Rack Power Density Explained: Why AI Racks Hit 120kW

Rack power density explained: watts-per-U and kW-per-rack

Rack power density is simply how much electrical power a rack turns into heat in a fixed physical space. Two figures capture it. Watts-per-U measures the intensity of a single slot: a standard rack has 42 rack-units (42U) of 1.75-inch mounting height, and a legacy 1U server drawing 500 W is a 500 W/U slot. Kilowatts-per-rack is the whole-cabinet total: multiply the average slot load by the U's you fill and you get the number your power feed, PDU and cooling must actually support.

The two are not interchangeable. You can hit a high kW-per-rack total with modest per-U loads spread over 42U, or you can concentrate an extreme per-U load into a handful of slots. AI kit does the second, and that is what breaks conventional halls. A rack is not limited by empty space - it is limited by how many watts you can feed in and pull back out as heat before the slot, the busbar or the air handling gives up.

Getting these units right is the whole game when you size a deployment. Our AI GPU calculator works in exactly these terms, translating a target model or GPU count into the kW-per-rack and cooling class you will need to provision.

The old normal: UK halls built for 5-15 kW

For two decades the British colocation and enterprise standard barely moved. A basic cabinet was provisioned at 3-5 kW, a well-loaded enterprise rack at 5-10 kW, and a modern high-containment rack pushed 10-15 kW while still being cooled by air. The Uptime Institute's 2025 Global Data Center Survey puts the worldwide mean rack density at just 7.6 kW, up only fractionally from 6.8 kW the year before - and it notes that few facilities exceed 30 kW at all.

That number matters because it describes the installed base most UK operators are still running. Raised floors, computer-room air handlers and three-phase power distribution were all engineered around single-digit kilowatts per rack. A hall commissioned in 2015 for a 5 kW average is a perfectly good hall - for the workloads it was designed for. The problem is not that these facilities are broken. It is that a new class of hardware has arrived that ignores their assumptions entirely.

This is also why the refurbished and prior-generation market stays healthy: plenty of mainstream compute still fits comfortably inside 5-15 kW envelopes. If your workload is virtualisation or general enterprise IT rather than large-model training, kit from the refurbished server tier will slot into an existing hall without any power or cooling drama.

The culprit: a GPU that eats 1,200-2,300 watts

The density jump starts at the chip. A mainstream CPU server draws a few hundred watts. A single NVIDIA Blackwell B200 GPU is rated at around 1,000 W air-cooled and up to 1,200 W in its liquid-cooled form - roughly a 43% jump over the 700 W Hopper H100 it replaced. The next generation goes further: NVIDIA's Vera Rubin GPU (the VR200 part) is reported to have been pushed to a 2,300 W thermal design power in its top 'Max-P' configuration, nearly double a liquid-cooled Blackwell.

Put that in perspective. One Rubin GPU at 2,300 W draws more power than a whole 1U enterprise server did a few years ago - and an AI compute tray holds several of them. When each accelerator is a kilowatt-plus space heater, and you pack dozens into a rack to feed a single training job over a high-bandwidth fabric, the rack total climbs into territory no air-cooled hall was ever meant to handle.

For the full component picture - GPU specs, power draw, vendors and system-level numbers - our AI server data study keeps a cited reference dataset you can dig into.

How 72 GPUs become 120 kW: inside the NVL72

The flagship example is NVIDIA's GB200 NVL72. It packs 72 Blackwell GPUs and 36 Grace CPUs into a single liquid-cooled rack across 18 compute trays, wired together as one giant accelerator by an NVLink fabric. NVIDIA specifies the rack at 132 kW nominal thermal design power, with observed full-load draw of 130-132 kW and a peak electrical design point around 192 kW (about 1.5x the nominal). The commonly quoted 120 kW is the working figure operators plan around.

That 120-132 kW is roughly 8-15 times a well-filled UK enterprise rack of 8-15 kW, and about 25 times a legacy 5 kW cabinet. It is around 3x a fully loaded air-cooled HGX H100 rack, which NVIDIA's own reference design caps near 41 kW. In one 600mm footprint you have concentrated the power of a small server room.

And this is not the ceiling. The Rubin-based VR200 NVL72 shipping through 2026 is projected to draw roughly 190-230 kW per rack, and the 2027 'Kyber' Rubin Ultra rack is specified at around 600 kW, with 1 MW-class racks openly discussed behind it. Rack power density is not settling at a new normal - it is still climbing steeply.

The cooling cliff: why air runs out near 20 kW

Density is ultimately a thermal problem: every watt in becomes a watt of heat to remove. Traditional air cooling has a hard practical ceiling. ASHRAE's TC 9.9 guidance classifies anything above 20 kW per rack as high-density and points operators towards water-based cooling; in the field, 15-20 kW is where well-contained air realistically tops out. Enhanced techniques - hot/cold-aisle containment and rear-door heat exchangers - stretch that to roughly 40-50 kW, but at rising cost and complexity.

Above about 50 kW there is no air option. Direct-to-chip (DLC) liquid cooling, where coolant is piped onto cold plates sitting directly on the GPUs, spans roughly 50-150 kW per rack and, in purpose-built designs, up to ~200 kW. The GB200 NVL72 at 132 kW is a liquid-cooled machine by necessity, not choice - about 115 kW of its heat is carried away by liquid, with only the remainder handled by air.

