The bottleneck isn’t power. It’s heat.

For a year the whole argument about AI infrastructure has been about power. The interconnection queue, the capacity auction, the moratorium. Megawatts, megawatts, megawatts. Power is real and it is scarce. But there is a second constraint sitting directly underneath it, and it is closer than the megawatt fight makes it look. It is heat.

Here is the arithmetic almost nobody in the capital stack has run. A conventional data center hall was designed for 5 to 10 kilowatts a rack. The good ones, with contained aisles and serious engineering, push to 15 or 20. A single NVIDIA GB200 NVL72 rack draws about 120 kilowatts. The generation after it goes higher, and the one being designed now is aimed past 500. Air does not move that much heat out of a rack at any fan speed. Somewhere north of 40 kilowatts, moving air quits, and you have to put liquid on the chip. There is no version of the next three years where the racks get cooler.

So the real question for an operator is not how many megawatts they hold. It is how many of those megawatts they can actually deliver at 120 kilowatts a rack, with liquid to the chip, this quarter. Most contracted capacity in this country cannot answer yes to that. It was built for a workload that no longer exists.

And that is the part the market has not priced. Operators are valued on megawatts under contract and megawatts in the pipeline. A megawatt that needs an eighteen-month retrofit before a single NVL72 can power on is carried at the same value as a megawatt that is liquid-ready next quarter. Those are not the same asset. One earns now. The other is a construction project wearing a lease.

Three things I see from inside the supply chain that the filings do not show. Retrofit is slower than build: converting a live hall to liquid means working around tenants, and operators quote a year or more to do it. Water is becoming the second permit: the cheap way to reject that much heat is to evaporate millions of gallons a day, and the places with power are exactly the places starting to say no to the water, which is the whole reason we run dry. And tenants have stopped buying space. Every GPU buyer I talk to opens with kilowatts per rack and time to power on, not square feet. Their chips are depreciating in a warehouse while they wait for a floor that can hold them.

Put those together and the mispricing is obvious. The scarce thing is not a megawatt. It is a megawatt you can deliver as dense, liquid-cooled, water-light capacity, now, on power that already exists. The operators, builders and owners who have that are sitting on a premium the market is still handing out for free.

That is the entire reason our machine is liquid-cooled from the first drawing, not retrofitted into. A complete 2 megawatt block, roughly a thousand GPUs, liquid to the chip, zero water in every operating mode, live in weeks on a pad that already has power. It does not ask an operator to convert a hall or win a water permit. It arrives already on the right side of the ceiling.

Power gets the headlines because power is easy to count. Heat is the ceiling nobody is pricing, and it is the one that decides which megawatts are worth anything.

Monuments are measured in megawatts. Machines are measured in what you can actually cool.

Henri Francois is the co-founder of Borealis by Stilllife, a data center fabrication company building modular, zero-water AI infrastructure in Corona, California. henri@borealis.cool