Microsoft Invests in Advanced Cooling to Address AI’s Growing Thermal Demands
As artificial intelligence workloads surge, data centers are facing a critical challenge: managing the escalating heat generated by powerful new hardware. Traditional cooling methods are struggling to keep pace, prompting Microsoft to explore innovative solutions to prevent cooling costs from crippling data center budgets and hindering the deployment of next-generation GPUs.
The increasing thermal density is driven by the rapid rise in GPU power consumption.According to Danish Faruqui, CEO at Fab Economics, “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.” This trend is accelerating: Nvidia’s Hopper H100 GPU required 700 watts in 2024, a figure that doubled with the Blackwell B200 and is projected to reach 1000W and 1400W with the Blackwell Ultra B300.
“Modern accelerators are throwing out thermal loads that air systems simply cannot contain, and even advanced water loops are straining,” explains Sanchit Vir Gogia, CEO and chief analyst at Greyhound Research. “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.” Gogia points to the critical bottleneck in the “last metre of the thermal path, between junction and package,” where performance is lost due to thermal interface resistance.
Looking ahead to 2026, Nvidia’s Rubin and Rubin Ultra GPUs are expected to demand 1800W and a staggering 3600W respectively, further intensifying the thermal pressure. To unlock the full potential of thes GPUs,hyperscalers and neocloud providers must overcome these thermal limitations.
Microsoft is focusing on microfluidic-based direct-to-silicon cooling as a potential solution. Faruqui believes this technology could reduce cooling expenses to less than 20% of the overall data center power budget. However, realizing this potential requires meaningful research and progress focused on optimizing the microfluidic structure’s size, placement, and analysis of non-laminar flow within the microchannels. Successfully developing this technology is seen as crucial; Faruqui states that microfluidic cooling “could be the sole enabler for Rubin Ultra GPU TDP budget of 3.6kW per GPU.”
