I built liquid-cooled GPU farms before anyone called it AI infrastructure

Past a certain rack density, thermal design stops being a line item and becomes the whole design. Everything else, the power, the fabric, the software, arranges itself around the problem of getting the heat out. That is the lesson the mine farms taught me years before an AI rack made it fashionable.

Between 2012 and 2016 I designed and delivered crypto mine farms end to end. Not buying pre-made miners and racking them. Designing the whole thing: the GPU clusters, the cooling, the power, and the software that fed the machines work and collected what they produced.

High-density compute, honestly dense

A mining rack is a preview of an AI rack. Multiple GPUs per board, each pulling two hundred watts and more, packed as tightly as the thermals would allow. High-density compute sounds like a spec sheet phrase until you are the one deciding how many kilowatts you can put in a single frame before the room turns against you.

Air only takes you so far. Past a certain density the fans are just moving hot air into more hot air. So I went to liquid, both ways. Direct-to-chip cold plates on the cards that could take them, with manifolds and quick-disconnect fittings so a failed card could come out without draining the loop. And full immersion cooling for the densest builds, boards sitting in dielectric fluid, no fans at all, the heat pulled straight into the tank. Leak detection under every rack, because the lessons were expensive and occasionally wet.

Power is the other half of thermal

You cannot design the cooling without designing the power, because every watt you deliver is a watt you have to remove. I laid out the power distribution from the supply in: the boards, the PDUs, the circuits and breakers, and the very real problem of inrush current when a room full of GPUs comes on at once. Balancing phases, sizing for sustained draw rather than nameplate, and watching the cost per kilowatt-hour against what the machines actually returned. Thermal design and power design are the same conversation, just measured in different units.

The system that fed the machines

The hardware was only half the build. The other half was the mining pool architecture: the software layer that took a huge pool of hashrate and turned it into coordinated, useful work. Miners connected over the stratum protocol, the pool handed out jobs, tracked shares, and distributed rewards. My job was to make that job distribution fast and fair, and above all low-latency, because a slow or badly scheduled pool means stale work and lost revenue every second it lags. Scheduling, failover, watchdogs, auto-restart when a rig fell off the bus. The operational discipline people now call SRE, learned because every minute offline cost money.

Why this still matters

The rigs are long gone. The craft stuck. When I size an AI cluster today, I am not starting from a vendor whitepaper. I am starting from the delta-T across a cold plate, the phase balance on a busy PDU, and the arithmetic of getting forty kilowatts out of a rack before it throttles.

The GPUs got bigger. The TDPs went up. The workload changed from hashing to training and inference. The thermodynamics did not move an inch. Heat still flows from hot to cold, coolant still needs flow rate and delta-T, and the first question about any dense cluster is still the cooling plant, not the model. I just happened to learn that a decade before the phrase AI infrastructure existed.