The thermodynamics of AI data centers got quietly settled about eighteen months ago. Air cooling is finished for accelerator-dense workloads. Direct-to-chip liquid cooling is a transitional technology that will dominate the next five years and then hand off to two-phase immersion. The trade press is running several iterations behind on this story, because the iteration that actually matters happens in plumbing diagrams, and plumbing diagrams don't make press releases.
So here's what's settled, what's still contested, and why the holdouts are wrong. The holdouts deserve a respectful argument, because some of them carry legitimate operational concerns. Most of those concerns have answers. The few that remain are real, and they're still smaller than the operational benefits of getting past air.
The thermodynamics that ended the argument
Air at 25 degrees Celsius carries about 1.2 kilojoules per kilogram per kelvin of temperature delta. Water at the same temperature carries about 4.18, roughly 3.5 times more. A two-phase fluid sitting at the silicon interface and transitioning from liquid to vapor carries roughly 200 kilojoules per kilogram, because it's harnessing the latent heat of vaporization. That's something like 170 times the energy-carrying capacity of air per unit mass.
That number is the whole argument. When an accelerator dissipates 700 watts inside a volume the size of a sandwich, the working fluid pulling that heat out has to move it without letting the silicon climb past its throttle point. Air can't do that above a certain power density. Liquid can, and two-phase does it more efficiently than single-phase liquid does.
The mainstream commercial answer to this physics has been direct-to-chip liquid cooling. Cold plates sit on the hot components, and a secondary loop carries the heat away to a chiller plant. It works at scale. It also drags its own constraints along: the cold plates have to be matched to the specific accelerator package, the manifold has to be plumbed to every node, and the leak risk is real and produces dramatic failure modes when a leak finally happens.
Two-phase immersion sidesteps the lot. You submerge the entire compute node in a dielectric fluid. The hot components boil the fluid locally, the vapor rises and condenses on a heat exchanger above the bath, and it falls back as liquid. No plumbing to the chip. No cold plates, no manifold. The fluid does the work on its own.
What the holdouts say
There are four serious arguments against two-phase immersion. I want to take each one in turn, because they deserve specific responses, not a wave of the hand.
Argument one. The fluid is expensive and supply-constrained. True today. A typical engineered coolant for two-phase immersion runs $80 to $150 per kilogram, and a 10-foot container's worth of two-phase immersion uses 800 to 1,200 kilograms. That's a six-figure fluid cost stacked on top of the hardware. Supply sits with a small number of vendors running limited production capacity.
The answer is that the cost amortizes over a long deployment life, and the supply constraint is already solving itself. New suppliers are entering the market. The 3M historical monopoly on the most common fluids has loosened. Price per kilogram is falling at roughly 20 percent a year.
Argument two. The fluid is environmentally problematic. Partly true, and worth taking seriously. Some of the historical engineered coolants carry global warming potentials in the thousands. The newer fluids come in under 10. Fluid choice matters here more than almost anything else. The category is moving toward fluids that are dramatically less problematic than the first generation, and regulatory pressure is pushing that shift faster than market forces would on their own.
Argument three. The maintenance procedures are unfamiliar. True, and a genuine operational burden. Pulling a node out of an immersion tank is a different procedure than swapping a node in an air-cooled rack. Technicians have to be trained. Procedures have to be documented. Per replacement, the downtime is measurably longer.
The answer is that the higher per-replacement cost gets offset by a much lower replacement rate. Immersion-cooled nodes have shown MTBF improvements of three to five times over air-cooled nodes in equivalent thermal environments. Fewer fans to fail, fewer thermal cycling events, far less dust ingress. The total maintenance burden comes out lower, even though each individual maintenance event takes longer.
Argument four. The construction cost is higher. True at small scale and false at large scale. A pilot deployment of two-phase immersion carries higher capital cost per kilowatt than the equivalent air-cooled build. A production deployment at 50 megawatts or above carries lower total cost of ownership, because the cooling infrastructure simplifies so dramatically. No chiller plant. No CRAC units, no row-level CDUs. The construction looks more like plumbing than HVAC.
The crossover lands somewhere between 5 and 15 megawatts of deployed compute, depending on the climate and the energy cost. Below it, air or direct-to-chip liquid is still defensible. Above it, two-phase immersion is the rational choice, and most new AI data center capacity coming online in 2026 and beyond sits above the crossover.
What the supply chain is telling you
The supply chain has been telegraphing this for eighteen months. Hyperscalers are building immersion-compatible facilities. Fluid suppliers are expanding production. Accelerator vendors are publishing immersion compatibility validations. The reference designs from the major OEMs now ship immersion variants. The conference circuit is starting to treat immersion as the default rather than the exotic option you have to justify.
If you're a data center operator making a five-year capacity decision in 2026 and you're not seriously evaluating immersion, you're making a decision the rest of the industry has already moved past.
This isn't a fashion. It's a thermodynamic conclusion the industry has slowly but durably arrived at. The trade press lags because the people whose opinions actually matter aren't usually the ones writing for the trade press.
Why this matters for transportable AI
I'm writing this as a transportable AI vendor rather than a hyperscaler because the transportable case is where the argument cuts sharpest. A 10-foot container deploying serious AI compute can't use air. The thermal density won't fit the form factor, and the chiller plant won't fit at all. The transportable case forces immersion as the only viable option, and building transportable immersion deployments teaches lessons the larger fixed-site deployments are still working through.
Those lessons transfer. A hyperscaler that ignores them repeats the same mistakes the transportable vendors made and fixed two years earlier. The hyperscalers paying attention to the transportable space accelerate their own deployments by skipping the mistakes entirely.
A closing observation
Two-phase immersion is boring the way good infrastructure is boring. It works. It isn't exciting, and it isn't the next breakthrough. It's the engineering answer the field converged on after several years of chewing through alternatives. The vendors and operators who've already converged will spend the next five years quietly building out facilities that run better than the air-cooled ones they replaced. The ones who haven't will spend the next five years explaining to their finance teams why the cooling line item came in bigger than projected.
There's no glamorous case for immersion. There's also no defensible case against it at scale. The category has moved past the argument. The open question is no longer whether to deploy immersion. It's which fluid, which form factor, which integration partner.
That's a procurement question, not a strategy question. The strategy question got answered eighteen months ago.