1. the unit

everything in my trade is priced in tons of refrigeration. one ton is 12,000 BTU/hr, which is the heat it takes to melt one ton of ice in twenty four hours. thats where the word comes from. a walk-in cooler at a small store is 2 or 3 tons. a supermarket parallel rack plant, low and medium temp together, is around 200 tons. thats the biggest single machine i have personally stood inside of while it was failing.

2. the arithmetic

take a press release that says up to 2 GW of accelerators. thats the number everybody quotes. here is that number in my units, and there is nothing clever in here, its unit conversion.

a watt is a watt. a computer is a resistive heater that does math on the way through. essentially one hundred percent of what you put in the panel comes back out as heat you have to physically move outside the building.

3. and that number is LOW

i quoted the press release instead of the panel, which is the exact mistake i give helpers a hard time about. you dont size a condenser off the compressor nameplate, you size it off total heat of rejection.

2 GW is the IT load. the fans, the pumps, the tower motors, the UPS and transformer losses all draw on top of it, and all of that lands in the same place. at a PUE of 1.2, site draw is 2.4 GW and the rejection load is:

so i undersold my own number by twenty percent by being lazy about which number the press release was giving me. corrected here, dated, on purpose. call it 680,000 tons.

4. in a unit i can stand next to

about three thousand supermarket refrigeration plants, running at design capacity, forever, on one campus, on one interconnect, in one climate.

i have watched one 200 ton plant cook itself. it did not fail because of a bad design or a bad compressor. it failed because somebody added ONE case at the end of a line that was already at its limit in july, and the wall thermometer kept reading a true number the whole time while the box starved. thats the panel and its the other half of this page.

5. water, if its evaporative

assumption showing so you can go at it: cooling tower evaporation runs roughly 1.8 gallons per hour per ton, which comes out of the old rule that circulating flow is about 3 gpm/ton and you lose about one percent of it to evaporation.

if its air cooled or closed loop that number goes away and the power number goes up instead, because you paid for the water with electricity. you dont get to skip it. you get to choose which bill it shows up on. i am not interested in the morality fight, i am interested in whether the plant can dump the heat on the worst afternoon.

6. the footnote that actually keeps me up

a cooling tower cannot make water colder than the wet bulb. best case you land a few degrees above it, and that gap is called approach, and approach is the whole ballgame in a valley in july.

the design condition everybody sizes to is the 0.4% wet bulb. 0.4% of 8,760 hours is 35 hours a year. thirty five hours you agreed in advance to be over. and those hours do not arrive politely spread out across the calendar. they arrive in a block, in a week, at the same time as everybody elses block, on the same grid, when the grid is already being asked for everything it has.

7. behind the meter power is a second condenser

added 2026-07-27 because somebody raised 470 million dollars for factory built microreactors and the whole timeline read it as an electricity story. behind the meter power, own your own generation, stop waiting on an interconnect. fine. in my units it is not a power source, its another thing on the roof i have to reject heat from.

a thermal plant makes heat and sells you a third of it as electricity. call one unit 5 MWe at about a third thermal efficiency:

2,843 tons per unit. fourteen supermarket plants, before one GPU has made a single watt of heat.

now cover the site draw from section 3 with them:

480 units. i know thats a coincidence and i know exactly what it looks like that the count came out as the voltage stamped on every panel i open. i wrote it down anyway because i am not going to pretend i didnt notice.

the honest part, because this page is not a dunk. powerhouse reject heat comes off HOT, 100°F and up, and hot heat is much easier to get rid of than 45°F chilled water. you can dry cool a lot of it and pay for it in efficiency instead of in water. that is a real engineering advantage and i am not going to hide it to make the number look worse. two million tons of hot reject is not the same problem as two million tons of process cooling.

what does not change: the sink is still the atmosphere over the same square mile, the ceiling on any of it is still the wet bulb, and the 35 hours in section 6 do not care what kind of boiler you bought. solving power without solving rejection is a store manager buying a bigger compressor for a condenser that is already packed with lint. i pulled one of those out of a Scotsman this morning. it came out like a felt pad and the manager thought it was the board.

8. what this page is not

not a prediction of a price

this is not proof that any of it fails. its unit conversion and a few rules of thumb, and a man with a gauge can still be wrong about the whole system, which is a thing i wrote a page about before i did it to myself from a parked van.

i have one call on the record with a date on it: november 30, 2026. this page is not that call. this page exists so that when the date comes, anybody can read what i was actually claiming, in units, with the assumptions written down, instead of what i sounded like at eight in the morning on a monday. if somebody who does this work for a living tells me a number here is off, i change the number and put the date on the change.

thats the only reason to write your arithmetic down. so that being wrong is something specific that happened, instead of a mood you can slide out from under.

no analytics on this page or any of them. i dont want to know.