Microsoft announces that all new Azure datacenters from 2027 will use liquid cooling as standard to support next-gen AI chips. PUE hits record lows.
Why Azure is standardizing liquid cooling
Microsoft has said that every new Azure datacenter starting in 2027 will use liquid cooling as the default design, not an optional rack feature. The driver is next-generation AI chips: denser accelerators put more heat into a smaller footprint than air systems can remove without oversized fans, deep cold aisles, and high power draw just for moving air. Liquid carries heat away more efficiently, so the cooling plant can keep silicon within safe temperatures while the facility itself uses less energy for cooling.
Making liquid cooling standard for new builds is a facilities decision as much as a silicon one. Once every hall is designed around coolant loops, manifolds, and leak-safe rack plumbing, Azure can place high-density AI capacity without redesigning each site from scratch. That matters for customers who need predictable capacity for training and inference clusters rather than waiting for one-off “high density” pods.
What liquid cooling changes in the hall
Air cooling relies on large volumes of conditioned air, raised floors or containment, and chillers sized for worst-case heat load. Liquid cooling moves heat at the chip or cold plate into a closed loop, then rejects it at a heat exchanger or outdoor plant. Less air movement usually means lower fan energy, quieter halls, and tighter temperature control at the device—useful when accelerators run near thermal limits for long jobs.
Standardization also forces operators to treat coolant, fittings, and maintenance as first-class operations work. Teams plan for fluid quality, isolation valves, and procedures that do not apply to pure air designs. For multi-tenant cloud regions, a single cooling architecture simplifies how Microsoft builds, monitors, and staffs sites over the life of the building.
- Higher rack density for AI without proportional growth in airflow infrastructure
- Lower cooling energy relative to IT load when the plant is designed around liquid rejection
- More uniform thermal conditions across dense accelerator trays
- Operational skills and tooling that match a liquid-first footprint from day one
PUE and what “record lows” signal
Power Usage Effectiveness (PUE) compares total facility power to power used by IT equipment. When PUE falls, more of each watt goes to compute rather than cooling, lighting, and other overhead. Microsoft reports that its Azure fleet has reached record-low PUE figures as efficiency work and denser, better-cooled designs take hold. Liquid cooling is one lever in that trend: removing heat with fluid instead of massive air movement can shrink the non-IT share of the energy bill when the rest of the plant is well tuned.
PUE is a facility metric, not a guarantee of cheaper cloud bills or greener apps by itself. Workload mix, utilization, and how customers size clusters still dominate cost and carbon at the service layer. Still, a lower baseline overhead means each region can host more useful compute per unit of grid power—relevant as AI demand grows faster than traditional enterprise VM fleets.
What builders and buyers should plan for
If you run AI or high-density workloads on Azure, treat 2027 as a planning horizon for new capacity rather than a switch that flips existing regions overnight. Existing air-cooled halls will keep serving current generations of hardware; the commitment applies to new datacenters. When you request capacity, ask how the region’s cooling design maps to GPU or accelerator density, and design job placement around where liquid-ready capacity will land first.
On the application side, liquid cooling does not change APIs or model code. It changes how densely providers can pack the chips those models need. Prefer architectures that scale horizontally and tolerate placement across availability zones, so you can follow density as it comes online. Monitor thermal-related throttling and job duration the same way you already watch queue time and instance health—cooler, denser racks only help if your software uses the hardware without long idle periods that waste the efficiency gains the facility was built for.