Every input to the AI build-out is fought over. Two of its outputs are not. One of them turns out to be the cleanest test we have of whether the whole thing is real.
Power, land, transformers, high-bandwidth memory, water rights, the electricians who can terminate a 1,400-amp bus bar. Every input to this build-out has more buyers than sellers, and every one of them gets written about weekly.
Two outputs are not contested at all. A data centre produces waste heat continuously, and retired hardware on a schedule. Both volumes are set by decisions that have already been made — by how much capacity is switched on, and by how fast the owner has told its auditors it writes the equipment down.
That makes them the only forecastable quantities in the entire build-out that nobody is competing to buy. And in the last fortnight, one of them stopped being a curiosity and became the most useful number on our weekly scoreboard.
Start with the physics, because it is not a metaphor. Nearly every kilowatt-hour that enters a rack leaves as heat. A 100 MW hall of servers is, thermodynamically, an 85–90 MW heat plant that also does matrix multiplication.
What changed recently is the grade of that heat, not its existence. Air-cooled halls exhaust air at 30–40 degrees, which a district heating network will not take without a heat pump in between. The direct-liquid-cooled racks that AI hardware requires return coolant at 55–65 degrees — already hot enough for a modern heat network. The architecture AI forced on the industry is the architecture that makes its waste heat sellable.
Stockholm Exergi's Open District Heating scheme pays, in its own published words, indicatively 2 million SEK per year for a heat delivery corresponding to 1 MW.1 The payment is set by avoided cost — what the same heat would have cost Exergi to make in its own plants — and it is temperature-indexed, rising as the weather gets colder.
The scheme has 27 suppliers, 16 of them data centres, delivering about 120 GWh a year. That is roughly 1.5% of Exergi's customer demand against a 10% long-term target, in a 12 TWh network. Eleven thousand apartments are heated this way.
Now the counter-example, a few hundred kilometres south. Meta's data centre in Odense delivers 165,000 MWh a year of surplus heat to Fjernvarme Fyn — enough for up to about 9,000 households — free of charge.2
A decade-old tariff in Stockholm with sixteen data centres on it. One of the largest operators in Europe giving the same commodity away for nothing. That is not a market clearing at a low price. That is an absent market.
Which raises the obvious question: if Stockholm pays, why does Meta give it away? The honest answer is that in the Nordic cases the permit and the social licence are the prize, not the revenue. Stockholm's utility is willing to pay because recovered heat displaces production it would otherwise have to make itself. Meta's problem in Odense was permission and reputation in a country where the alternative heat source was coal and where the planning system can stop you.
Anyone telling you waste-heat revenue will change where American data centres get built is selling a Nordic institutional fact as though it were a law of physics.
Germany's Energy Efficiency Act does something no market has managed: it turns heat reuse into a construction requirement. A data centre commissioned from 1 July 2026 must reach a share of reused energy of at least 10 per cent; one commissioned from 1 July 2027, 15 per cent; from 1 July 2028, 20 per cent.8
Two details that summaries of this law routinely flatten, and both matter. The dates are commencement-of-operation triggers, not compliance deadlines — the statute gives operators up to two years after switch-on to actually hit the number, so a hall energising in July 2028 owes its 20 per cent by July 2030. And the obligation attaches at a non-redundant rated connected load of 300 kilowatts, which is a different and broader measure than the IT load a data-centre operator would normally quote.
Now the part that matters most for this piece. The law lets an operator escape the reuse obligation entirely by signing an agreement with a nearby municipality or heat-network operator. And it specifies what that agreement must contain: an investment plan, a provision on who bears the cost of the connecting pipe, and a provision on the price at which the waste heat is supplied.9
So a heat price is legally required to exist in every one of those agreements in Germany — and nothing requires anyone to publish it.
That is the clearest possible statement of what we mean by an absent market. The price is not missing because the commodity is worthless. It is missing because it is negotiated bilaterally, written into a private contract to satisfy a statute, and filed away. Every one of those agreements is a data point that exists and that nobody outside the two signatories can see.
The second stream has a published release schedule, and it is hiding in plain sight in annual reports.
