How Two-Phase Immersion Cooling (2PIC) Reduces Energy Costs for Data Centers
Modern AI infrastructure is pushing server rack power densities far beyond what traditional air cooling or liquid cooling was designed to handle. Some racks already pull up to a megawatt, and next-generation designs are pushing toward that ceiling fast. If your data center was built, or is being built, around conventional air or liquid cooling infrastructure, it’s worth rethinking your cooling strategy before energy costs get ahead of you.
Power density hasn’t spiked overnight. It has climbed gradually, one higher-output chip at a time, beyond what air- and liquid-cooling architectures can physically support. The data center thermal management industry has responded in turn with a progression of solutions: optimized airflow management, then direct-to-chip liquid cooling, then single-phase immersion, and now two-phase immersion. Each step changes where power is consumed, and where the real savings are.
Air Cooling & Direct-to-Chip Liquid Cooling
The Problem with Air Cooling
AI chips that draw 500 watts fit comfortably inside a standard air-cooled server and rack architecture. But a rack built to use one megawatt would need thousands of those chips in the same footprint, and air cooling cannot work effectively at that power-to-space ratio. Air cooling can handle roughly 30 to 40 kilowatts per rack before it runs out of room to work, a ceiling set by how little heat air can carry compared to liquid cooling.
The ARPA-E COOLERCHIPS program has set a public goal of cutting total cooling energy to under 5% of a data center’s IT load, pushing the industry further away from air cooling toward more efficient solutions.
What Cold Plate Solved, and What It Didn’t
Direct-to-chip cold plates were the industry’s first real answer. A single-phase fluid circulates directly over the chip surface to maximize the heat transfer out of the most power-dense component, while relying on air cooling to manage the rest of the server components. It’s a genuine fix for density, without a major shift in data center rack architecture, which is why most data centers building for AI today are already running some version of it.
A cold plate lowers power usage effectiveness (PUE) compared to air. But the fluid still has to reach the chip, which means pumps, manifolds, hoses, and quick disconnects routed through the rack, and each one of those is a joint that can fail, an added construction complexity, and a power inefficiency. Furthermore, a cold surface running below the dew point risks condensation on moisture-sensitive components, and circulating fluid at high volumes requires chemical knowledge, fluid management, and microchannels that can filter particulates and clog.
True, fluid chemistry matters just as much to the success of immersion cooling. But going to full immersion, single- or two-phase, entirely eliminates the costs and risks of routing pressurized liquid through thousands of chips, servers, and racks.
Single-Phase Immersion Closes the Gap
Single-phase immersion goes a step further than cold plate. Instead of routing fluid only to the hottest components, the entire server is submerged in a dielectric bath that never boils, and a pump circulates the fluid through an external heat exchanger to carry the heat away. That removes the per-chip plumbing failure points the cold plate still carries, and it’s a real jump in density over air or direct-to-chip alone.
The pump is still doing the work, though. Single-phase systems move heat only as fast as the pump can circulate the fluid, which is exactly why most operators skip this step entirely. Single-phase immersion takes on the same tank, the same dielectric fluid inventory, and the same new rack architecture as two-phase, without shedding the pump that two-phase eliminates. It asks for the infrastructure cost of immersion without its biggest efficiency win, which is why the market tends to jump straight from cold plate to two-phase instead.
Air Cooling | DTC Cold Plate | Single-Phase Immersion | Two-Phase Immersion (2PIC) | |
|---|---|---|---|---|
Cools all server components | Yes | No, chip only | Yes | Yes |
Requires a pump | No | Yes | Yes | No |
Fan power draw | Full | Reduced | None | None |
Typical PUE | 1.4–1.8 | 1.10–1.20 | 1.03–1.10 | ~1.02 |
How Two-Phase Immersion Cooling Reduces Energy Costs
Cooling Leaves the Room
Air-cooled and cold plate facilities still need a Computer Room Air Conditioner (CRAC) or Computer Room Air Handler (CRAH) running just to keep the data hall itself at a safe temperature, on top of whatever’s cooling the servers. Server fans push that heat straight into the room, and CRAC and CRAH units are what pull it back out.
Two-phase immersion removes most of that load. The dielectric fluid carries chip heat straight to a sealed condenser loop instead of into the room, so there’s almost no server heat left for a CRAC or CRAH to handle. Whatever capacity remains is there for the room itself, not the servers in it, so a smaller CRAC or CRAH runs far less often, and that’s less facility power going toward cooling a room that no longer has to fight server heat.
Boiling Replaces Pumping
Two-phase immersion cooling (2PIC) eliminates the pumping loop entirely. Server hardware sits submerged in a dielectric fluid engineered to boil at a controlled, low temperature. Heat from the chip turns that fluid to vapor, which rises, condenses against cooler condenser coils above, and falls back into the tank on its own.
The cycle runs passively on its own, with no pump working against pressure and no cycling on and off to catch up with the load. That’s what gets PUE below 1.03, close to its theoretical floor.
That fixed boiling point is also a ceiling the chip never crosses. Because the fluid maintains the chip at a controlled temperature independent of chip heat rejection, the junction temperature stays level under load instead of climbing with it, exactly what keeps the chip from ever needing to throttle with changes in demand.
The Fluid Cost Comparison
Novec™ 7100 and Novec™ 649 were the qualified fluids for two-phase immersion long before AI density made it a mainstream conversation, engineered for exactly this kind of controlled, low-temperature boil.
