The cooling crisis in AI data centres has moved from an engineering challenge to an infrastructure emergency in the space of eighteen months. The shift from conventional server racks drawing five to ten kilowatts per rack to GPU clusters drawing fifty to one hundred kilowatts per rack has broken the fundamental assumption underlying most data centre cooling design: that air cooling -- moving chilled air through the facility to absorb heat from equipment surfaces -- can remove heat at the rate it is generated. It cannot, at the densities that AI training and inference infrastructure requires. The consequence is that data centre operators who built or bought facilities designed for conventional compute are finding them unsuitable for AI workloads, and the operators investing in new AI-native facilities are discovering that the cooling infrastructure can be as costly as the compute infrastructure it is protecting.
Industry announcements in October 2026 have underlined the commercial urgency. Multiple hyperscale operators have disclosed that cooling infrastructure is the critical path item in their AI data centre expansion programmes -- not chip supply, not power connection, but the physical infrastructure to remove heat from GPU clusters at the density and reliability that AI workloads require. The engineering solutions exist: liquid cooling in various forms has been demonstrated at scale. The challenge is deployment speed, cost, and the operational complexity of managing cooling infrastructure that is fundamentally different from the air systems that facility management teams have operated for decades.
The Physics of the Problem
Air cooling works by passing chilled air over equipment surfaces, absorbing heat, and exhausting that heat to the outside via cooling towers or air handling units. The heat removal rate of air cooling is limited by the heat capacity and flow rate of air -- and air, being a gas with low density and low specific heat capacity, is not an efficient heat transfer medium at very high power densities. At rack power densities above twenty to thirty kilowatts, the volume of air required to remove heat exceeds what conventional data centre airflow management can deliver without significant energy penalty and floor space consumption for air handling infrastructure.
GPU clusters for AI training regularly operate at fifty to one hundred kilowatts per rack. The latest Nvidia GB200 NVL72 rack system is rated at one hundred and twenty kilowatts. Deploying these systems in facilities designed for air cooling at ten kilowatts per rack creates a physics mismatch that facility engineering cannot solve incrementally -- it requires a fundamentally different cooling approach. The industry has two primary options: direct liquid cooling, which brings chilled liquid directly to the server level via manifolds and cold plates; and immersion cooling, which submerges server hardware in dielectric fluid. Both are far more efficient at heat removal than air cooling. Both require infrastructure investment that fundamentally changes the design and cost profile of data centre facilities.
Direct Liquid Cooling: The Near-Term Path
Direct liquid cooling is the approach most rapidly being deployed in high-density AI facilities. Cooling Distribution Units -- CDUs -- connect to building chilled water systems and distribute liquid through manifolds to server-level cold plates attached to CPUs and GPUs. The heat is absorbed by the liquid directly at the chip surface, where temperatures are highest, and the warmed liquid returns to the CDU for cooling. Direct liquid cooling achieves substantially better thermal efficiency than air cooling and can support the rack power densities that current generation AI hardware requires.
The infrastructure investment for direct liquid cooling includes the CDUs themselves, the in-rack and in-row piping manifolds, the cold plates that must be matched to specific server and GPU hardware, and the building-level water distribution systems. For existing facilities converting from air to liquid cooling, the retrofit cost is substantial -- estimates vary but typically run to hundreds of thousands of dollars per rack row for a fully plumbed deployment. For new facilities designed from the start for liquid cooling, the infrastructure integration is more cost-effective but still represents a significant capital premium over air-cooled design.
The operational challenge of liquid cooling at scale is water management -- leaks in a data centre carrying liquid to within millimetres of electrical equipment are a catastrophic failure mode that air cooling does not have. The industry is developing engineering standards, connector specifications, and monitoring systems to manage this risk, but the operational discipline required is new to most data centre facility teams. Training, standardisation, and tooling are all still maturing, which is a constraint on deployment speed even when engineering and capital are not.
Immersion Cooling: The Long-Term Premium Segment
Immersion cooling achieves the highest thermal efficiency and the highest rack density of any current cooling approach. Single-phase immersion immerses servers in non-conductive mineral oil or engineered fluid, removing heat as the warmed fluid circulates to a heat exchanger. Two-phase immersion uses engineered fluids that boil at low temperatures, removing heat through phase change -- the most efficient heat transfer mechanism available. Two-phase systems from Submer, Asperitas, and LiquidStack can support rack densities exceeding two hundred kilowatts and are being specified for the most demanding AI training installations.
