A year ago cooling hardware was sold in watts per square centimeter and degrees per watt. This year the pitch deck ends in tokens per watt. The argument runs: a cooler chip leaks less power, so more of the rack's budget goes to compute, so the same megawatt produces more tokens. Cold plate makers have been quickest to adopt it, and they have published the chain they use.
We think the chain is right. We also think it has one more term in it than the cold plate pitch mentions, and that term is ours.
The chain, as a cold plate maker writes it
Frore Systems' July white paper on the thermal stack lays it out plainly. The chip's junction temperature is the coolant inlet temperature plus heat flow times total thermal resistance. The coolant inlet temperature is the free-cooling floor plus whatever a chiller adds. And the free-cooling floor, in the paper's own glossary, is set by ambient temperature, humidity, and the efficiency of the heat exchanger inside the coolant distribution unit.
That last item is the approach: the temperature gap the exchanger needs to move full load. Every degree of it is a degree the coolant cannot have, so it is a degree the chip pays for. A cold plate changes one resistance in that chain. The distribution unit's exchanger sets where the chain starts.
What one degree is worth, in watts
Write the chain as an equation and the approach term has a coefficient of exactly one. A degree removed at the heat exchanger arrives at the junction undiminished, because nothing between the two attenuates it. Turn it around: at a fixed junction limit, one degree of approach is one over the cold plate's thermal resistance in chip power.
Liquid-cooled accelerator cold plates sit around 0.015 to 0.021 C per watt from junction to water. That makes one degree of approach worth 48 to 67 watts per GPU, call it 57 watts, or about 4 percent of a 1,400 watt part's budget. The figure needs nothing from the operator: no climate, no electricity price, no revenue model. It holds in a new hall and a twenty-year-old one.
Where it is worth nothing
The headroom only becomes tokens if the rack is thermally limited, meaning power is available and junction temperature is what binds. In a power-limited hall, where every watt the utility will sell is already spoken for, the gain in tokens is exactly zero. The degree buys reliability margin instead, which is worth something, but not the thing on the pitch deck.
Tokens also scale less than one for one with chip power, and how much less depends on the model and the interactivity target. So the token gain is smaller than the watt gain, and it is a property of the workload, not the hardware. That is why we quote watts, not tokens or dollars.
Where we come in
A narrow approach at full load takes more heat transfer surface, or better surface, inside the distribution unit's exchanger. Stamped plates cap how much fits in the cabinet. We 3D print ours, so the channels are shaped for heat transfer and easy flow together, and more surface fits in the same box. Our 1 MW exchanger for the Open Compute Project's Deschutes design is built on that basis.
What operators should ask for
Two numbers, from two vendors. Ask the cold plate supplier for junction temperature at rated load and inlet temperature; that back-solves the resistance. Ask the distribution unit supplier for the approach at full load. The ratio is the headroom in watts per GPU per degree. Whether it turns into tokens is a question about the building's power contract, and only the operator can answer it.