The microfluidic cooling system tested by Microsoft promises up to 3x better heat removal than cold plates, improved power usage, and reduced operational costs.
Microsoft has announced a new cooling technology for AI chips. The new cooling system uses microfluidics to channel liquid directly inside silicon chips, a breakthrough, it says, that could dramatically reshape how data centers manage the rising heat from artificial intelligence workloads.
Microsoft validated the design by cooling a server running simulated Teams meetings.
“Tiny channels are etched directly on the back of the silicon chip, creating grooves that allow cooling liquid to flow directly onto the chip and more efficiently remove heat. The team also used AI to identify the unique heat signatures on a chip and direct the coolant with more precision,” Microsoft said in a blog post.
As per the lab-scale tests conducted by Microsoft, depending on workloads and configurations involved, microfluidics performed up to three times better than cold plates at removing heat. Microfluidics was also successful in reducing the maximum temperature rise of the silicon inside a GPU by 65%, which can also vary by the type of chip.
The team expects the advanced cooling technology would also improve power usage effectiveness, a key metric for measuring how energy efficient a data center is, and reduce operational costs.
For prototyping, Microsoft has teamed up with Swiss startup Corintis to use AI to help optimize a bio-inspired design to cool chips’ hot spots more efficiently than straight up-and-down channels.
Rising thermal pressure on AI hardware
AI workloads and high-performance computing have placed unprecedented strain on data center infrastructure. Thermal dissipation has emerged as one of the toughest bottlenecks, with traditional methods such as airflow and cold plates increasingly unable to keep pace with new generations of silicon.
“Modern accelerators are throwing out thermal loads that air systems simply cannot contain, and even advanced water loops are straining. The immediate issues are not only the soaring TDP of GPUs, but also grid delays, water scarcity, and the inability of legacy air-cooled halls to absorb racks running at 80 or 100 kilowatts,” said Sanchit Vir Gogia, CEO and chief analyst at Greyhound Research. “Cold plates and immersion tanks have extended the runway, but only marginally. They still suffer from the resistance of thermal interfaces that smother heat at the die. The friction lies in the last metre of the thermal path, between junction and package, and that is where performance is being squandered.”
Cooling costs: the next data center budget crisis
Cooling isn’t just a technical challenge but also an economic one. Data centers spend heavily to manage the immense heat generated by servers, networking gear, and GPUs. Hence, the cost of cooling a data center is also a significant expense.
“As per 2025 AI infra buildouts TCO analysis, over 45%-47% of data center power budget typically goes into cooling, which could further expand to 65%-70% without advancement in cooling method efficiency,” said Danish Faruqui, CEO at Fab Economics. “In 2024, Nvidia Hopper H100 had 700 watts of power requirements per GPU, which scaled in 2025 to double with Blackwell B200 and Blackwell Ultra B300 to 1000 W and 1400 watts per GPU. Going forward in 2026, it will again more than double by Rubin and Rubin Ultra GPU to 1800W and 3600W.”
The thermal budget per GPU is at least doubling every year, therefore, in order to deploy the latest GPU and best compute performance, it is imperative for hyperscalers and neocloud providers to solve thermal bottlenecks.
Faruqui added microfluidics-based direct-to-silicon cooling can limit cooling expense to less than 20% within data center power budget but would require significant technology development optimization around microfluidics structure size, placement and non-laminar flow analysis in micro channels. If achieved, microfluidic cooling could be the sole enabler for Rubin Ultra GPU TDP budget of 3.6kW per GPU.
Scaling microfluidics
The challenge is universal. Hyperscalers, including AWS, Google, Meta, and Oracle, are all grappling with extreme chip heat as AI hardware power density soars.
“The escalating thermal load from new generations of AI silicon means that relying on today’s solutions, such as cold plates, could impose a ‘hard ceiling on progress’ within as little as five years. This makes thermal constraints a universal challenge for all hyperscalers deploying high-power AI chips,” said Brady Wang, associate director at Counterpoint Research. He added that while microfluidics is not a new idea, and multiple approaches exist, making it work at scale has proven difficult for the entire industry.
Those difficulties include manufacturing, implementation, and operational risks. “Fabricating micron-scale channels increases process complexity and may raise yield loss due to wafer fragility. Ultra reliable sealing is critical, as even minor leaks or particulate contamination could degrade chip performance,” said Manish Rawat, analyst at TechInsights. “Unlike replaceable cold plates, silicon integrated cooling makes chip replacement the only maintenance option, escalating service costs and logistical complexity. Additionally, long-term exposure to coolant, even dielectric, can induce chemical and mechanical stress, necessitating extensive qualification to ensure 5–10 year reliability.”
For microfluidics to take root, the approach demands careful management of fabrication, reliability, and maintenance risks. It must also become standard practice across the ecosystem.
