Microsoft Brings Microfluidics To Datacenter Cooling With 3X Performance Gain

Microsoft has successfully tested a microfluidic cooling system that removed heat up to three times better than cold plates currently used in datacenters. The technology etches tiny channels directly into silicon chips, allowing cooling liquid to flow directly onto the heat source. In lab tests announced September 23, 2025, the system reduced the maximum temperature rise inside GPUs by 65%. The channels, roughly the width of human hair, were optimized using AI to create bio-inspired patterns resembling leaf veins.

Microsoft collaborated with Swiss startup Corintis on the design. The cooling fluid can operate at temperatures as high as 70C (158F) while maintaining effectiveness. The company demonstrated the technology on servers running Microsoft Teams services, where the improved cooling enables overclocking during demand spikes that occur when meetings start on the hour and half-hour. Microsoft is investigating incorporating microfluidics into future generations of its first-party chips as the company plans to spend over $30 billion on capital expenditures this quarter.

Point of interest: Failure modes

[Gravis Zero]

Liquid cooling has been around forever but what matters a lot is how often it fails and how. If this was a silver bullet then MS wouldn't be spilling the beans but rather, promoting that the benefits of their secret technology and that you should invest in them. There is no benefit to them sharing this information which leads me to believe it's another pipe dream.

Re:

[Anonymous Coward]

Microfluidic channels aren't a new idea and people have discussed their possibility for years. It's not a "secret." This is microsoft saying they tested the concept and found it to work.

Re:

[PPH]

Microsoft isn't a hardware company. Their primary interest is to convince others to use the technology.

Re:

[bloodhawk]

Microsoft IS a hardware company, despite most of their money coming from software they have massive investment in hardware development that dwarfs most purely hardware oriented companies. Hardly suprising considering the hundreds of billions they spend on Datacenters and servers.

Re:

[PPH]

For those data centers, they are the customer. How much hardware do they sell? Pretty much just the XBox.

Re:

[bloodhawk]

They sell, Xbox, accessories, mice, keyboards. Their surface revenue is now around $7B. They also had a multi billion dollar hololens business (which obviously they are now exiting). For those Datacentres they are both the customer AND the developer, the hardware is mostly custom specs and boards similar to how Google and AWS do it, they develop the machines then farm out their production to vendors.

Re:

[thegarbz]

That is non-sequitur. Either they are a hardware company so they can get others to use the technology, or they aren't a hardware company and then why would they have interest in getting others to use tech?

Your post is ignorant of how these companies work. Not only does Microsoft have a consumer hardware division, but like all major tech companies with cloud infrastructure they develop a shitton of their own hardware for their own purposes, and this is one of them. MS is developing this for MS, not for other

Re:

[fahrbot-bot]

... which leads me to believe it's another pipe dream.

Or "heat pipe dream" anyway. :-)

Which is easier?

[MachineShedFred]

So which is easier for Microsoft: pushing the state of the art in thermal transfer by etching fluidic channels into hardware and developing / testing / productionalizing "microfluidics" for moving heat away from that hardware... ... or making Teams less shitty and bloated, making the hardware and cooling you already have sufficient to the task without having to play stupid games with variable overclocking at the top and bottom of the hour to keep up?

I guess we now know, since we see which direction they dec

Re:

[Locke2005]

My experience at Microsoft is that they are much better at buying hardware that works than they are at designing hardware themselves that works. They have better software talent than hardware talent. That being said, I'm not the person to judge how great the Microsoft “butt hinge with butt straps” patent is... (That's a real patent, but apparently it was filed with the wrong company name.)

Microsoft hardware

[algaeman]

Their mice are very good, but I've never had any problems with them overheating.

Re:

[Locke2005]

Microsoft mice and keyboards are now made by Incase, not Microsoft. I have like the Microsoft mouse and keyboard in the past, but like I said, I'm pretty sure they bought the rights to those designs from someone else. My only problem with Microsoft keyboards have been related to "Pepsi syndrome", i.e. they stop working when you spill champagne on them.

Re:

[Xenx]

The alternative to dynamic overclocking is to have more, or more powerful, servers to handle the peak load. This problem exists regardless of how (un)optimized you believe their code to be.

Re:

[thegarbz]

What makes you think it's one or the other.

Step 1. Build cloud infrastructure
Step 2. Build a CPU intensive app.
Step 3. Make cloud infrastructure efficient to run to keep cost down through better cooling.
Step 4. Charge customer CPU hours to run the bloated app.
Step 5. Profit.

Yeah I identified the missing steps. Finally.

Accidental revelation of code bloat?

[Misagon]

The company demonstrated the technology on servers running Microsoft Teams services, where the improved cooling enables overclocking during demand spikes that occur when meetings start on the hour and half-hour.

To me, that problem sounds like is something that would better have been resolved with more (actual) innovation in software rather than using "AI" for doing it in hardware.

But they've got to showcase "AI" for something I suppose, for marketing reasons.
Something that everyone knows about, but which st

Re:

[Xenx]

Peak load is going to be a concern, regardless of code optimization. The peak load problem is one based on human factors, not software. All optimization would change is the scale. That is, you'd still have the same concern just with potentially fewer servers overall. It might affect the cost/benefit of the work, but that is it.

teams?

[drinkypoo]

"The company demonstrated the technology on servers running Microsoft Teams services"

Teams services uses measurable CPU? It should be I/O bound. What an amazing anti-commercial for their software.

Re:

[korgitser]

I wonder if they tested it on all three of Teams' servers, or only two of them...

Re:

[Xenx]

With a reported 320+ million daily users, I would imagine there would have to be at least some CPU usage on those servers.

Re:

[tzanger]

I'd love to know how many of the 320M users are forced to use it because the bean counters saw "Teams messaging" as a free line item with their Office360 subscriptions and decided to make everyone use it instead of practically any other team messaging/interactive platform on the planet.

Re:

[allo]

Video transcoding?

Direct die with more surface area

[SafeMode]

Pretty sure this sounds like direct die liquid cooling with the added post processing of etching some channels that increase surface area and direct flow.

Seems pretty logical for an improvement to direct die cooling (which has been around forever). Just takes the work done to water blocks and implements them in micro scale on the silicon itself.

My concern would be with such tiny channels, how frequently they would get clogged by inevitable partially dissolved minerals picked up over time from the radiator

Obligatory AI reference?

[ByTor-2112]

"Optimized by AI"? I call bullshit. Optimized by FEA, whatever the marketing says.