Data centre waste heat: Why almost nobody is using it

This is article 2 of 8 in The Third Number series. Here is article 1.

A hundred megawatts of computing, running hard the way the machines going up this year are meant to run, will draw something like 745,000 megawatt-hours across a year before a single fan turns or a drop of water evaporates. And almost nobody outside the industry knows what happens to it. Essentially all of it comes back out as heat. Not most of it. Nearly all of it. Electricity goes in one end and warmth comes out the other, and the site's entire cooling plant exists to move that warmth somewhere else as fast as it appears.

Vincent Thibault put it to Nicholas Sokic of Sustainable Biz Canada in March 2023 in a sentence I have not been able to improve on. "It's a one (to) one ratio. So the amount of energy that's going to be consumed by the computer is the amount of energy that's going to be retrofitted in the form of heat."

So a hundred megawatts of computing is also, at the same instant, a hundred megawatts of heating.

Thibault's company bet roughly a billion Canadian dollars on that sentence.

Their reasoning was the kind any operator in this audience would have signed off on. The heat is free, it is enormous, it is continuous, and northern agriculture is starving for exactly that. Site the campus near farmland to keep the pipe short. Lock in hydroelectric power at four and a half cents. Let a pipeline company move the heat, because you are in the computing business and they are not. Build the hard part first, at your own cost, before anybody asks you to, and the buyer at the far end will follow, because the arithmetic is obvious to anyone who looks at it or at least it should be.

Three years on, the data centre is running and the far end of that pipe has not been financed.

What happened to them is the subject of this piece, and it is not a story about a bad bet. It is a story about a good one that nobody could write down in a recognised ledger no matter how real it is.

A Danish town answered this sixty years ago

Kalundborg has about seventeen thousand people in it. It also has a refinery, a power station, a pharmaceutical plant, a plasterboard factory, and a dozen other operations that between them now trade water, steam, gas, sulphur, fly ash and leftover heat with each other through a web of pipes. What one plant finishes with, the next one buys.

It began in 1961 with a single pipeline carrying water from a nearby lake to a new refinery, and it grew outward for six decades, one bilateral deal at a time.

The order matters, and it is the opposite of what you would expect. Nobody designed Kalundborg. It was assembled by plant managers doing deals that paid, each of them solving a problem on their own site with a counterparty a few kilometres away. The ecological reading came afterward. In 1989 Frosch and Gallopoulos wrote in Scientific American about what they called the industrial ecosystem, in which the use of energy and materials is optimised, waste and pollution are minimised, and every output of a process has an economically viable role somewhere else.

They were describing, in the language of biology, something Danish engineers had already built without needing the metaphor.

And the biology behind it is not a slogan, though it has been worn down into one. One organism's waste is another's food is a structural fact about how every ecosystem on earth stays open for business. It is the reason a forest floor is not waist-deep in the corpses of last century's trees. Nothing in a living system is discarded, because there is always something for which the discard is a meal. That arrangement is what lets a forest run for ten thousand years on sunlight and rain with nobody hauling the residue away.

So we know it works. We have known since 1961. And the heat coming off the world's data centres is still, overwhelmingly, thrown at the sky.

Why?

The answer is a calendar

A data centre rejects heat flat. All 8,760 hours of the year, at essentially the same rate, because the servers do not know it is July.

A district heating network, which is the obvious buyer and is common across northern Europe, does not work that way at all. Demand peaks in the cold and collapses in the summer to little more than the hot water people wash in.

So even a facility sitting on top of a willing utility, with a pipe already in the ground, cannot sell more than a fraction of the heat it makes. For half the year the buyer does not want it at any price.

This half year of 'no place to go' is the constraint that quietly kills most of the heat reuse projects that get announced, and it cannot be buried. Meta's facility at Odense in Denmark was purpose-built for heat recovery, sits beside a willing Danish utility, and delivers about 165,000 megawatt-hours a year, enough for roughly nine thousand households. That is the best-documented project of its kind in the world, and it recovers a minority of what the site produces. Fifteen to twenty-five per cent of annual heat sold is what the good Nordic projects actually achieve. Anyone modelling half is modelling something nobody has built…yet.

Temperature is the second tall wall, and it costs money to climb. Air-cooled halls throw off heat at twenty-five to thirty-five degrees C, which is warm rather than hot, and a district network wants sixty and above. Somebody has to install heat pumps and pay to run them. Yandex's site at Mäntsälä in Finland is the cleanest published illustration: seventy-five megawatts of computing, a four-megawatt heat pump, exhaust leaving at thirty to forty-five degrees and boosted to fifty-five or sixty, heating about twenty-five hundred homes.

And then the part that should stop anyone building a business case. In most of these projects no money changes hands at all. Meta gives Odense its heat for nothing. Amazon gives Tallaght its heat for nothing. The one place on earth where waste heat is a traded commodity is Stockholm, where Exergi's open district heating market buys from more than thirty data centres across sixteen providers, and by 2022 was recovering enough to warm on the order of thirty thousand apartments.

Stockholm is the exception that shows how rare the arrangement is. Everywhere else, the operator gives the heat away, because giving it away is easier than valuing something the books have always treated as waste.

And then the thing that springs the trap

It is so simple that I have watched very smart people miss it.

A radiator does not want heat in July.

