AI is digital in function, but its growth is increasingly constrained by physical reality. The faster AI compute expands, the more electricity enters the data center — and nearly all of that electrical energy eventually becomes heat.
AI Is Getting Hotter
The AI boom is usually described as a race for better models and more powerful accelerators. Yet behind every advance in large-scale AI is a less glamorous engineering question: how do we remove the heat?
As AI servers become denser and more powerful, cooling is no longer a background utility. It is becoming a core part of data-center architecture, capital planning and operating reliability.
The change can be seen at rack level. The IMMERSEKOOL AI Data Center Technical White Paper cites an average rack density of about 7.8 kW for conventional facilities outside the high-density segment. By contrast, new AI platforms are moving into the 100–200 kW class per rack, with future architectures expected to move still higher.
That is not simply “more electricity.” It means dramatically more heat inside nearly the same physical footprint.
A 200 kW Rack Is Also a 200 kW Heater
From a thermal point of view, a rack drawing 200 kW behaves almost like a 200 kW electric heater. The compute is useful, but the energy still has to leave the rack as heat.
This is why the cooling question changes as GPU density increases. A traditional server room can move moderate heat loads with conditioned air. A dense AI rack can demand a completely different thermal architecture.
Why Air Cooling Reaches a Practical Limit
Air has served the data-center industry for decades because it is simple, familiar and inexpensive. The limitation is its heat-carrying capacity.
The 2026 technical white paper illustrates the scale of the problem with a 200 kW rack. Using a 15°C air-temperature rise, the required airflow works out to roughly 24,300 CFM. Across a typical rack face, that corresponds to air velocity of around 10 m/s.
At that point, fan power, pressure loss, noise, airflow management and physical layout all become increasingly difficult. The issue is not that air suddenly stops working at one exact number. The issue is that the engineering and economic penalties rise rapidly as rack density increases.
Liquid changes the equation because it can transport far more heat in a much smaller flow volume. The same white-paper calculation shows that roughly 200 kW can be removed with water flow in the range of about 240 LPM at a 12°C temperature rise.
Liquid Cooling Is Not One Technology
“Liquid cooling” is often used as if it described a single solution. In practice, the market is developing across four major architectures:
- Direct Liquid Cooling — Single Phase: water or water-glycol remains liquid while carrying heat away through a cold plate mounted directly on the processor.
- Direct Liquid Cooling — Two Phase: a refrigerant boils inside a sealed evaporator or vapor cold plate, removing heat through latent heat.
- Immersion Cooling — Single Phase: the server is submerged in dielectric fluid that remains liquid and removes heat by convection.
- Immersion Cooling — Two Phase: dielectric fluid boils on the heated surface and the vapor is condensed back into liquid.
IMMERSEKOOL’s liquid-cooling portfolio is structured around these same four quadrants, covering cold plates, vapor cold plates, immersion heat sinks and two-phase boilers.
Single-phase direct liquid cooling currently has strong momentum in mainstream AI infrastructure because it can be integrated into conventional rack-based operating models. Two-phase technologies remain important because they offer a different path to very high heat flux, although refrigerant selection, sealing, service procedures and common interface standards require further development.
The Real Cooling System Is Bigger Than the Cold Plate
When liquid cooling is discussed, the cold plate often receives most of the attention. But a production AI rack is a complete fluid loop.
A typical path includes:
CDU → Rack Manifold → Hose Assembly → Quick Disconnect → Cold Plate → Return Line
The rack-level demonstration in IMMERSEKOOL’s product material makes this system view clear: manifold, server, quick disconnect, cold plate, CDU and hose kit operate as one connected thermal network.
This matters because every component influences the system. Pressure drop changes pump requirements. Hose routing changes serviceability. Air ingress can affect stability. A poor crimp or weak interface can create a leak risk. A manifold that does not match the rack layout can make an otherwise good cooling design difficult to install and maintain.
Fluid Interfaces Are Becoming Critical Infrastructure
One of the most underestimated parts of liquid cooling is the interface itself: the hose, fitting, adapter, manifold connection and quick disconnect.
In a dense rack, a small pressure loss at one connection may look insignificant. Multiply that loss across dozens of cooling branches and it becomes part of the total pump-energy and flow-balancing problem.
The same is true for leakage. Data-center liquid cooling demands repeatable assembly quality because the cooling loop operates beside expensive electronics. Reliability therefore depends not only on the headline components but also on the “last meter” of infrastructure that connects the rack together.
Standards Help — but They Do Not Standardize Everything
Open interfaces are essential to a scalable data-center ecosystem. They allow server, rack and cooling suppliers to work around common mechanical and performance requirements.
But a standardized coupling interface does not automatically standardize everything behind it. Hose stems, crimp sleeves, adapters, threads, hose length, bend radius and rack-specific routing still have to be engineered.
This is why pre-terminated hose assemblies are becoming more important. IMMERSEKOOL’s EPDM hose-kit approach combines rack-specific length and routing with pre-installed quick disconnects and leak testing before shipment.
Its published EPDM hose qualification data includes a working pressure of 10.34 bar, burst testing to 89.1 bar, an operating temperature range of −40°C to +100°C, UL 94 V-0 flame performance, and additional testing for cold bending, vacuum, adhesion, ozone exposure and coolant compatibility.
Rack Manifolds Are No Longer Just Plumbing
The manifold is becoming a central distribution layer inside the liquid-cooled rack. Port count, port pitch, orientation, coupling type and supply/return configuration must match the server and CDU layout.
That means the manifold increasingly belongs to the compute-platform design rather than to generic building plumbing.
The same principle applies to hose kits. A good hose assembly is not simply a piece of hose cut to length. Bend radius, connector orientation, routing space, identification, installation force and test procedure all affect the rack’s usability over its lifetime.
The Next Step: Physical AI Meets Data-Center Service
There is another reason the mechanical interface matters: automation.
Today, most server replacement and cooling-line maintenance is carried out by people. But future AI facilities may contain enormous numbers of accelerators and cooling connections. As maintenance becomes more automated, robots may need to grip, align, connect, disconnect and verify those interfaces.
That changes the design question. A hose-end sleeve may need a repeatable gripping surface. Connection geometry may need better self-alignment. Visual markings may need to support machine vision. A correct connection may need to be easy for a robot to verify.
This is where liquid cooling begins to intersect with Physical AI: the data center that runs AI may increasingly depend on physical automation to maintain the infrastructure that keeps AI running.
Cooling Is Now a Strategic Layer of AI Infrastructure
The AI market is often viewed through semiconductors, cloud platforms and model performance. But the next phase of competition will also be shaped by power delivery and thermal infrastructure.
Who can operate more compute within the same power envelope? Who can maintain dense racks reliably? Who can minimize cooling energy and service time while preserving thermal headroom?
These questions are no longer peripheral. They are becoming central to the economics of AI infrastructure.
AI may be digital in function, but its growth is increasingly constrained by physical realities — and cooling is now one of the most important.
Source basis: IMMERSEKOOL AI Data Center Technical White Paper 2026 and IMMERSEKOOL Liquid Cooling for AI Data Centers product and qualification materials. DATAAD editorial synthesis.