Thoughts on the data center thermal management shift. [[Hyperscalers]] are deploying $300B+ in 2025 capex (median 53% YoY increase), but AI chips now exceed 1000W [[Thermal Design Power - TDP]]. Air cooling can't keep up. Liquid is mandatory. ![[Screenshot 2026-01-27 at 12.00.24.png]] The physics set hard limits. Air cooling tops out at 1000W TDP. [[Direct-to-Chip Cooling]] handles 1000-1600W. Beyond 1600W, only [[Two-Phase Cooling]] or [[Immersion Cooling]] works. Nvidia's latest chips push these boundaries, making the transition less a choice than a forced migration. ![[Screenshot 2026-01-27 at 12.01.13.png]] Two approaches dominate. Immersion submerges entire servers in dielectric fluid, capturing all waste heat uniformly. Requires specialized equipment and facility redesign. DTC targets individual processors with cold plates, integrating with existing [[Racks]] infrastructure for faster deployment and easier maintenance. ![[Screenshot 2026-01-27 at 12.01.51.png]] The market validated both. Over $1B in venture capital flowed into cooling startups since 2023. Submer (immersion, $131M raised), Iceotope (DTC, $78M raised). Strategic acquirers paid 15-25x revenue multiples. Schneider bought Motivair for $1.1B at 25x revenue. KKR took CoolIT at 23x. Liquid cooling penetration in AI servers jumps from 36% (2024) to 59% (2029). HPC follows similar trajectory: 40% to 62%. Once physical limits force change, adoption accelerates regardless of operational inertia. ![[Screenshot 2026-01-27 at 12.02.23.png]] DTC wins on serviceability and deployment speed. Immersion wins on total energy efficiency and heat capture. The answer depends on [[Rack power Density]] requirements and existing infrastructure constraints. Most hyperscalers deploy both, optimizing by workload and facility. > Capital flows to atoms, not bits. Data centers become thermal management companies that happen to run compute. The constraint shifted from silicon performance to heat removal. This connects directly to [[What do hyperscalers care about]]: scale and speed matter, but now you can't have either without solving thermal. The [[AI Capex Super-Cycle]] makes this urgent. $400B+ chasing compute supremacy, but that compute is worthless if you can't cool it. Every 1% improvement in [[Power Usage Effectiveness - PUE]] cascades through operating costs. The infrastructure buildout also touches [[Geothermal Energy and the Future of Data Centers]] for waste heat recovery and [[SMRs x Data Center Opportunities]] for reliable baseload power. Related: [[Cooling and Energy Efficiency Technologies]], [[Data Centre Energy Demand]], [[The infrastructure layer and AI capex]], [[The Unsustainable Data Center]] --- #deeptech #datacenters #AI #infrastructure