In the relentless push to tame the insatiable appetite of hyperscale datacenters, Microsoft has turned to liquid cooling to see what might stick at cloud scale.
The basic idea is simple: move beyond air and put the servers on ice. But in practice, the implementation of advanced cooling methods (cold plate, one-phase immersion, and two-phase immersion) is a study in trade-offs and unintended consequences.
Microsoft’s own study of the three options, outlined in Nature, lays bare the reality that every degree shaved off a CPU core temperature comes with a corresponding cost, whether in energy, water, or regulatory headaches.
And sadly, no single method holds the crown.
Cold plate cooling, also known as direct-to-chip cooling, is arguably the most conservative choice. It doesn’t ask datacenter operators to rip out racks or rethink layouts. Instead, it uses metal plates embedded with microchannels to siphon heat directly from processors and GPUs, like a high-tech, liquid-cooled band-aid.
It is effective. GHG emissions drop by 15 percent, energy demand shrinks by the same amount. Blue water consumption falls by 31 percent. And there are other benefits, too. Cold plates deploy without major disruption to air-cooled datacenters.
But as with all cooling options, there’s a catch: installing them is a different story. They require intricate plumbing and still leave hot spots that conventional air systems must handle.
"Cold plates deploy with less disruption to existing air-cooled datacentres. However, installing them can be complex, and air cooling may still be necessary," the study warns.
One-phase immersion cooling, on the other hand, invites operators to take the plunge.
Servers are submerged in tanks of hydrocarbon-based fluid, which has the advantage of being less hazardous than PFAS. Cooling happens by natural or forced convection, with the fluid carrying heat away and cooling in a loop.
The method is simpler than two-phase immersion, with fewer moving parts and no complex boiling cycles. But alas, it’s also less powerful. GHG emissions fall by 16 percent, energy demand by 15 percent and water consumption by 45 percent. Overclocking potential is limited to a modest 10 percent bump, lower than what cold plates or two-phase can achieve.
Yet in an era of increasingly stringent environmental regulations, one-phase immersion offers a safe bet.
"One-phase immersion uses cheaper coolants and simpler tanks than two-phase immersion but performs less effectively at the chip level," Microsoft’s researchers note. Less effective, yes. But safer in the long run.
Then there’s two-phase immersion cooling, the black sheep and the prodigy of liquid cooling.
This is the method that gets the engineers excited and the legal team on edge.
Servers get dunked in dielectric fluid, which boils off the heat and recondenses it like a miniature weather system in a box and the performance metrics are dazzling.
GHG emissions drop by 21 percent, energy demand by 20 percent and blue water consumption by 52 percent. Overclocking potential climbs by 20 percent, turning a dense server rack into a high-performance hot rod.
The downside? The fluids are PFAS. And PFAS are a problem.
"Two-phase immersion can support very high tank power densities (+500 kW) but uses polyfluoroalkyl substances (PFAS) that have been under legislative scrutiny," Microsoft’s report notes.
PFAS restrictions in the EU and the US could gut two-phase cooling before it even gets off the ground, leaving datacenters caught between the promise of radical thermal efficiency and the specter of future fines and retrofits.
Yet, for all the reasons above, the allure remains. For AI-heavy datacenters chasing higher power densities and wringing every ounce of compute from every square foot, two-phase immersion is the most seductive option. But it’s also the riskiest.
Microsoft’s study is less a blueprint than a set of hard truths for operators facing the big cooling dilemma.
Two-phase immersion is a regulatory gamble. One-phase immersion is the sensible, middle-ground play. Cold plates are the least disruptive but won’t cool tomorrow’s hyperscale servers when the heat gets serious.
"Our analysis has provided quantitative evidence of the sustainability aspects of relevant cooling technologies so they can be properly evaluated in decision-making," the report states, tying a bow on its findings.
We've pulled one of the most illustrative charts from the study to highlight the three options Microsoft explored:
Cold-plate cooling is a delicate operation in which every server wears a custom-fit metal jacket, each plate laced with veins of coolant—75% water, 25% polyethylene glycol—that siphon heat away from the electronics and out through a circulating loop. It’s an intricate, methodical system that looks deceptively simple: a grid of metal plates, a series of cooling lines, and a pump, all synchronized to keep CPUs from boiling over.
Then there’s one-phase immersion, a more visceral approach. Entire servers are dunked in vast tanks of mineral oil, their circuit boards bobbing in a viscous, dielectric sea. The heat flows from silicon to fluid, then off to the coolant loop. The simplicity is its strength—no complex chambers, no vapor management. Just immersion and convection, a big, wet heat sink.
But two-phase immersion is where it gets almost alchemical. Here, the servers are not just immersed; they are submerged in a liquid that boils at a whisper of a temperature. The heat doesn’t just diffuse; it erupts, transforming the liquid into vapor, which rises, condenses, and rains back down in a closed-loop microclimate. It’s a fever dream of controlled chaos, a boiling sea of tech that promises the ultimate heat rejection—if you can handle the PFAS-laden fallout.
And so, the choice isn’t between good and bad cooling methods. It’s a matter of which trade-offs you’re prepared to live with.