Microsoft extended the useful life of its server and network equipment from four years to six beginning in its 2023 financial year, worth about $3.7bn of operating income that year.3 Alphabet did the same thing in January 2023, expecting to cut depreciation by about $3.4bn and eventually reporting $3.9bn.4
Those disclosures are the closest thing that exists to a schedule for the world's future supply of second-hand accelerators. If a machine is written down over six years, the company is telling you when it expects to stop wanting it.
In its annual report filed in February 2025, Amazon disclosed that it was changing the useful lives of a subset of its servers and networking equipment, effective 1 January 2025, from six years to five. In the same filing it recorded about $920m of accelerated depreciation on equipment it had decided to retire early.5
Amazon's own stated reason: "an increased pace of technology development, particularly in the area of artificial intelligence and machine learning."
The realised effect the following year was $1.4bn more depreciation and $1.0bn less net income, primarily at AWS.6
Thirteen months earlier, the same company had lengthened the life of the same asset class to six years. Every other lengthening in this industry has been justified by hardware lasting longer. Amazon's shortening is justified by technology moving faster.
It is also worth noting what Amazon has not disclosed: which subset, or how large it is. Not as a dollar figure, not as a share of the fleet.
The accounting cannot move. It is a straight line committed to in a filing, fixed for years. The second-hand market re-prices every week. The gap between those two is the cleanest available test of whether the compute shortage is real.
If demand is genuine, obsolete chips keep clearing near what the books say they are worth, because there is more work than hardware. If it is not, the resale price cracks first — before any income statement can show it, because the depreciation schedule is frozen and the bid is not.
We added this as the fifth test on our weekly scoreboard. The rule was written on 20 September, before we had looked at any of the data, and we wrote down what we expected to find: strengthening, because the loudest available evidence pointed that way. We were wrong. Amazon had already shortened, in a filing from February 2025, and our own rule marked the cell against the answer we anticipated.
This piece began as four streams. Two did not survive contact with the sources, and saying so is more useful than quietly publishing the two that did.
The volume is real — every gigawatt sits behind oil-filled transformers, and dielectric fluid is reclaimable rather than consumable, which is exactly the condition under which a market should form. Two things killed it.
First, and decisively: there is no published price series for reclaimed transformer oil anywhere. There are quality standards that define what the reclaimed product must be, and there are trading-site listings carrying seller-quoted asks. There is no index, no assessment, no series. The price is not merely low; it is not published. That absence is itself the finding, and the route to a number would be a utility rate case rather than a market data vendor.
Second, the volume does not scale with compute the way the premise assumed. Indoor halls increasingly use dry-type transformers because fire codes and insurers dislike oil indoors, so the oil lives at the campus substation rather than beside the rack.
American data centres consumed an estimated 66 billion litres of water directly in 2023, with hyperscale and colocation facilities accounting for 84% of it, and average site water-use effectiveness staying just over 0.36 litres per kilowatt-hour through that year.7
But water is not a waste stream you can sell. It is a priced liability, and the industry is designing it out: closed-loop cooling with dry coolers drives on-site consumption toward zero. Water use therefore scales with the share of the fleet still on open evaporative towers — a share that is falling — not with installed compute. The genuinely interesting object here is the inverse of the premise: treated municipal waste water sold as an input, which is a water utility's product with a data centre as the customer.
Heat is real, physically guaranteed the moment a hall is switched on, geographically stranded, and — outside one Swedish scheme — unpriced in public even where the law requires a price to be written down. The thing to watch is announced gigawatts in a city that has a heat network, with no thermal connection filed within eighteen months of switch-on. If heat were valuable, the pipes would get poured where pipes are possible.
There is also a specific document worth asking for. Any German data centre relying on a heat-offtake agreement to satisfy its reuse obligation is holding a contract that states a heat price. A municipality is a public body. Someone will eventually publish one of those numbers, and when they do, the first real price outside Stockholm will exist.
Retired silicon is real, dated years in advance by depreciation policy, and it has already started talking. The thing to watch is simple and it sits in annual reports: does any other large buyer shorten the life it expects from its own servers? Microsoft, Alphabet, Meta and the specialist clouds have so far only lengthened. Amazon went the other way and said why.
The hyperscalers have told their auditors these machines last six years. The second-hand market gets to vote every week. Watch the gap.