For a UK operator this is the decisive fork. A hall built for 5-15 kW air-cooled racks cannot simply have a 120 kW rack wheeled in. You are looking at facility water loops, coolant distribution units and structural changes - effectively a new build or a deep retrofit. The bill for that is exactly the kind of capital decision our IT finance calculator is designed to model before you commit.

Grid to chip: the power-delivery chain and the 800 VDC shift

Getting 120 kW into a rack is as hard as getting the heat out. Power arrives from the grid as high-voltage AC - typically an 11 kV or 33 kV feed - and is stepped down by a site transformer to around 400 V three-phase, passed through UPS and switchgear, then carried to the rack over busway or PDUs at roughly 415 V AC. Inside today's AI racks, power shelves rectify that AC to 54 V DC at about 97% efficiency, and a heavy copper busbar shuttles it up the rack to each compute tray, where point-of-load VRMs drop it to the sub-1 V core voltage the GPU silicon actually runs on.

Every conversion loses a little, and at 132 kW even a 3% loss in the power shelves alone is around 3.6 kW of extra waste heat. That is why the industry is moving the goalposts again. NVIDIA has announced an 800 VDC facility-to-rack architecture for the 600 kW-1 MW racks coming after Blackwell; running direct current at 800 V pushes over 150% more power through the same copper than 415/480 V AC, cutting conductor mass and conversion stages. A single 1 MW rack on the old 54 V scheme would otherwise need up to 200 kg of copper busbar.

The practical takeaway for buyers: the constraint has shifted from the chip to the electrical plant around it. Whole-rack thinking - transformer capacity, busway rating, redundancy and conversion losses - now decides whether a deployment is even feasible in a given building.

The UK squeeze: grid queues, not just hot racks

In Britain the density problem collides with a grid problem. More than 80% of UK data-centre capacity sits in and around London, and the West London cluster - including Slough's 35-plus facilities and the Heathrow corridor - is approaching saturation on both land and power. Some new connections in the area have been quoted as unavailable until the mid-2030s because local substations have hit their ceiling. Data centres already consume around 2.5% of UK electricity, a figure that could climb towards 10-15% as AI build-out continues.

That means a UK buyer cannot treat a 120 kW rack as purely a kit decision. The binding constraint is often the incoming supply: securing a high-capacity grid connection routinely takes over a year even when approved, which is why operators are pushing north and into the government's designated AI Growth Zones, each earmarked for up to 500 MW. Where you can physically get the megawatts is now a first-order design input.

It also reframes what to do with existing estates. Many organisations are consolidating older, lower-density workloads to free power and floor space for a small number of high-density AI racks - a calculation that runs alongside decisions about virtualisation platform costs and when to retire kit that is past its useful life.

What UK buyers should do now

The honest starting point is to know your real numbers. Measure per-U and per-rack draw for what you run today, then map planned AI workloads onto the density bands: sub-20 kW stays on air, 20-50 kW needs enhanced air or rear-door exchangers, and anything above that means committing to liquid. Do not let a single 120 kW rack ambush a hall designed for 7.6 kW averages.

Second, separate the workloads. General enterprise compute and virtualisation do not need AI-class density and should stay in efficient, air-cooled, cost-effective racks - often on prior-generation or right-sized configured hardware. Reserve the expensive liquid-cooled envelope for the accelerators that genuinely require it, and check lifecycle timing against end-of-life support dates so you are not retrofitting a hall for gear that is about to age out.

Third, plan the plant, not just the servers. Grid connection, transformer headroom, cooling medium and busway rating decide feasibility long before the rack ships. Model the capital and operating cost across scenarios, and get independent advice on whether a retrofit, a colocation move or a new high-density build is the right answer for your specific supply and timeline.

Sources

Every figure in this guide is drawn from the vendor documentation, industry surveys and technical reporting below, current as of July 2026.

• •Uptime Institute — Global Data Center Survey 2025 (mean rack density 7.6 kW): https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2025
• •Tom's Hardware — Nvidia boosts Vera Rubin power to 2,300 W: https://www.tomshardware.com
• •NVIDIA — DGX GB200 Rack Scale Systems User Guide (132 kW, power shelves, 54 V DC): https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html
• •SemiAnalysis — GB200 Hardware Architecture (AC to 54 V DC, 97% efficiency, compute trays): https://newsletter.semianalysis.com/p/gb200-hardware-architecture-and-component
• •NVIDIA Technical Blog — 800 VDC architecture for AI factories: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/
• •Data Center Dynamics — NVIDIA prepares industry for 1 MW racks and 800 V DC: https://www.datacenterdynamics.com
• •Introl — Liquid vs air cooling for AI data centres (air ~20 kW, DLC 50-150 kW): https://introl.com/blog/liquid-vs-air-cooling-ai-data-centers
• •IntuitionLabs — NVIDIA HGX data centre requirements (HGX H100 rack ~41 kW): https://intuitionlabs.ai/articles/nvidia-hgx-data-center-requirements
• •techUK — Data Centre Programme: what we achieved in 2025 (UK grid, London saturation): https://www.techuk.org/resource/data-centre-programme-what-we-achieved-in-2025.html
• •Introl — NVIDIA B200 vs GB200 deployment guide (B200 1,000-1,200 W): https://introl.com/blog/nvidia-b200-vs-gb200-deployment-guide