Microsoft microfluidic cooling
Last reviewed
Jun 2, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,774 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
Jun 2, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,774 words
Add missing citations, update stale details, or suggest a clearer explanation.
Microsoft's microfluidic cooling is an experimental chip-cooling technique, disclosed by Microsoft in September 2025, that etches tiny channels directly into the back of a silicon die and pushes liquid coolant through them so the fluid touches the silicon itself rather than reaching it through an external cold plate. In laboratory prototypes, Microsoft reported that the approach removed heat up to three times more effectively than conventional cold plates and reduced the maximum temperature rise of the silicon inside a GPU by 65 percent.[1][2] The channel layouts were optimized with AI in collaboration with the Swiss startup Corintis, producing branching, bio-inspired patterns that resemble the veins of a leaf. As of mid-2025 the work was a research and prototype effort rather than a deployed product.[1][3]
Microsoft announced the microfluidic cooling work through its Source publication on September 23, 2025, framing it as a response to the rising power density of AI accelerators.[1] The central idea is to bring coolant as close as physically possible to the transistors that generate heat. Instead of bolting a metal cold plate onto the top of a packaged chip, Microsoft's prototype carves microscopic channels into the back side of the silicon and circulates coolant through them, so the thermal path from the hot transistors to the liquid is measured in microns rather than millimeters.[1][2]
The company collaborated with Corintis, a startup based in Lausanne, Switzerland, on the design and validation of the channel geometry.[1][3] Microsoft reported two headline results from lab testing: heat removal up to three times better than cold plates, and a 65 percent reduction in the peak temperature rise of the silicon inside a GPU.[1] Microsoft cautioned that both figures vary with the type of chip and with workloads and configurations.[1]
The technique is not new in principle. Direct in-silicon liquid cooling traces back to research by IBM and academic groups decades earlier, and Microsoft itself had been experimenting with microfluidic channels for several years before the 2025 disclosure.[4][5] What Microsoft presented was an engineering effort to make the approach manufacturable and to pair it with AI-optimized channel routing.
Each new generation of AI accelerator draws more power and dissipates more heat in roughly the same physical footprint, which makes thermal management one of the limiting factors in data center design.[1] The dominant production technique, direct-to-chip liquid cooling with cold plates, attaches a metal plate to the lid of a packaged processor and runs coolant through the plate. The problem Microsoft identified is structural: the coolant in a cold plate is separated from the active silicon by several layers, including the chip lid and a thermal interface material, so heat must conduct through those layers before it reaches the fluid.[1][2]
Microsoft argued that this stacked structure is becoming a bottleneck. Sashi Majety, a senior technical program manager at the company, said that within about five years a data center relying heavily on traditional cold-plate technology would be "stuck," because the insulating layers between coolant and silicon cap how much heat can be pulled away.[1] A cold plate also applies cooling uniformly across the top surface of a chip, whereas heat on a real die is concentrated in localized hot spots tied to whichever circuit blocks are active for a given workload.[2][3]
Microfluidic cooling removes the intervening layers by putting the coolant inside the chip. In Microsoft's prototype, channels roughly the width of a human hair are etched into the back side of the silicon die, and coolant is routed through those channels directly across the silicon.[1][2] Because the liquid contacts the silicon itself, the interface layers that a cold plate relies on are largely eliminated, shortening the thermal path between the transistors and the fluid.[2]
The engineering challenge Microsoft described centered on channel depth. The channels must be deep enough to carry sufficient coolant and to avoid clogging, yet shallow enough that they do not weaken the silicon and cause mechanical failure.[1][6] Microsoft said it worked through this balance along with leak-proof packaging, the choice of coolant, the etching method, and integration into a manufacturing process, iterating through four design generations in the year before the announcement.[1]
The following table summarizes how Microsoft contrasted the two approaches.
