There is a photograph that circulates in data-center engineering circles: a rack of servers fully submerged in a clear bath, columns of bubbles rising off the processors like a pot coming to boil. That is not a failure. That is the cooling system working exactly as designed. The liquid is an engineered dielectric fluid, the bubbles are the fluid flashing to vapor as it pulls heat off a chip, and the whole arrangement is the leading candidate for cooling the AI accelerators that now draw too much power to cool with air.

Ask an American which company owns the most US patents on that technique and the guesses come fast: Nvidia, whose GB200 racks pull roughly 120 kilowatts and started this fire. Microsoft, which has publicly dunked servers in boiling fluid at its own data centers. Amazon. Intel. All wrong.

The answer is Baidu.

One in eight

Between January 2020 and mid-June 2026, the US patent office granted 268 patents on immersion and two-phase cooling for servers and data-center racks. Baidu USA holds 36 of them. That is one in every eight grants in the category, and it is more than any other single company by a wide margin. Microsoft is second with 19. Then the drop-off is steep: Nvidia has 6, Intel 4, IBM 5 across its filings, Amazon 2, Google 2, Dell 2. Add up that entire roster of American hyperscalers and chipmakers and you land at 40 grants. Baidu, by itself, is at 36.

This is a category that barely existed a decade ago. Through 2014 it produced a handful of grants a year, mostly from mainframe-era thermal engineers at IBM and Fujitsu. Then it climbed: 44 grants in 2023, 51 in 2024, 69 in 2025, and 24 more by the middle of this year. The line tracks the power draw of AI silicon almost perfectly, because that is what is driving it. A rack of modern accelerators produces heat that a fan physically cannot move fast enough. You either put the chips in liquid or you throttle them.

Baidu saw this early. Its first grant in the set, from May 2020, describes a full immersion tank filled with “thermally conductive dielectric liquid” in which server blades sit completely submerged, with a heat-recovery unit spliced into the return line so the waste heat can be captured rather than dumped. Read the claims and the design intent is unmistakable: this was never a lab curiosity. It was infrastructure, specified down to the bypass valves.

The gravity trick

The reason this is a story and not a keyword count is that Baidu’s 36 patents share an unusually specific engineering thread. Roughly two in five carry “two-phase” or “phase change” in the title itself, and the rest lean the same direction in their claims. Two-phase is the harder, higher-performance cousin of plain immersion. Instead of relying on a fluid that merely warms up, two-phase cooling picks a fluid that boils at a low temperature right at the surface of the chip. The phase change from liquid to vapor absorbs enormous amounts of heat, and the vapor carries it away on its own.

The trouble with two-phase systems is that you now have to manage vapor, and vapor wants to escape. Baidu’s patents are a running argument about how to tame it. One 2024 grant describes a cooling plate bolted to server electronics with intermediate outlets machined along each fluid channel, so a bubble forming mid-channel can exit as vapor instead of choking the flow behind it. Another, granted in February 2025, does away with the pump entirely: vapor rises off the servers, a separator strips the vapor from the remaining liquid, and gravity alone recirculates the cooled liquid back down to the chips. A “passive internal return loop,” the filing calls it, pump-free by design. In a data center, a pump you can delete is a pump that cannot fail at 2 a.m.

Delete the shared label and these are not the same patent restated. A submerged tank, a micro-channel cold plate, a pump-free thermosiphon, a modular two-phase rack with a common vapor-return interface: they are distinct machines that happen to be solving the same physical problem from different angles. That is what a real portfolio looks like, as opposed to a pile of filings that all say “AI-powered” and mean nothing.

The man in the middle

Nearly all of it traces to one engineer. Tianyi Gao is listed as an inventor on the bulk of Baidu’s cooling grants, and his résumé reads like the technology’s own origin story. Gao earned his PhD in electronics-packaging and data-center thermal management at Binghamton University between 2012 and 2015, in the lab of Bahgat Sammakia, one of the most cited figures in the field and the director of Binghamton’s NSF-funded center for energy-smart electronic systems. Gao’s dissertation-era work, published in Applied Thermal Engineering, was on liquid cold plates for cooling high-density processors and GPU accelerators. He was building the plumbing for GPU-class heat loads years before “GPU-class heat load” became a phrase venture capitalists could say.

He carried that directly into Baidu USA’s Sunnyvale office, where he became a senior thermal engineer and, judging by the patent record, a one-man filing machine. The arc from a university heat-transfer lab to the single largest US immersion-cooling portfolio runs through one career. That is the recombination Steven Johnson would recognize: a niche academic specialty, incubated for a decade in an obscure corner of mechanical engineering, walks across the street to a search company right as the rest of the industry discovers it needs exactly that skill.

The catch

Here is where the story turns, because Baidu bet heaviest on the one flavor of immersion cooling whose supply chain just collapsed.

Two-phase immersion only works if you have the right fluid, one that boils at a precisely engineered temperature and does not conduct electricity. For years, essentially one product line filled that role: 3M’s Novec and Fluorinert fluids. In December 2022, 3M announced it would exit all “forever chemical” PFAS manufacturing by the end of 2025. The last orders shipped in early 2025, and production has now stopped. As Data Center Dynamics reported, 3M’s exit “vaporized the supply chain for two-phase immersion cooling.” Before it left, 3M is estimated to have supplied around 80 percent of the fluorinated coolant market, and its withdrawal has already caused shortages for chipmakers like Samsung and TSMC that use the same fluids in wafer chillers.

So the most valuable US patent moat in immersion cooling now surrounds a method whose working fluid you can no longer buy at scale. Chemours is racing to ship a replacement, Opteon 2P50, targeted for commercial volume this year. Solvay’s Galden line and AGC’s fluorinated solvents are being pressed into service. Two-phase vendors like ZutaCore and LiquidStack, which built their products around Novec, are scrambling to reformulate. The patents describe the hardware beautifully. Whether there is a fluid to pour into it is a separate question, and it is being answered right now in a handful of specialty chemical plants.

That tension is the interesting part for anyone deciding where to spend. A cold-plate design is durable intellectual property; a fluid formulation is a moving target under regulatory pressure. Baidu spent the early 2020s locking down the geometry of how boiling coolant moves through a server. The open race, the one still worth watching, is who supplies the liquid that makes the geometry matter.

Method note

Counts come from the US patent grant corpus, drawn from USPTO bulk grant data (roughly 9.3 million utility grants), searched for immersion and two-phase cooling combined with server, rack, electronic, or data-center terms, over grants issued from January 2020 through mid-June 2026. Assignee counts combine variant spellings and subsidiary filings; Baidu’s are held by Baidu USA LLC, its US research arm. The field trajectory reflects grant year, which lags filing by two to three years, so the recent surge understates how early the underlying work was done. Patent claims and abstracts were read directly. Career and supply-chain details come from Binghamton University, ResearchGate, and reporting by Data Center Dynamics; the 3M PFAS exit is from 3M’s own December 2022 announcement. Patent grants document what was filed and cleared, not what is deployed in production.