GPU Rack Power Density 2026: 700W Chips to 600kW Racks

The AI Power Surge: Why 2026 Is the Inflection Point

Rack power density has stopped being a footnote in data-centre planning and become the constraint that decides whether a facility can host 2026-class AI hardware at all. The average data-centre rack reached 27 kW in 2026, a 69% year-on-year jump from 16 kW in 2025, according to AFCOM's infrastructure research. That average masks a far steeper curve for AI-specific estates: dedicated AI/HPC racks now routinely run at 80-120 kW, several times a standard enterprise rack's 10-15 kW envelope.

The driver is unambiguous. AI-optimised servers are forecast to account for 31% of total data-centre power consumption in 2026, up from roughly 20% in 2025, and Gartner puts global data-centre electricity consumption at 565 TWh for the year. Individual AI data-centre sites now demand between 100 MW and 750 MW each. For UK buyers weighing colocation against on-premise builds, the practical question is no longer how much power the IT load needs, but whether the building, the grid connection and the cooling plant can physically deliver it.

The Wattage Ladder: GPU-by-GPU Power Draw in 2026

The per-accelerator wattage driving this curve is now well documented across vendors, and it forms a clear ladder. Intel's Gaudi 3 sits at the lower end: 600W for the PCIe card, 900W for the OAM module. AMD's Instinct MI300X OAM module draws 750W. NVIDIA's Blackwell B200 runs at 1000W TDP in its standard air-cooled SXM/HGX configuration, rising to a 1200W peak when liquid-cooled, and the Blackwell Ultra B300 pushes further still to 1400W.

This is the figure every UK buyer should anchor their planning to, because it is the input for every downstream calculation — PDU sizing, busway capacity, cooling plant tonnage and floor loading all cascade from the per-chip TDP multiplied across a rack's accelerator count.

• •Intel Gaudi 3 PCIe card: 600W TDP
• •AMD Instinct MI300X (OAM module): 750W TDP
• •Intel Gaudi 3 OAM module: 900W TDP
• •NVIDIA B200 SXM, air-cooled HGX/DGX: 1000W TDP
• •NVIDIA B200 SXM, liquid-cooled peak: 1200W TDP
• •NVIDIA B300 Blackwell Ultra: 1400W TDP

From Rack to Room: A Worked Power-Density Calculation

Take a typical 8-GPU tray as the base unit. Eight AMD MI300X modules at 750W each sum to 6,000W (6 kW) of GPU-silicon load alone. Eight NVIDIA B200s in air-cooled HGX form sum to 8,000W (8 kW). The same eight B200s at their liquid-cooled peak of 1200W each sum to 9,600W (9.6 kW). Eight B300 Blackwell Ultra chips reach 11,200W (11.2 kW) on GPU silicon alone.

These are compute-only subtotals — before networking switches, CPUs, memory, storage and power distribution overhead are added, and before multiple trays are stacked into a single rack. That is exactly why real deployed AI/HPC racks measure 80-120 kW rather than the 6-11 kW of GPU silicon in one tray, and why NVIDIA's own GB200 NVL72 rack — a multi-tray, rack-scale system — is rated at approximately 120 kW. Buyers sizing a deployment should use an AI GPU requirements calculator and understand rack power density before assuming a colocation hall's existing power allocation per rack will suffice.

Cooling the Beast: Matching Liquid Cooling to Density Tier

Traditional air cooling struggles to handle racks above 30-40 kW, which puts almost every AI/HPC deployment discussed above beyond air's practical ceiling. That is why liquid cooling has become the default architecture for AI-centric deployments in 2026, not an optional upgrade — and why UK facilities still relying on hot/cold-aisle air containment for anything above roughly 30-40 kW per rack are already operating at the edge of physics.

The efficiency case reinforces the thermal one: liquid cooling can improve energy efficiency by up to 15% compared with air cooling, a gain that compounds across a facility's annual electricity bill once density passes the point where air simply cannot move enough heat. For a fuller comparison of direct-to-chip, rear-door and immersion approaches by density tier, see our dedicated guide to AI server cooling solutions.

The Legacy Wall: UK Infrastructure and Retrofit Realities

For UK IT buyers, the escalating densities mean that infrastructure built for traditional enterprise workloads can become obsolete far faster than its depreciation schedule assumes. Facilities designed around 10-15 kW racks and air cooling were never engineered for 80-120 kW AI trays, and retrofitting power distribution, containment and fire suppression for that jump is rarely a bolt-on project — it typically means redesigning electrical risers, busways and cooling plant from the switchgear up.

