Why the rack forced the decision
The mechanism is simple. Every watt a chip consumes ends up as heat in the room, and the cooling system has to move it out. NVIDIA's GB200 NVL72 packs 36 Grace CPUs and 72 Blackwell GPUs into one rack that NVIDIA describes as a rack-scale, liquid-cooled design. Estimates of its power cluster around 120 kW, and HPE's specification for its version lists 132 kW nominal. The next generation goes further: supply-chain reports put Vera Rubin NVL72 racks at roughly 190 to 230 kW.
Set that against the installed base. Uptime Institute's 2026 survey puts the average rack at about 7.8 kW and the most common at 11 kW, as reported by Data Center Knowledge. The IEA says AI server power density rose about 11 times between 2020 and 2025 and expects another fourfold rise by 2027. Air cooling scales with how much air you can move, and nobody can move fifteen times more air through the same rack. Water carries vastly more heat per unit volume than air, so a modest flow through a cold plate does the job.
The options, compared
| Approach | How it works | Practical range | Strengths | Weaknesses |
|---|---|---|---|---|
| Rear-door heat exchanger | Water-cooled coil replaces the rack's rear door; server fans push air through it | Up to ~75 kW per rack (Motivair ChilledDoor rating) | No server changes; retrofits existing halls | Still air-cooled at the chip; cannot handle 120 kW rack-scale systems |
| Direct-to-chip, single-phase | Cold plates on GPUs/CPUs; water-glycol loop to a CDU | 100 kW+ per rack; the design point for GB200 NVL72 | What NVIDIA and server OEMs ship; broad vendor base | Residual heat still needs air; leak management and water chemistry become your problem |
| Direct-to-chip, two-phase | Refrigerant boils in the cold plate, carrying heat as vapour | Aimed at the highest-power chips | High heat flux; non-conductive fluid | Smaller vendor base; refrigerant regulation |
| Immersion, single-phase | Servers submerged in a tank of dielectric oil | Very high; captures nearly all heat | No fans; very uniform temperatures | Hardware support and warranties; heavy tanks; servicing is messy |
| Immersion, two-phase | Servers submerged in fluid that boils and recondenses | Very high | Excellent heat transfer | Fluid cost and regulatory pressure on some fluorinated fluids; niche |
Rear-door heat exchangers: the bridge
A rear door swaps the back of a standard rack for a water-cooled coil. The servers do not change, which is why it is the usual first step for an existing hall. The ceiling is the issue: Motivair, one of the established makers, rates its ChilledDoor at up to 75 kW per rack. That covers a dense rack of PCIe GPU servers. It does not cover a rack-scale Blackwell system.
Direct-to-chip: the default
Cold plates sit on the GPUs and CPUs, coolant flows through them, and a coolant distribution unit (CDU) hands the heat to the building's facility water through a heat exchanger. The CDU also holds the coolant at a set temperature and pressure and, per the Open Compute Project's CDU methodology, keeps it at least 2 °C above the dew point so nothing condenses on the electronics. Memory, drives and power supplies are often still air-cooled, so these rooms still have air handling, just much less of it.
Direct-to-chip won the default for a supply-chain reason, not a physics one: it is what NVIDIA and the server OEMs design, test and warranty. That means spare parts, trained technicians and integrators who have done it before. If you are deploying GPU hardware in 2026, this is the path with the fewest surprises.
Immersion: real, but a niche
Immersion submerges whole servers in dielectric fluid, either oil that stays liquid (single-phase) or a fluid that boils off the hot components and recondenses (two-phase). It removes nearly all the heat and all the fans. The catch is everything around it: hardware validated for immersion, warranty terms, the weight of the tanks, and maintenance that involves lifting dripping servers out of a bath. It fits operators who control the whole stack. For most enterprise and colocation deployments of NVIDIA rack-scale systems it is not the supported route.
Water temperature changes the building
One underrated benefit of cold plates is that they work with warm water. ASHRAE's liquid-cooling classes run up to W45, meaning facility water supplied at about 45 °C, which can be cooled by a cooling tower or dry cooler without mechanical chillers. Vendors are now building CDUs for that range: DCX, for example, announced an 8 MW CDU for 45 °C water aimed at Vera Rubin deployments. Fewer chillers means less energy, and with dry coolers it can mean little water. The water side of that trade-off is covered in AI data center power and water use.
