AI hardware has changed the physics of the data center floor. A single rack of current AI accelerators now draws more power than a full row of traditional servers did a few years ago, and that shift breaks the assumptions most buildings were designed around. AI data center cooling has moved from a background utility to the constraint that decides which facilities can host modern hardware at all.
NVIDIA rates its GB200 NVL72 rack at roughly 120 kW, and deployed systems draw 130 kW or more. Air cooling reaches its practical limit near 30 kW per rack. Most existing facilities handle only 5 kW to 15 kW, so they cannot host this hardware without a deep retrofit.
Why AI racks broke the old power budget
AI accelerators pack more compute, and therefore more heat, into the same rack footprint than any earlier generation. NVIDIA rates the GB200 NVL72 at about 120 kW per rack, and operators report deployed racks drawing 130 kW to 132 kW under load. For comparison, the Uptime Institute finds that most production racks still sit in the 10 kW to 30 kW range. Axe Compute builds capacity on infrastructure already engineered for these densities, so enterprise teams do not inherit the retrofit problem.
The gap between those two numbers is the whole story. A facility designed around 10 kW racks came with air handlers, floor loading, and power feeds sized for that world. Drop a 120 kW rack into it and the cooling system cannot keep the silicon within its thermal limits. Consequently the hardware throttles, or it simply will not fit. This is a physical ceiling, and marketing language does not raise it.
Air cooling stops working long before 120 kW
Air has a hard limit as a heat transfer medium. Traditional air cooling becomes ineffective somewhere between 20 kW and 35 kW per rack, depending on airflow design and inlet temperature. Beyond that point, operators reach for water. Rear-door heat exchangers push viable density to roughly 40 kW to 60 kW. Direct-to-chip liquid cooling, where coolant runs through cold plates pressed onto the processors, supports 80 kW to 120 kW. Immersion cooling, which submerges hardware in dielectric fluid, reaches higher again.
The industry has already recognised this shift in its standards. ASHRAE Technical Committee 9.9, the body that sets thermal guidelines for data center equipment, updated its framework to address rack densities above 40 kW, an admission that the previous air-first model no longer matches where hardware is going. In practice, any facility that wants to host current AI accelerators has to plan for liquid at the rack.
Most existing facilities were engineered for a different workload
The efficiency numbers show how little the installed base has moved. The Uptime Institute reported a global average power usage effectiveness of 1.54 in its 2025 survey, a figure that has held flat for six consecutive years. By contrast, facilities purpose-built with direct-to-chip cooling reach a power usage effectiveness of 1.03 to 1.20, close to the theoretical floor. Adoption tells the same story: in the 2025 survey, only 22 percent of operators used direct liquid cooling, while 75 percent still relied on perimeter air.
Retrofitting the rest is not a quick upgrade. It means installing coolant distribution units, running liquid loops to every rack, reinforcing floors for heavier hardware, and rebuilding power distribution. Many older buildings also sit behind grid connections that cannot deliver the additional power a dense AI hall requires. The result is a large stock of data center space that cannot host the hardware enterprises actually want to run.
Power delivery is being rebuilt from the rack up
Cooling is only half of the physical problem. Engineers are redesigning power delivery inside the rack as well. The legacy 54 V distribution standard targeted kilowatt-scale racks, and it has become a bottleneck as racks approach the megawatt scale. NVIDIA and more than twenty infrastructure partners are moving to an 800 VDC architecture for this reason. According to NVIDIA, 800 VDC transmits over 150 percent more power through the same copper, which removes the need for heavy busbars feeding each rack and improves efficiency across the power chain.
This transition arrives with the next hardware generation. The 800 VDC power portfolio ships in the second half of 2026, aligned to the rollout of NVIDIA Vera Rubin platforms. For enterprise teams, the takeaway is direct: the buildings that will host the next wave of AI hardware are being re-engineered at the level of copper and coolant, not merely furnished with new servers.
