AI Racks Are Outgrowing Air — What Liquid Cooling Changes for BMS in GCC Data Halls

Why Air Runs Out

Air is a poor carrier of heat. Moving a given quantity of heat with air rather than liquid needs roughly three orders of magnitude more volume, and in a data hall that volume has to be pushed through a raised floor, a containment aisle and a rack full of obstructions. Up to around 30kW a rack, good containment and correctly commissioned CRAH control handle it. Past that, the fan energy needed climbs steeply, static pressure control gets brittle, and a single blanking-panel gap turns into a genuine hot spot. AI and GPU racks start where that curve is already unpleasant and go well beyond it.

Liquid removes the constraint by bringing the coolant to the chip. Direct-to-chip runs a cold plate on the processor with a closed coolant loop; immersion submerges the hardware in dielectric fluid; rear-door heat exchangers sit between the two in both cost and capability. All three are in service in the region now, and all three put equipment into the hall that the facilities team is accountable for but which did not exist on the old CRAC-era points list.

What Actually Gets Added to the BMS

The practical addition is smaller than people fear and more specific than most tender documents describe. Per coolant loop: supply and return temperature, flow rate, and differential pressure, each trended rather than merely alarmed, because the useful signal is drift over days. From the Coolant Distribution Unit: pump duty and standby status with rotation, filter differential pressure, setpoint and valve position. Across the hall: zone-level leak detection, and dew point.

Dew point deserves its own sentence. If coolant supply temperature drops below the hall dew point, water condenses onto pipework and manifolds inside a room full of live electronics. The control logic has to hold coolant supply above dew point at all times, which means the humidity sensing has to be trustworthy and the interlock has to be real — not a dashboard warning somebody notices on Monday.

The Alarm Philosophy Has to Change

This is the part most often carried over unexamined from air-cooled practice, and it is the part that matters most. An air-cooled hall fails slowly. Lose a CRAH unit and the room has thermal mass; temperature ramps over minutes, and a supervisory BMS polling on a normal cycle catches it with time to spare. A liquid loop has almost no thermal mass at the cold plate. Lose flow and the affected processors reach throttling and then thermal shutdown in seconds, not minutes. A leak behaves the same way — the damage is done long before a poll cycle notices.

The consequence is architectural: safety interlocks belong local to the plant, hardwired or resident in the CDU controller, where they act in milliseconds. The BMS owns supervision, correlation, trending and escalation — it should not be in the safety path. Designs that route a leak trip through a supervisory head-end are the ones that look fine at commissioning and fail the first time they are genuinely tested.

The Hard Part in This Climate Is the Other End

Liquid cooling is often sold on efficiency, and at the rack it delivers. But the heat still has to leave the building. Whatever the coolant absorbs is rejected to ambient through dry coolers or chillers working against a GCC summer design day, and that side of the system is no easier than it ever was — arguably harder, because higher coolant return temperatures are exactly what make dry rejection viable, and getting that trade-off wrong turns a claimed efficiency gain into a chiller running flat out in August. Regional water scarcity pushes the same way, away from evaporative rejection and toward closed-loop designs. The control strategy that ties coolant setpoints to ambient conditions is where the actual operating cost is won or lost.

Hybrid Halls Are the Realistic Near Term

Very few UAE operators are building greenfield all-liquid halls. The common case is an existing facility adding a few high-density rows while the rest stays on CRAH units — and a hybrid hall is harder to control well than either pure approach. Two cooling systems draw on the same heat rejection capacity, and if they sit on separate control islands with independent setpoints they will work against each other, each responding to a condition the other created. Bringing both onto a single head-end, with one sequence arbitrating between them, is precisely the retrofit integration problem — just with newer equipment on one side of it.