Vendors after the consolidation
Two years ago liquid cooling was a field of specialists. It is now mostly owned by large industrial and HVAC companies:
- Schneider Electric completed its purchase of a 75% controlling stake in Motivair, a CDU and rear-door maker, in February 2025, after agreeing the deal in October 2024 at a reported $850 million.
- Ecolab agreed in March 2026 to buy CoolIT Systems from KKR-managed funds for about $4.75 billion in cash; CoolIT makes cold plates and CDUs.
- Trane Technologies announced in February 2026 an agreement to acquire LiquidStack, which sells direct-to-chip and immersion systems.
- Vertiv bought CoolTera, a coolant distribution specialist, in 2023, and Flex bought JetCool, a cold-plate maker, in 2024.
- Independent and immersion specialists, including nVent for CDUs and Submer and GRC for immersion, remain in the market.
For a buyer, consolidation cuts both ways. Bigger parents mean better balance sheets and service networks. They also mean fewer independent options and lead times set by the same factories everyone else is waiting on.
What to ask a colocation provider
- What is the maximum supported density per rack today, with liquid, and in which hall?
- Is the CDU yours or mine, who maintains it, and what is the response time for a leak?
- What facility water temperature do you supply, and does it match the hardware's specification?
- How is heat rejected: evaporative towers, dry coolers or chillers? What is the site WUE?
- Have you deployed this exact rack-scale system before, and can I talk to that customer?
The last question filters most of the market. Many providers sell "AI-ready" space that has never held a 120 kW rack. If you would rather not answer these questions yourself, renting finished capacity from a neocloud moves the cooling problem to someone who has.
Related
- AI infrastructure — the cluster hub
- Power and water use — what cooling choices do to the water bill
- The largest AI data centers — where GB200-class racks are being installed at scale
- Neoclouds compared — renting liquid-cooled capacity instead of building it
Frequently asked questions
Why do AI data centers need liquid cooling?
Because the newest AI racks produce more heat than air can carry away. NVIDIA's GB200 NVL72 puts 72 GPUs in one rack drawing roughly 120 to 132 kW, and it is designed as a liquid-cooled system. Water and other coolants carry far more heat per unit volume than air, so cold plates on the chips can remove that heat with modest flow, where air would need impractical volumes and fan power.
What is direct-to-chip liquid cooling?
Direct-to-chip (also called direct liquid cooling or cold-plate cooling) mounts metal cold plates on the hottest components, mainly GPUs and CPUs, and pumps coolant through them. A coolant distribution unit separates that sealed loop from the building's facility water through a heat exchanger and controls flow, pressure and temperature. The remaining components are still air-cooled, so a direct-to-chip room still needs some air handling.
Immersion vs direct-to-chip cooling: which is better?
Immersion submerges entire servers in a dielectric fluid and captures nearly all of the heat, with no fans. Direct-to-chip captures most of it and leaves the rest to air. Immersion can be more thermally efficient, but the major accelerator platforms are designed and supported for cold plates, so direct-to-chip is the lower-risk default for GPU clusters in 2026. Immersion makes most sense where you control the whole stack, for example in edge or crypto-derived sites.
What is a rear-door heat exchanger?
A rear-door heat exchanger replaces the back door of a standard rack with a water-cooled coil. Server fans push hot air through the coil, which removes the heat before it enters the room. It needs no changes to the servers, which makes it a common retrofit. Motivair rates its ChilledDoor at up to 75 kW per rack, well short of a GB200 NVL72 rack.
Does liquid cooling use more water?
Not necessarily. The coolant loop to the chips is closed. Water consumption depends on how the building rejects heat: evaporative cooling towers consume water, dry coolers do not but use more energy. Liquid cooling can run with warm facility water, around 45 °C in the ASHRAE W45 class, which makes chiller-free and lower-water designs easier.
Sources: NVIDIA; HPE; Uptime Institute via Data Center Knowledge; IEA (April 2026); Open Compute Project; Motivair; Schneider Electric; Ecolab via Business Wire; Trane Technologies; DCX via Business Wire. Rack power for future platforms is as reported, not confirmed by NVIDIA.