| Aspect | Conventional cold plate | In-silicon microfluidic cooling |
|---|---|---|
| Coolant location | External metal plate on top of chip package | Channels etched into the back of the silicon die |
| Thermal path to silicon | Through chip lid and thermal interface material | Coolant contacts silicon directly |
| Cooling distribution | Uniform across the chip surface | Routed toward specific on-die hot spots |
| Channel scale | Not applicable | Channels roughly the width of a human hair |
| Status (mid-2025) | In production across the industry | Laboratory prototype |
Sources: Microsoft Source; Tom's Hardware.[1][2]
Rather than running coolant through simple straight, vertical channels, Microsoft and Corintis used AI to optimize the channel layout so that fluid is steered toward the regions of the die that run hottest for a given workload.[1][3] The resulting geometry is biomimetic: Microsoft described it as resembling the veins in a leaf or the structure of a butterfly wing, branching patterns that distribute fluid efficiently across a surface.[1][2] Microsoft tested both the bio-inspired layouts and straight vertical channels, and reported that the optimized branching designs cooled hot spots more effectively.[1]
Corintis specializes in this kind of design. The company describes building networks of microscopic channels, with some passages as narrow as about 70 micrometers, and it has used additive manufacturing to produce copper cooling components with such channels.[5] In 2025 Corintis raised a $24 million Series A round and stated it had produced more than 10,000 copper cold plates, with a target of one million by the end of 2026, indicating an effort to commercialize microfluidic cooling more broadly than the Microsoft prototype alone.[5]
Microsoft attributed the following figures to its own laboratory prototypes, and noted that the values depend on the chip and the workload.[1]
| Reported metric | Value | Comparison baseline | Source |
|---|---|---|---|
| Heat removal effectiveness | Up to 3x better | Conventional cold plates | Microsoft Source[1] |
| Reduction in peak GPU silicon temperature rise | 65 percent | Conventional cold plate operation | Microsoft Source[1] |
In its testing, Microsoft ran the microfluidic cooling on a server executing the core services for a simulated Microsoft Teams meeting, an internal workload chosen to represent realistic, variable demand.[1] A separate IEEE Spectrum account of the Microsoft and Corintis work cited heat removal roughly three times as efficient as existing methods and described chip temperatures lowered by more than 80 percent relative to air cooling, a different baseline from the 65 percent figure Microsoft reported against cold plates.[5] Secondary coverage from Tom's Hardware, HPCwire, and Data Center Frontier reported the same up-to-3x and 65 percent figures attributed to Microsoft.[2][6][7]
The work was presented as research and engineering, not as a shipping product. Microsoft said it was investigating how microfluidic cooling could be incorporated into future generations of its own chips, and that bringing the technique to production would require further work with fabrication and silicon partners.[1] Coverage by Data Center Frontier likewise characterized the results as coming from prototype testing under specific workloads and configurations, with no production deployment timeline and no published fleet-level reliability data.[3]
Several obstacles stand between the prototype and broad deployment. Microsoft and outside observers pointed to manufacturing yield and throughput, the lack of industry standardization or multi-vendor commitments to produce microfluidic-ready silicon, and the difficulty of guaranteeing leak-free operation when coolant runs inside the chip package.[1][3] Etching channels into the die also has to be reconciled with the structural integrity of the silicon and with existing chip-packaging processes.[1][6] Because the channels are cut into the silicon itself, the approach is most naturally suited to chips designed for it from the start rather than retrofitted onto existing parts.[3]
Microsoft positioned microfluidic cooling as a way to lift several constraints at once. Judy Priest, the company's corporate vice president and chief technical officer for cloud operations and innovation, said microfluidics would allow more power-dense designs, enabling more capability in a smaller amount of space.[1][3]
The company described three broad implications. First, denser servers: removing more heat per chip would let operators pack hardware more tightly without hitting thermal limits.[1] Second, three-dimensional chip stacking, where cooling could flow between vertically stacked silicon layers, an architecture that is difficult with surface-only cold plates and relevant to memory-heavy designs such as those using high-bandwidth memory.[1] Third, sustained higher performance, sometimes described as a form of overclocking, in which a chip could safely run at elevated power for bursts because the cooling can absorb the extra heat without damaging the silicon.[1] Microsoft also noted that the ability to run with warmer coolant could reduce the energy a data center spends on cooling and improve its power usage effectiveness.[1][3]
The reported figures come from Microsoft's own prototypes and have not been independently reproduced at scale, and Microsoft itself stated that the results vary by chip type, workload, and configuration.[1] The 65 percent and up-to-3x numbers are best-case laboratory results rather than guaranteed production performance. The IEEE Spectrum "more than 80 percent" figure uses air cooling as its baseline, so it is not directly comparable to the cold-plate comparison and should not be conflated with it.[5] No commercial product, ship date, or named chip using in-silicon microfluidic cooling had been announced as of mid-2025, and significant questions about manufacturing scalability, standardization, and long-term reliability remained open.[1][3]