UK colocation providers are responding by investing heavily in high-density racks and liquid-cooling infrastructure, and that capital outlay will feed through into higher per-kW service pricing over time. On-premise buyers face the same maths without a landlord to share it: electrical and cooling system upgrades, and in many cases a full facility redesign, are now a precondition for hosting current-generation GPU hardware rather than a nice-to-have. Compounding this, grid interconnection delays exceeding three years are cited as a primary business risk for AI infrastructure expansion, and more than 50% of US data-centre builds planned for 2026 face delays tied to four-year lead times for power transformers — a global supply-chain constraint that UK projects sourcing the same equipment are not immune to.

TCO and the UK Angle: Grid, Cost and Compliance

The financial picture scales with the physical one. Data-centre CapEx per MW can increase by 2-4 times for builds supporting 600 kW and above, reflecting the heavier electrical infrastructure, liquid-cooling plant and structural reinforcement that extreme density demands. Combined with multi-year transformer lead times and interconnection queues, this makes power availability — not chip supply — the binding constraint on how fast a UK operator can stand up new AI capacity.

Regulatory and sustainability pressure adds a further layer: as facilities push density higher, PUE and energy-efficiency compliance become harder to hold without the roughly 15% efficiency gain liquid cooling offers over air. Specific GBP cost benchmarks for UK electricity and build costs were not available in the data reviewed for this tracker, but the underlying global cost and lead-time pressures — CapEx multipliers, transformer scarcity, grid delay risk — apply directly to any UK site planning a high-density AI deployment, whether hosting high-density AI racks in colocation or building on-premise.

2027-2030 Outlook and a Checklist for UK IT Leaders

The curve does not flatten from here. NVIDIA is developing the Rubin Ultra Kyber rack, a test unit projected to reach 600 kW, with a release slated for around summer 2027 — five times today's GB200 NVL72. Data-centre consultants are already designing racks for 2.2 megawatts within a five-year timeframe from 2026, signalling that today's 'extreme' density will look routine by the turn of the decade.

For UK IT leaders evaluating colocation against on-premise strategy, the decision now hinges on a small set of hard questions: does the facility's grid connection have realistic headroom given multi-year interconnection queues; is the cooling plant liquid-ready above the 30-40 kW air ceiling; can the provider absorb a 2-4x CapEx multiplier without passing an unsustainable premium to tenants; and does the roadmap account for GPU generations that may double per-rack draw again before 2030. Model your own configuration with an AI GPU requirements calculator and compare it against current NVIDIA DGX systems before committing to a facility.

Methodology

This tracker compiles per-GPU thermal design power figures and rack-level power density benchmarks published between January and July 2026 by vendor documentation (NVIDIA, Intel), hardware databases (TechPowerUp), industry research bodies (AFCOM, Gartner) and specialist data-centre infrastructure publications (Data Center Trends 2026, ETDatacenters, Enki AI, Data Gravity, Introl Blog, Jiumeng Power Supply, Spheron Blog). Figures span vendor-published TDP specifications, industry survey averages and named forward-looking engineering projections, each retained with its original scope — GPU model, form factor, cooling method and calendar period — rather than merged into a single blended number.

Every figure in this piece is reproduced with the scope stated by its original source: for example, NVIDIA B200 TDP is reported separately for air-cooled (1000W) and liquid-cooled peak (1200W) configurations, and rack density figures are kept distinct between 'average data centre rack' (27 kW) and 'AI/HPC-specific rack' (80-120 kW) rather than treated as interchangeable. Where the underlying research did not provide UK-specific GBP cost data, that gap is stated explicitly rather than estimated, and no figure has been averaged, extrapolated or combined across sources beyond the simple GPU-count multiplication shown in the worked example.

Sources

Every figure in this article traces to the sources below.

• •AFCOM — average data centre rack density 2026
• •Jiumeng Power Supply — AI/HPC rack power density 2026
• •Spheron Blog — NVIDIA B200 TDP (air-cooled SXM)
• •Alican Kiraz / Tech Bytes — NVIDIA B200 liquid-cooled peak TDP
• •TechPowerUp GPU Database — AMD Instinct MI300X TDP
• •TechInsights — Intel Gaudi 3 OAM/PCIe TDP
• •Intel — Gaudi 3 PCIe card TDP
• •Introl Blog — NVIDIA GB200 NVL72 rack power
• •Data Center Trends 2026 — Rubin Ultra Kyber, air-cooling limits, 2.2MW racks
• •How Liquid Cooling is Reshaping AIDC Performance in 2026 — efficiency gains