Liquid Cooling Joins the Loop

Air is a poor carrier of heat. Water and engineered coolants carry far more heat per unit of volume. As accelerators grew denser, the industry reached the point where air could not remove the heat from a rack fast enough, and liquid came to the chip. That change in cooling medium is also a change in control. It moves the facility much closer to the workload.

Look at a published design. In October 2024, Vertiv released a reference power and cooling architecture, co-developed with NVIDIA, for the GB200 NVL72, NVIDIA's liquid-cooled rack-scale system. Vertiv described it as a seven-megawatt design supporting up to 132 kilowatts per rack. The architecture uses chillers, room cooling and coolant distribution units, or CDUs. A CDU sits between the building's water loop and the loop that runs through the servers. It pumps coolant to the cold plates on the chips and exchanges heat back into the facility.

The CDU is where IT and operational technology, or OT, physically meet. On one side is the chip, governed by the scheduler and by accelerator telemetry. On the other side is the plant, governed by the building management system. The CDU controls flow and temperature on the boundary, and its sensors report on both worlds at once.

Here is why that changes the control problem. An air-cooled room has large thermal inertia. The air and the room absorb a temporary surge, and the building management system can respond over minutes. A liquid loop at the chip has much less inertia at the point of heat generation. When thousands of accelerators ramp up together, the heat arrives at the cold plates almost immediately. If the flow is not there, the chip protects itself by throttling, which lowers its clock speed and slows the job. The cost appears not as an alarm in the facility but as lost throughput in the cluster.

So thermal state becomes a scheduling input. Meta's Llama 3 paper reported a daily throughput variation of one to two percent that it attributed to midday temperatures affecting the chips' dynamic voltage and frequency scaling. That was reported on Meta's H100 cluster, not a dense liquid-cooled hall, and it was still measurable. In a fully liquid-cooled hall with tighter margins, a scheduler that ignores coolant temperatures and flow will place work where it cannot run at full speed.

The data needed to close this loop exists on both sides. NVIDIA's Data Center GPU Manager, or DCGM, reports each device's temperature and power draw and flags thermal violations. The CDU and the building management system report supply and return temperatures, flow rates and pump status. What has been missing is a shared model that relates one to the other, so that a scheduler can know that a given row has thermal headroom and another does not.

The design envelope is set by standards. ASHRAE's Technical Committee 9.9 publishes the thermal guidelines that define allowable temperature ranges for data center equipment, including classes for liquid-cooled systems. Operators design to those classes. Running warmer saves energy in the plant, because chillers work less. Running cooler protects performance and component life. The right point on that curve depends on the workload, which is exactly the information the facility has historically lacked.

The equipment vendors have organized around this. Schneider Electric acquired Motivair, a liquid-cooling specialist, and has said it is an approved CDU vendor for NVIDIA. Vertiv markets its liquid-cooling line inside its 360AI portfolio of reference designs. Both companies also sell monitoring software that gathers thermal and electrical data. The commercial logic is plain: the vendor that owns the cooling loop wants to own the telemetry that describes it.

Power behaves the same way. Liquid-cooled racks at these densities draw large, fast-moving loads. The electrical power monitoring system sees those swings at the switchgear and the power distribution units. If that data reaches the scheduler, a job can be capped or delayed before a breaker is stressed. If it does not, the facility's only tool is a hard limit that slows everything.

Faraday's read is that liquid cooling is the point at which the facility stopped being a landlord and became a participant in the computation. An air-cooled data center could be run to a contract: this much power, this much cooling, per rack. A liquid-cooled AI hall runs better as a single control system in which coolant, current and computation are scheduled together. Operators who still run it as two systems will leave throughput on the table, and they will not see it in any facility report.