Immersion cooling has a higher upfront capital cost than direct liquid cooling and requires purpose-designed tanks rather than standard server racks. Hardware vendors have been slow to validate their equipment for immersion -- warranty coverage for immersion-installed servers has been a limiting factor. But the hardware vendor position is changing: Nvidia has published immersion cooling guidelines for its H100 and H200 GPUs, and OEM server manufacturers are designing systems intended for immersion deployment rather than treating it as an aftermarket modification. As hardware validation improves, immersion will move from a niche solution for the most extreme density applications to a mainstream option across high-performance AI facilities.
Energy Efficiency and the Water-Power Trade-off
Liquid cooling's superior thermal efficiency translates directly to improved data centre Power Usage Effectiveness (PUE) -- the ratio of total facility power consumption to IT equipment power consumption. Air-cooled data centres typically achieve PUE values between 1.3 and 1.6, meaning thirty to sixty percent of total facility power is consumed by cooling infrastructure. Liquid-cooled facilities can achieve PUE values below 1.1, because the thermal efficiency of liquid cooling allows waste heat to be rejected at lower temperatures with less auxiliary power consumption. At the scale of a large AI data centre consuming hundreds of megawatts of IT power, the difference between PUE of 1.5 and 1.1 represents tens of megawatts of power saved and millions of dollars in annual energy cost reduction.
Water consumption is the trade-off. Direct liquid cooling systems that use evaporative cooling towers to reject heat consume significant volumes of water -- a concern in water-stressed regions. Immersion cooling with dry coolers or heat exchangers can achieve very low water consumption. The trade-off between electricity efficiency and water consumption is becoming a key specification criterion for data centre operators in regions with water scarcity concerns, and is driving adoption of heat exchanger-based rejection systems that sacrifice some electrical efficiency for water savings. The ability to reuse waste heat -- injecting it into district heating networks or using it for process heat in adjacent facilities -- is also being evaluated at scale, creating the possibility that AI data centre heat becomes a resource rather than a disposal problem.
Top 10 Companies in This Space
- Vertiv Holdings -- critical digital infrastructure supplier, CDU and liquid cooling distribution equipment for enterprise and hyperscale data centres
- Schneider Electric -- data centre infrastructure and cooling systems, EcoStruxure platform managing hybrid air and liquid cooling environments
- Stulz GmbH -- precision cooling specialist, high-density data centre cooling systems and CDU deployment
- Submer Technologies -- two-phase and single-phase immersion cooling systems for high-density AI compute
- LiquidStack -- two-phase immersion cooling systems, partnerships with hyperscale and colocation operators for AI workloads
- Asperitas BV -- Dutch immersion cooling specialist, natural convection single-phase systems for efficient AI cluster cooling
- Iceotope Technologies -- chassis-level precision immersion cooling, targeted at highest-density GPU rack deployments
- Allied Control (Vertiv) -- two-phase immersion cooling technology, HPC and AI data centre focus
- CoolIT Systems -- direct liquid cooling specialist, CDU and cold plate systems for GPU workloads
- Equinix Inc -- global colocation provider, investing in liquid cooling infrastructure across its International Business Exchange data centre network
Our Take
Datacentre cooling has become one of the most active capital deployment areas in technology infrastructure, and the opportunity set is real and growing. The physics of AI compute -- GPU clusters at densities that air cooling cannot handle -- creates unavoidable demand for liquid cooling infrastructure that cannot be met by incremental improvements to existing air systems. Vertiv and Schneider Electric are the large-cap expressions of this thesis, with established positions in data centre infrastructure that are being extended into liquid cooling. The pure-play immersion cooling companies -- Submer, LiquidStack, Iceotope -- offer higher growth potential with higher execution risk as they compete for share in a market that is still standardising. The energy efficiency benefit of liquid cooling is the business case that facility operators understand best: the PUE improvement pays back the infrastructure premium in electricity savings over relatively short timescales at today's energy prices. The near-term constraint on deployment speed is operational capability -- trained workforces and validated procedures. Expect that constraint to ease as the industry matures its standards over the next eighteen months. -- Elena Vasquez, MarketsNXT