Immersion Cooling Has the Best Numbers, So Why Isn't It Winning the AI Buildout?

Summary: Immersion cooling posts the best efficiency and density numbers in the industry, but direct-to-chip cold plate is winning the AI buildout on warranties, serviceability, and a hardening regulatory line against the fluids two-phase systems depend on. The spec-sheet advantage is real. What decides where immersion lands is everything the datasheet leaves out, including whether you can prove the numbers once the fluid is in the rack.

Put the major cooling architectures side by side on a spreadsheet and immersion cooling wins almost every column. Lower PUE, higher rack density, fewer moving parts. Then watch which one the hyperscalers actually order, and it is the architecture with the worse efficiency numbers. Direct-to-chip cold plate is quietly becoming the default for the AI buildout, while the technology that looks best on paper keeps losing the deal. The reasons have almost nothing to do with thermodynamics, and understanding them tells you a lot about how AI capacity really gets built.

What immersion cooling actually does better

Immersion cooling submerges live servers in a dielectric fluid, a liquid that does not conduct electricity, so the electronics can sit in it and keep running.

Single-phase systems pump that fluid through a heat exchanger. Two-phase systems let the fluid boil off the hot components and condense again on a coil overhead, exploiting the latent heat of vaporization for tighter thermal control. In both cases the fluid is in direct contact with the silicon, which is why heat capture approaches 100 percent.

Those numbers are hard to argue with. Vertiv reports immersion deployments delivering up to 100 kW of cooling per rack, with potential to cool racks approaching 900 kW, densities that air cannot touch and that even cold plate struggles to reach. Single-phase immersion commonly reports PUE in the 1.03 to 1.10 band, with two-phase pushing toward 1.02. Direct-to-chip typically lands between 1.15 and 1.30. Lawrence Berkeley National Laboratory’s Center of Expertise for Data Center Efficiency frames liquid cooling broadly as the path past the efficiency ceiling that air-cooled halls keep running into. On a spec sheet, submersion cooling wins every time.

Why direct-to-chip won the buildout

If immersion is so far ahead, why is the technology with the worse PUE taking the market? In direct-to-chip, coolant is piped across metal cold plates bolted to the CPU and GPU. The liquid never touches the board.

Liquid cooling has climbed toward 37 percent of deployments in 2026, up from roughly 3 percent in 2021, and direct-to-chip accounts for about 70 percent of that liquid-cooled share. The reason is compatibility. Cold plate drops into a more or less conventional server and a more or less conventional rack, and the major GPU platforms ship with direct-to-chip reference designs ready to go. IEEE Spectrum’s survey of data-center liquid cooling keeps arriving at the same operator conclusion.

The AI buildout is moving too fast to wait on an architecture that asks you to redesign the whole data hall. Cold plate is the path of least resistance, and at hyperscale, least resistance usually wins.

The warranty and serviceability tax on tank cooling

Most enterprise server warranties and OEM support contracts are written around air-cooled or direct-to-chip operation. Drop a server into a tank of dielectric fluid and, for many vendors, you are outside the supported operational specs, which can complicate or void the warranty. Operators end up shifting to immersion-specific servers and separate support arrangements before the first rack goes wet.

Then there is the day-to-day work. A failed drive in an air-cooled rack is a two-minute hot-swap. The same job in a tank means hoisting the server out of the fluid, letting it drain, working on it wet, and re-immersing it. That is a slower, messier workflow that needs trained hands and a place to do it. Two-phase systems add vapor management requirements on top. None of this shows up in the marketing literature, and all of it shows up in the operations budget.

The regulatory cloud over two-phase dielectric fluids

This next problem is more serious, because it threatens an entire branch of the technology rather than just its running costs. Two-phase immersion depends on engineered fluorochemical fluids, and those fluids fall under the umbrella of PFAS.

PFAS are the “forever chemicals” that regulators are moving to restrict, because they do not break down in the environment.

In December 2022, 3M, the dominant maker of the fluorinated fluids many two-phase systems were built around, announced it would exit all PFAS manufacturing by the end of 2025. The demand side is moving too. In 2023, five European countries submitted a universal PFAS restriction proposal to the European Chemicals Agency, one of the broadest chemical restrictions ever contemplated. For an operator weighing a multi-year, two-phase rollout, sourcing the fluid years from now is a genuine open question.

This is largely a two-phase problem. Single-phase systems typically run on hydrocarbon, synthetic, or bio-based dielectric fluids that sit outside the PFAS debate. The catch is that two-phase, the branch under the regulatory cloud, is the one posting the very best efficiency numbers.

How AKCP monitors liquid and immersion cooling

Here is the thread that runs through every point above. The numbers only count if you can see them on your own floor, and the moment you move to liquid, you inherit new failure modes that a spec sheet never mentions. That is a measurement problem, and it is the one AKCP is built to solve, whichever cooling architecture you land on.

Start with the fluid itself. Any liquid cooling loop, cold plate or immersion, now runs coolant close to live silicon, so leak detection stops being optional. AKCP supports spot and rope leak detection plus flow sensing, so a drip or a loss of circulation raises an alarm before it reaches the hardware, not after.

Then prove the efficiency claim. AKCP computes real-time PUE as a first-class virtual sensor, graphed and alertable like any other reading, so the 1.05 on the datasheet becomes a live number you can defend to finance instead of a figure you reconstruct once a quarter. For the racks that are still partly air-cooled, and in hybrid halls where liquid and air share the room, the Thermal Map sensor reports per-rack inlet and outlet temperatures with ASHRAE scaling and color heat maps, so you can see stratification and hot spots as they form.

Power gets the same treatment. A contactless current meter clamps around the conductors already feeding a rack and reads the draw with no rewire, no PDU swap, and no maintenance window, which matters when the racks are dense and you were never going to schedule an outage on them. Power Train maps the distribution path from the mainline down to the individual outlet, with per-outlet current and kWh, so high-density load stops being a guess. And because your cooling distribution units and other gear speak Modbus, SNMP, or MQTT, AKCP ingests them as virtual sensors, landing cooling and IT telemetry together in Quicklime DCIM. For the airflow you still have, sensorCFD runs an AI-assisted CFD simulation fed by live sensors and returns a thermal report you can act on.

We don’t just tell you where you have a problem, we tell you how to fix it. That is the same whether the heat leaves your racks through a cold plate or a tank of fluid.

What it means for operators planning AI capacity

Both architectures are going to keep shipping. Cold plate will handle the bulk of enterprise, cloud, and AI deployments because it is OEM-supported and serviceable. Immersion cooling earns its place in the specialized, ultra-dense corners where the density and PUE advantages outweigh the operational cost.

So the question for an operator is not which architecture wins on paper. Read the fluid datasheet and the support terms before the thermal spec, model the wet-maintenance workflow honestly, and if you are looking at two-phase, get a straight answer on fluid sourcing past 2025. Immersion cooling’s numbers are real. So are the reasons it is not the default, and those reasons, not the thermodynamics, are what will decide where it lands. Whichever way you go, the number that matters is the one you can measure on your own floor.