Signal & Supply
← Archive
November 20, 2026 COOLANT CHEMISTRY

Two-Phase Immersion Cooling Works by Letting an AI Chip Boil Its Own Coolant. The Only Company That Made That Coolant at Scale Just Stopped.

Next-generation AI chips were supposed to graduate to the most aggressive cooling trick available: submerging them in a liquid engineered to boil directly off the hot silicon. Almost all of that liquid came from one company. Right on schedule, at the end of 2025, it finished exiting the entire chemical family the liquid belongs to.

Key takeaway Two-phase immersion cooling isn't efficient because of clever plumbing — it's efficient because boiling a liquid pulls away far more heat, at a nearly fixed temperature, than simply warming one up. Only a narrow class of liquids can do that safely around bare chips: electrically non-conductive, non-flammable, and boiling at a precisely tuned 30–60°C. For decades those liquids came almost entirely from 3M's Novec and Fluorinert lines — both PFAS, the "forever chemicals" 3M committed in 2022 to stop making by the end of 2025. It confirmed in early 2026 that it hit that deadline. Replacement fluids exist on paper; none is a confirmed drop-in yet, so anyone already running two-phase systems is drawing down a finite stockpile of a liquid nobody is manufacturing anymore.

Drive the physics: boiling a chip cool vs. just warming it up.

Pick a cooling mode, set the chip's power draw, then press Run. This is a simplified teaching model, not real chip data — but the shape is real: two-phase cooling holds the chip at the fluid's boiling point (34°C, 3M's published figure for its Novec 7000 grade) across a wide power range, until heat flux exceeds the fluid's limit and the boiling film collapses ("dryout"), a real two-phase failure mode. The 95°C line is a common illustrative throttle threshold, not any specific chip's spec.

300 W
CHIP TEMP 25 °C
STATE Idle
SIM. TIME 0.0 s
COOLANT STOCK N/A (air)
Two-phase mode selected. Raise or lower chip power, then press Run to see the chip's temperature settle at the fluid's boiling point instead of climbing — right up until the power exceeds what the boiling film can carry away.

The plain version

Think about why sweating cools you down better than just sitting in a warm breeze. A breeze (air cooling) carries heat away slowly. Sweat works because evaporating liquid soaks up a huge amount of heat without your skin's temperature climbing much at all. Two-phase immersion cooling uses the same trick on AI chips: instead of air or a pumped liquid just getting warmer, the chip sits directly in a special fluid that's engineered to boil at a low, precise temperature right off its surface. The vapor rises, hits a condenser coil, turns back to liquid, and drips back down — no pump needed in that loop. Because boiling soaks up so much heat at a nearly constant temperature, this method can carry away far more heat per chip than blowing air or pumping ordinary liquid past it, which is exactly what the newest, hungriest AI chips need.

There's a catch most people cooling their laptop never think about: the fluid has to be electrically non-conductive, so a bare chip can sit directly in it without short-circuiting, and it has to not catch fire. For decades, essentially one company made fluids like that at serious commercial scale: 3M, selling them under the names Novec and Fluorinert. Those fluids belong to a chemical family called PFAS — "forever chemicals" that barely break down in nature and have been tied to environmental contamination and health concerns, and to billions of dollars in lawsuits against 3M. Back in 2022, facing that pressure, 3M announced it would quit making every PFAS product it sold, with a deadline of the end of 2025. In January 2026, the company confirmed, and independent reporting backed it up, that it hit that deadline on schedule.

That leaves every data center already running two-phase immersion cooling living off a dwindling stockpile of a fluid nobody makes anymore. Rival chemical makers, Chemours and Solvay among them, are racing out replacement fluids, but as of now none is confirmed as a true like-for-like swap — a new coolant has to survive years of direct contact with hundred-thousand-dollar chips before anyone will trust it at scale, and that kind of qualification takes time nobody budgeted for. Meanwhile, the newest AI accelerators keep drawing more power every generation, at exactly the moment the industry's most powerful cooling trick ran short of the one ingredient that made it safe.

The expert version

Two-phase immersion cooling exploits latent heat of vaporization rather than sensible heat. A dielectric fluid in direct contact with a chip package is formulated to boil at a temperature below the chip's maximum junction temperature — 3M's datasheet lists a 1-atm boiling point of 34°C for Novec 7000, one of several grades spanning roughly 30–60°C depending on formulation. Instead of the fluid's bulk temperature rising as it absorbs heat, it undergoes a phase change at a nearly fixed temperature, removing far more energy per unit mass for the same temperature delta than single-phase convective cooling. Vapor rises to a condenser coil, condenses, and returns by gravity: a closed loop requiring no pump in the primary circuit. This lets two-phase systems sustain much higher heat flux per unit chip area than air or single-phase liquid cooling, which matters as AI accelerator package power climbs past several hundred watts and rack densities exceed 100 kW.

The fluids that make this safe are engineered perfluorocarbons (3M's Fluorinert line) and hydrofluoroethers (3M's Novec line): dielectric, chemically inert, non-flammable, and precisely tuned in boiling point so bare silicon can sit directly in them without electrical shorting. Both chemistries are PFAS by regulatory definition — per- and polyfluoroalkyl substances whose carbon-fluorine bonds resist environmental breakdown, driving EPA scrutiny and multibillion-dollar litigation exposure. 3M announced in December 2022 that it would exit all PFAS, fluoropolymer, and fluorinated-fluid manufacturing by the end of 2025; the company's own PFAS-stewardship disclosures and independent reporting (Star Tribune, Fox 9) confirmed in January 2026 that the exit was completed on schedule, though 3M notes that eliminating PFAS from its own downstream products, as distinct from halting its own manufacture of PFAS, remains an ongoing, separate effort.

That removes the dominant commercial source of qualified two-phase dielectric fluid. Chemours' Opteon line, including 2P50, a hydrofluoroolefin fluid introduced in 2025 and marketed as fully PFAS-free, and Solvay's Galden, a perfluoropolyether that itself still meets the EPA's structural PFAS definition, are the leading replacement candidates. As of 2026, industry technical guides describe no commercially available fluid as a confirmed drop-in for the full Novec product range: qualifying a new dielectric against chip packaging materials, solder, and seals for long-term reliability typically runs from months to multiple years. Single-phase immersion cooling, using mineral or synthetic dielectric oils that never boil and carry no PFAS dependency, is unaffected by the shortage and continues to scale freely, creating a fork in liquid-cooling roadmaps between the two approaches.

Why it matters for tech + supply chain: The cooling method next-generation, highest-density AI racks were supposed to lean on hardest just lost its only proven fluid supplier — right as chip power keeps climbing and nobody has a confirmed replacement ready at scale.

Why it matters for tech + supply chain: Two-phase immersion's thermal advantage was never just engineering — it was downstream of a single company's chemical manufacturing footprint. Regulatory pressure on PFAS has now severed that footprint before replacement fluids have cleared the multi-year reliability qualification that chip-contact coolants require.

Get it in your inbox every morning