenger cannot bring down, in a field where almost all the others fall at the first check.

What's missing, and it's one thing

All of this is work on public sources. No measurement on a live site. We say it first because it is the first question a serious reader asks, and because it is also the answer to what comes next.

One heat meter, on one real site, for one full heating season. At the delivery point. Recording three quantities and no more: the temperature actually delivered, hour by hour; the GWh actually absorbed by the customer; and the hourly emission factor of the grid feeding the site.

Those three readings turn a defensible dossier into an enforceable one. They cost a meter and a season — the best ratio in the whole dossier between what a measurement costs and what it unlocks.

What we don't promise

That heat becomes profitable. It won't: 0.3 to 2.0% of revenue, and no technical development changes that order of magnitude. That the climate gain is a given — it can be zero where the heat network is already decarbonised, precisely in Northern Europe where the technical conditions are best. That aquaculture solves the volume problem. It doesn't.

What we do claim

That every figure here comes from an open source or a written calculation, and that none of them falls under scrutiny — because we spent more time trying to bring them down than producing them.

In a market where 96% announced is worth 57% delivered, that is the only thing separating a dossier from a brochure.

Full dossier: 8 parts — pitch, fact sheet, technical, calculations, simulation, sources, what is not measured, and the seven working documents in full. Available on request.