READ MORE: What Is Novec™? History, Where It’s Used & How to Find a Replacement
That history comes at a price, though. Two-phase dielectric fluids can cost more than single-phase options, but comparing them purely by volume misses the point. Two-phase cooling relies on the latent heat of vaporization, the energy absorbed when the fluid boils, which pulls exponentially more thermal energy per liter than the sensible heat transfer single-phase systems use. It also allows tanks to pack server blades more tightly together. That lets two-phase systems run on substantially less fluid per kW of compute, offsetting the upfront cost differential.
There’s a density problem underneath the cost problem, too. As data centers shift to AI workloads, the cooling bottleneck isn’t just total power, it’s heat flux density. Individual GPUs are scaling from 400 watts to over 1,000 watts. Single-phase systems struggle to manage this concentrated heat without pumping unsustainable volumes of fluid. Two-phase systems isolate that extreme density directly at the phase-change level, allowing facilities to cool next-generation AI hardware without expanding their fluid footprint.
Where the Savings Show Up
Lower PUE
A lower PUE means less total facility power for the same compute load, so a two-phase facility can push more compute through the same grid draw. That efficiency holds at any load, whether a facility runs near capacity or well under it, because there’s no fan speed or chiller cycle that has to chase the demand curve. Air-cooled facilities keep fans and chillers cycling even as compute demand drops, and their PUE also drifts with the seasons since it depends on outside air temperature. Two-phase PUE doesn’t move with the weather or the load.
PUE only measures the facility, not the server, so that number tells half the story. For instance, a PUE of 1.15 means a facility draws 1.15 kW from the grid for every 1.0 kW that reaches the rack, with the extra 0.15 kW lost to chillers, pumps, and power distribution. Two-phase immersion brings that number down near 1.02. Inside the chassis, a second gain appears that standard PUE accounting never accounts for: traditional servers spend up to 10% of their own power just spinning cooling fans. Submerge the server, and that fan load disappears, freeing that power to go straight to the chips instead. One gain happens at the facility level, and the other happens inside the server, so they stack instead of overlapping.
Independent data backs this up. In a side-by-side study of two 30-megawatt hyperscale data centers, Page Southerland Page, Inc. found the two-phase facility used 11,700 megawatt-hours of mechanical cooling energy per year against 19,600 for the air-cooled build, a 40.3% reduction, with annualized PUE dropping from 1.11 to 1.07.
No Throttling
Heat doesn’t just cost money to remove, it costs compute, too. PUE captures steady-state efficiency, but it goes quiet the moment heat becomes the bottleneck. With air and cold-plate cooling, a hot chip doesn’t throttle right away. It forces the fans to spin harder first, the same fan power that was already stealing from compute, now working overtime just to buy the chip a little more headroom.
That buys the chip a little time, but not much. When fan speed hits its ceiling, and the chip is still running hot, throttling kicks in anyway: the chip drops its own clock speed to protect itself, and every watt still flowing into it produces less useful work. Air and cold plate systems are especially prone to this under sustained AI training loads, where thermal headroom shrinks as ambient conditions or airflow degrade over a shift, so the facility pays twice: once for the fans, then again in lost throughput.
Two-phase immersion skips both steps. Junction temperatures stay pinned near the fluid’s boiling point regardless of load, so there’s no fan to ramp up and no clock speed to back off. Chips hold their rated performance, and the energy paid for at the meter reaches the silicon as usable compute instead of getting spent twice on the way there.
Tokens Per Watt
PUE alone doesn’t answer the question that matters most for a facility running AI workloads: how much useful output comes out of every watt drawn from the grid. The industry has started measuring that directly as tokens per watt, and it’s becoming the number operators get judged on ahead of PUE or raw rack density.
Two-phase immersion pushes that number in the right direction from every angle at once: lower facility overhead, no parasitic fan draw, no thermal throttling. That matters beyond the spec sheet, because data center power contracts are fixed. A facility built for 10 megawatts can’t call the utility and ask for 11 when demand goes up, so the only lever left is getting more usable compute out of the same 10 megawatts, which is exactly what determines dollars per token: what it costs to produce the compute being sold.
At a 1.15 PUE with 10% of server power going to fans, a facility spends roughly 1.28 kW of grid power for every kilowatt that actually reaches useful compute. At 2PIC’s 1.02 PUE with no fan tax, that drops to about 1.02 kW per kilowatt of compute, a reduction of roughly a fifth in grid power per unit of useful output. It’s the same 1.15-to-1.02 PUE shift and the same 10% fan recovery, just converted into the number that determines dollars per token.
Conclusion
Air and cold plate cooling are running into real limits: parasitic losses inside the rack, throttling under sustained load, and a power ceiling that only gets tighter as AI chips grow more powerful. That’s what’s pushing more data center builders toward two-phase immersion cooling. Making that shift means building with vendors already equipped to execute: tank manufacturers, integrators, and fluid suppliers with a proven track record of qualification.
On the fluid side, that history runs through Standard Fluids™, built around the team that developed Novec™ and Fluorinert™ at 3M™, the same fluids that qualified two-phase immersion cooling in the first place and the fluids most of the ecosystem still runs on today.
Curious what the math looks like for your own facility? Contact the Standard Fluids team to walk through what it would cost and what you would save as you scale.