AI Data Center Cooling: Air, Liquid and What Changes in South Florida

The Density Problem, in Numbers

• Average rack density went from about 16 kW in 2025 to roughly 27 kW in 2026 — and forecasts put the average near 40 kW within a few years.
• AI deployments routinely land at 50–70 kW per rack, and only about one in five operators say they are ready to support that.
• The newest AI systems reach as high as ~246 kW in a single rack — roughly thirty times a conventional cabinet.

Put a 100 kW rack in a room designed for 8 kW cabinets and you do not get a hot spot. You get a room that physically cannot move enough air to carry the heat away, no matter how many CRAH units you add.

Where Each Cooling Method Actually Applies

• Air cooling — up to roughly 30–40 kW per rack. Still the right answer for the overwhelming majority of racks in the world, including nearly every enterprise room and edge site. Requires disciplined containment: hot and cold aisle separation, blanking panels, sealed floor cutouts, correct static pressure. Most rooms are leaving capacity on the table simply because containment was never finished.
• Rear-door heat exchangers — roughly 40–80 kW. A liquid-cooled door on the back of the cabinet captures heat before it enters the room. The least disruptive way to add density to an existing air-cooled hall, because the room's basic airflow design stays intact.
• Direct-to-chip (D2C) — roughly 50–150 kW. Cold plates sit directly on the CPUs and GPUs; a coolant distribution unit (CDU) circulates fluid through them and rejects the heat to a facility water loop. This is the current mainstream answer for new AI racks. Note that D2C typically captures 70–80% of the heat — the remainder still leaves as hot air, so the room still needs air cooling. Hybrid is the normal state, not a transitional one.
• Immersion — above roughly 100 kW. Entire servers submerged in dielectric fluid. Highest thermal ceiling, but it changes rack design, service procedures and floor loading, and it has to be designed in from the start rather than retrofitted.

What South Florida Adds to the Problem

• Wet-bulb, not dry-bulb, is the constraint. Evaporative and adiabatic strategies that carry the load cheaply in Phoenix or Denver lose most of their advantage in a climate that spends half the year near saturation. South Florida facilities lean harder on mechanical cooling, which means more compressor hours, more heat rejection capacity and a bigger penalty for any equipment that is not maintained.
• Almost no free-cooling hours. Northern facilities shut off compressors for months. Here, the plant runs essentially year-round — so preventive maintenance intervals that are generous elsewhere are aggressive here, and deferred maintenance shows up faster.
• Salt air and corrosion. Coastal condensers and dry coolers degrade measurably faster. Coil coatings, wash schedules and material selection are not optional details near the water — they are the difference between a seven-year condenser and a fifteen-year one.
• Hurricane season. Outdoor heat-rejection equipment has to survive wind and debris, and the facility has to have a plan for cooling during a utility event, not just power. A generator that carries the IT load but not the full cooling plant buys you minutes.
• Humidity in the white space. Any outside-air path — door seals, makeup air, a propped door during a service call — introduces moisture that has to be removed before it condenses on something expensive.