Every generation of AI accelerators pushes more compute into the same footprint. That’s good news for performance, but it means the power and heat a single rack has to handle keeps climbing, and facility planning hasn’t always kept pace.
The result is a growing gap between what data centers were designed to support and what’s actually showing up on the loading dock.
The density jump: from 16kW to 27kW in a year
Average rack density climbed from roughly 16kW in 2025 to 27kW in 2026. That’s not a gradual trend line. It’s a step change driven by newer GPU platforms packing far more silicon, and far more power draw, into a single rack than the generation before it.
For facilities built even two or three years ago, that shift alone can outpace what the original electrical and cooling design assumed.
Why 50–70kW racks catch operators off guard
Racks in the 50–70kW range are now common in AI deployments, but only about one in five operators report being prepared to support them. Some of the newest AI systems push even further, with power draw reaching up to 246kW per rack.
“Unprepared” here isn’t abstract. It shows up as power feeds sized for a lower draw, cooling loops without the capacity to handle the added heat, floor loading limits that weren’t built for denser hardware, and electrical distribution equipment specified for assumptions that no longer hold.
Air cooling hits a wall
Traditional air cooling can’t dissipate the heat load of a fully populated high-density AI rack. Platforms like NVIDIA’s GB200 NVL72 require direct-to-chip liquid cooling; facilities running these systems don’t have air cooling as a fallback.
That’s pushed liquid cooling from a niche option to a baseline requirement, whether through direct-to-chip cooling, rear-door heat exchangers, or immersion systems, each with different implications for how a rack row, a cooling loop, and a facility’s water strategy need to be designed.
It’s not just a cooling problem
Higher density also means bigger builds. The average facility size operators are now planning for has grown to roughly 38MW, up from 32MW the year before. Operators are increasingly designing for scale from day one instead of planning incremental expansion.
That shifts density planning into conversations that used to happen much later: site selection, utility interconnection capacity, and how much room a facility has to grow into future hardware generations without a full redesign.
Where planning needs to start earlier
Grid and power constraints are pushing the planning conversation earlier in the project timeline. Facilities need to plan liquid-cooling infrastructure and electrical distribution around grid and power availability before density requirements are finalized, not after.
That timing matters even more given how long procurement for major electrical equipment already takes. When switchgear and transformer lead times are measured in years, waiting until density numbers are locked in to start planning the infrastructure around them isn’t a viable sequence anymore.
What this means for installation and commissioning
For teams executing these projects, the practical implications touch nearly every phase of the work. Rack and stack sequencing has to account for weight and airflow assumptions that may not match the original building design. Structured cabling layouts shift when higher-density racks change port counts and cable paths. Power distribution has to be sized for the actual load, not the load the facility was originally built around.
Cooling system commissioning becomes its own validation step, confirming that a liquid cooling loop performs under real thermal load rather than a theoretical one. And documentation needs to reflect what the facility actually runs, so the next team working on that site isn’t inheriting outdated assumptions.
If your team is mapping out density requirements for an upcoming deployment, we’re glad to talk through what we’re seeing in the field.
Sources: Nautilus Data Technologies, “GPU Density Benchmarks 2026” · build.inc, “GPU Data Center Development: Site Criteria, Power Density, and Cooling Requirements” · Spheron Network, “Power-Bound, Not GPU-Bound” · AFCOM State of the Data Center 2026 (via Datacenter Knowledge) · Schneider Electric, “Planning liquid-cooled AI data centers around grid and power constraints” (Jul 28, 2026)