数据中心机架功率密度快速攀升及相关影响分析

Rack power density – how much sustained power a rack draws – has been steadily climbing.

Not long ago, many operators treated 10 kW per rack as “typical,” with higher densities seen mainly in specialized environments. Today, AI workloads, space constraints, and improved thermal options are pushing practical rack capacities toward the extreme end, with designs approaching, and in some cases targeting, 1 MW per rack.

The question for operators is: How far can this trend go, and what has to change to support higher-capacity racks?

The State of Rack Power Density

Rack capacities still vary widely. Loads around 10 kW (or lower) remain common in smaller-scale data centers. That said, there has been a clear trend toward much higher rack loads in recent years.

By 2024, it had risen to 12 kW.
As of 2026, it stands at 26 kW.

From 12 kW in 2024 to 26 kW in 2026, the average has more than doubled in two years.

What’s Driving Higher Loads

Several forces are converging:

The AI boom. AI accelerators and modern servers draw substantially more power than prior generations, with individual GPUs or other accelerators reportedly drawing up to 700 W per device in some configurations.
Steady demand for capacity. With data center real estate in short supply, operators are packing more compute into existing footprints, either by increasing rack density, adopting taller or deeper cabinets, or both. Either way, the result is higher per-rack power draw.
Better thermal solutions. Advanced thermal solutions – direct-to-chip liquid cooling, rear-door heat exchangers, and immersion – make higher rack densities feasible without runaway heat.
Facility design improvements. New approaches to room layout and equipment movement (e.g., ceiling hoists for constrained rooms) ease handling and service in confined areas.

Facility Impacts: Power, Cooling, Structure

Accommodating higher per-rack power draw is not as simple as slotting in more servers. It ripples across power, cooling, and building infrastructure.

Power availability and delivery. Even when total site capacity is adequate, traditional distribution architectures may struggle to deliver hundreds of amps to a single rack with appropriate redundancy.
Cooling at scale. More watts per rack mean more heat per rack. While liquid cooling can reduce thermal resistance and improve energy efficiency, it introduces new CAPEX, monitoring and maintenance needs, water management and leak detection requirements, and new skills. Not every facility can adopt these technologies quickly.
Structural and spatial considerations. Heavier, denser racks increase floor loads and handling complexity. Facilities designed around lighter racks may require structural reinforcement, alternative floor systems, or revised pathways for installation and service.

Will Rack Loads Keep Rising?

Whether or not rack loads keep increasing will depend on demand, supply, and policy:

Demand vs. capacity. It’s plausible that total data center capacity growth could catch up to demand, easing the pressure to densify. For now, demand for AI-capable compute remains strong, and there is little evidence of a broad slowdown.
Construction constraints. Where moratoriums, interconnect delays, or permitting constraints stall new builds, operators are likely to densify existing space.

What Operators Should Do Now

Racks drawing around 1 MW will probably remain outliers for the foreseeable future, but the days when 10 kW was “typical” are ending. To stay ahead, data centers should:

Plan power architectures that can deliver higher per-rack power draw with appropriate redundancy and fault tolerance.
Validate floor loading, pathways, and handling equipment for heavier, denser gear.

In short, higher rack loads are here and rising in AI-centric environments. Facilities that proactively upgrade power, cooling, and structural systems will be best positioned to capture the next wave of compute demand.

About the Author

Technology Analyst

Christopher Tozzi is a technology analyst with subject matter expertise in cloud computing, application development, open source software, virtualization, containers and more. He also lectures at a major university in the Albany, New York, area. His book, “For Fun and Profit: A History of the Free and Open Source Software Revolution,” was published by MIT Press.