What facility limits mean for enterprise AI strategy
Facility readiness has become a procurement question, not an operations footnote. An enterprise team can wait for a builder to retrofit or construct a facility, which takes quarters and depends on grid access, or it can place workloads on infrastructure that already meets the density requirement. That is a choice about time and control, and it belongs in the planning stage of any serious AI program. Teams weighing this decision should read our view on enterprise GPU strategy in 2026 and the case for bare-metal over virtualised GPU access.
Run modern AI hardware on infrastructure already built for it.
Frequently Asked Questions
Why does modern AI hardware require liquid cooling?
AI accelerators concentrate far more power, and therefore heat, into a single rack than earlier hardware. Air cooling reaches its practical limit near 30 kW per rack, while a modern AI rack can draw 120 kW or more. Liquid carries heat away far more effectively than air, so it becomes the only option at these densities.
How much power does an NVIDIA GB200 NVL72 rack use?
NVIDIA rates the GB200 NVL72 at roughly 120 kW per rack. Operators report deployed systems drawing 130 kW to 132 kW under full load, which is why direct liquid cooling is mandatory rather than optional.
What is the power limit of air cooling in a data center?
Traditional air cooling becomes ineffective between about 20 kW and 35 kW per rack. Rear-door heat exchangers extend that to roughly 40 kW to 60 kW, and direct-to-chip liquid cooling supports 80 kW to 120 kW, while immersion cooling reaches higher still.
Why can most existing data centers not host modern AI hardware?
Most facilities top out at racks of 5 kW to 15 kW, using air cooling and kilowatt-scale power distribution. Hosting AI hardware requires liquid cooling loops, coolant distribution units, higher floor loading, and rebuilt power delivery, which amounts to a deep retrofit rather than a simple upgrade.
What is the 800 VDC power architecture for AI data centers?
800 VDC is a power delivery standard promoted by NVIDIA and its partners to replace the legacy 54 V in-rack approach. It transmits over 150 percent more power through the same copper, which removes the need for heavy busbars and supports megawatt-scale racks in the Vera Rubin generation.
How does Axe Compute avoid the facility retrofit problem?
Axe Compute provides bare-metal GPU infrastructure across 200+ locations in 93 countries on 400,000+ GPUs that are already live. Enterprise teams reserve capacity on infrastructure engineered for high density, so they do not carry the cost or delay of retrofitting a building.
About Axe Compute
Axe Compute (NASDAQ: AGPU) provides bare-metal GPU infrastructure across 200+ locations in 93 countries. The platform operates 400,000+ GPUs with 48-hour provisioning, zero egress fees, no virtualisation overhead, and 99% uptime. Pricing runs significantly below hyperscaler rates. Contact us at info@axecompute.com.
Sources
– NVIDIA GB200 NVL72 rack specification: https://www.nvidia.com/en-us/data-center/gb200-nvl72/
– NVIDIA GB200 NVL72 by HPE (per-rack power breakdown): https://buy.hpe.com/us/en/compute/rack-scale-system/nvidia-nvl-system/nvidia-gb200-nvl72-by-hpe/p/1014890104
– Air, rear-door, direct-to-chip, and immersion cooling density thresholds (Introl): https://introl.com/blog/liquid-vs-air-cooling-ai-data-centers
– ASHRAE TC 9.9 update for rack densities above 40 kW (Tom’s Hardware): https://www.tomshardware.com/pc-components/cooling/the-data-center-cooling-state-of-play-2025-liquid-cooling-is-on-the-rise-thermal-density-demands-skyrocket-in-ai-data-centers-and-tsmc-leads-with-direct-to-silicon-solutions
– Uptime Institute 2025 Global Data Center Survey (PUE 1.54, liquid cooling adoption): https://mgrid.org/2025/10/01/uptime-institute-data-center-pue-stagnation-2025-liquid-cooling/
– NVIDIA 800 VDC architecture for AI factories: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/
– Flex 800 VDC power rack for NVIDIA Vera Rubin (H2 2026 timing): https://investors.flex.com/news/news-details/2026/Flex-Launches-800-VDC-Power-Rack-for-Next-Generation-NVIDIA-AI-Infrastructure/default.aspx