A tomato does.

A greenhouse buys warmth in August at the same rate it buys warmth in February, because a plant under glass needs its root zone and its air held steady all year, and the further north the greenhouse sits the more true that becomes. It wants that heat at exactly the temperature a data centre is slightly embarrassed to be producing, which is warm and not hot. No heat pump. No boosting. No capital stack to climb.

The seasonality that limits every district heating project in the world simply does not apply, because the buyer at the far end of the pipe is growing food, and food grows every month.

Put a greenhouse at the end of the pipe, and a hundred megawatts of computing stops being a cooling problem and becomes an agricultural input and a community export.

Somebody worked this out and built for it

QScale's campus at Lévis, across the river from Quebec City, sits on about a hundred acres and was sited near farmland deliberately, to keep the pipe short.

It has 142 megawatts reserved from Hydro-Québec on a grid that is essentially all hydroelectric, at electricity prices around four and a half United States cents per kilowatt-hour, which is below every state in the American industrial world. The climate gives it free cooling for up to eighty per cent of the year. The heat is designed to travel as warm water within a radius of about ten kilometres, moved by Énergir, Quebec's largest gas distributor, because QScale is in the computing business and would rather let a pipeline company run pipelines.

The published intention is a campus capable of heating more than fifteen thousand Quebec households, on roughly a billion Canadian dollars of investment, with claimed savings at completion of about 816,000 megawatt-hours and 146,576 tonnes of greenhouse gas a year. And the greenhouses were to grow 2,880 tonnes of raspberries and something over eighty-two thousand tonnes of tomatoes, annually, on the exhaust of a supercomputer.

Eighty-two thousand tonnes of tomatoes. That is what an ecosystem-level transfer looks like when you draw it on a page.

Now the information we are required to make.

The greenhouses are not built. The heat recovery system was expected to be running in early 2024. As of the middle of 2026 no operating greenhouse at Lévis appears anywhere in the public record, and the agricultural partner was described by the company itself as an as-yet-unannounced arrangement with an experienced greenhouse producer. The first two phases of the data centre are complete and running.

The tomatoes are a rendering.

Why the tomatoes are still a rendering

I want to be careful about what that means, because it is not an accusation, and QScale has done more here than anyone.

The physics works. The site is right, the climate is right, the power is clean and cheap, and the operator built the campus for this on purpose, at their own cost, before anyone asked them to.

Which is exactly where their reasoning had a hole in it, and it is a hole I would not have seen either.

They believed the constraint was physical. Get the heat to the right temperature, get the pipe short enough, get the power clean enough, get a partner who knows how to move energy, and the economics follow, because at that point the deal is simply obvious. Every one of those beliefs was correct. They executed all of them.

The constraint was never physical. It was that the value of the thing they were offering had no unit, and a deal with no unit cannot be underwritten no matter how obvious it looks to the people standing next to it.

So ask what a lender was supposed to price.

The greenhouse operator brings a crop, a market, and thirty years of horticultural risk. The data centre brings heat it currently gives away for free. There is no agreed way to say what a generation of guaranteed warm water is worth. There is no instrument that puts the operator's avoided cooling, the grower's avoided gas, the region's avoided emissions and the county's new jobs into one number that two boards and a bank can argue over.

So the deal that everyone can look at and call a good deal sits unfinanced. Not because anyone doubts it. Because nobody can write it down as a number everyone recognises.

The third number is what is missing, and it is missing in a way you can walk up to and touch. There is a pipe in the ground in Quebec with warm water in it and nothing at the far end.

The shape of the Third Number

It is starting anyway, small, where the arithmetic happens to be easy enough to do on the back of an envelope. Green Mountain in Norway sends its heat less than a kilometre to what is described as the world's largest land-based trout farm. EcoDataCenter at Falun in Sweden is building toward greenhouses and fish.

Those two have the same thing in common with Kalundborg. Short pipe. One counterparty. A deal simple enough that two plant managers could shake hands on it without anyone needing a shared standard first.

That is how a system like this always starts, and it is also why it nearly always stops. Bilateral deals do not scale. Kalundborg took sixty years to reach a few dozen exchanges because every single one had to be negotiated from scratch, privately, with no common way to say what was being traded. Multiply that across ten thousand data centres and the arithmetic of the negotiation itself becomes the bottleneck.

What turns a handful of handshakes into an industry is a shared way of counting that can change industries (see). Not a better pipe. Not a better heat pump. An agreed unit for what the warm water is worth over a generation, so that the hundredth deal takes a week instead of three years.

The map already tells us where the treasure is. What we have never built is the thing that tells a bank what the treasure is worth.

What we are asking for

Somebody reading this has a heat offtake deal that everyone agreed was a good deal and nobody could finance. The BiosVerse-Compatibility Index (BCI) scores whether a design fits. The Generational Cost of Ownership (GCOO) model prices what the fit or the misfit costs across generations. Both exist and both run today. What is not settled is how they are governed: who holds the standard, what counts as evidence, and who has the standing to fail a project. That is the work of Biomimetics26, on November 9 and 10 at the Minnesota Landscape Arboretum, and it is open. Bring that deal, and bring what the lender asked you for that you could not produce.

Next: walk a forest a season after a bad fire and you will not find a control room in the ashes, because there never was one.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.