AI rack power reference
What each AI platform actually draws per rack, what it forces on cooling and distribution, and where the published figures disagree. A typical enterprise cabinet runs 11 kW. A GB300 rack runs 137. That gap is the whole planning problem.
Last reviewed September 2026
Rack-scale platforms
HGX H100 / H200
NVIDIA · Hopper- GPUs
- 8 per node, typically 4–5 nodes
- Cooling
- Air with containment, or rear-door heat exchanger
- Distribution
- 208 V three phase
- Availability
- Widely deployed
The last generation that a conventional well-contained hall can cool without liquid.
GB200 NVL72
NVIDIA · Blackwell- GPUs
- 72 Blackwell GPUs, 36 Grace CPUs
- Cooling
- Direct-to-chip liquid, no air variant
- Distribution
- 208 V three phase, busway
- Availability
- In volume deployment
The step change. Roughly ten times a typical enterprise cabinet in the same footprint.
GB300 NVL72
NVIDIA · Blackwell Ultra- GPUs
- 72 Blackwell Ultra GPUs
- Cooling
- Direct-to-chip liquid, no air variant
- Distribution
- 208 V three phase, busway
- Availability
- Shipping
Vera Rubin NVL72
NVIDIA · Rubin- GPUs
- 72 Rubin GPU packages, 36 Vera CPUs
- Cooling
- Direct-to-chip liquid, CDU per rack or row
- Distribution
- Three phase; 800 VDC on the roadmap
- Availability
- Full production announced CES 2026, volume H2 2026
The widest disagreement in the table. See the note on conflicting figures below before designing to any single number.
Rubin Ultra NVL576 (Kyber)
NVIDIA · Rubin Ultra- GPUs
- 576 GPU dies, 144 packages
- Cooling
- Liquid, with separate cooling sidecar racks
- Distribution
- 800 VDC
- Availability
- H2 2027 target
Paired with dedicated power and cooling sidecar racks. Not a rack you retrofit into an existing hall.
Where the published figures disagree
Vera Rubin rack power is genuinely contested in public sources as of late 2026. Some coverage puts it at 120–130 kW, roughly level with Blackwell. Other coverage and at least one OEM sizing its cooling around the platform put it near 227 kW, with a commonly cited range of 190–230 kW. Those are not small differences — they are the difference between a hall you can retrofit and one you cannot.
Several widely repeated Rubin figures are also disputed as not appearing in NVIDIA’s own published specifications, including a 166 kW rack figure and a 2,300 W per-GPU TDP. Secondary coverage frequently quotes superseded numbers from earlier announcements.
What to do about it: for anything you are committing budget to, get the number from the vendor’s current published specification or from your OEM in writing. Design the electrical and mechanical envelope to the top of the plausible range, not the middle. Being oversized on a 200 kW rack costs money; being undersized strands the deployment.
Naming, which is a mess
The Rubin flagship rack was introduced as NVL144, counting GPU compute dies — two dies per package, 72 packages. It was renamed to NVL72 in late 2025 to count packages instead, and that name was confirmed at CES 2026. Both names refer to the same rack.
VR200 appears in partner and supply-chain material for the same system. NVL144 CPX is a different product — the long-context Rubin CPX variant, sold as a standalone rack rather than the flagship.
If you are comparing quotes or articles, check which name is being used before assuming two sources describe the same hardware. A good deal of the numerical disagreement above traces back to this.
Per-accelerator draw
Useful when you are building a load from a device list rather than buying a rack-scale system. Eight 1,000 W accelerators in a 4U chassis is an 8 kW thermal problem in one box before you count CPUs, memory, or fans.
| Accelerator | Generation | TDP | Memory |
|---|---|---|---|
| A100 | Ampere | 400 W | 40 / 80 GB HBM2e |
| H100 SXM | Hopper | 700 W | 80 GB HBM3 |
| H200 SXM | Hopper | 700 W | 141 GB HBM3e |
| B200 | Blackwell | 1000 W | 192 GB HBM3e |
| Rubin (per package) | Rubin | Not published | 288 GB HBM4 |
What each density band forces
Density decides the design. Total room load decides the utility conversation, but per-rack density decides whether the room works at all.
Up to 20 kW
- Cooling
- Air with containment. ASHRAE TC 9.9 calls above 20 kW high density.
- Power
- 208 V single or three phase, conventional whips
- Structural
- Standard raised floor, no special review
20–50 kW
- Cooling
- In-row units or rear-door heat exchangers. The practical ceiling for air.
- Power
- Three phase, higher amperage feeds
- Structural
- Check floor loading; cabinets get heavy
50–150 kW
- Cooling
- Direct-to-chip liquid. No air-only option exists in this band.
- Power
- Three phase busway; individual whips become impractical
- Structural
- Floor loading review required. CDUs add weight and need facility water.
Above 150 kW
- Cooling
- Liquid throughout, often with dedicated cooling sidecar racks
- Power
- 800 VDC. At 1 MW per rack a legacy 54 VDC system would need roughly 200 kg of copper busbar per rack.
- Structural
- Purpose-built. Not a retrofit into an existing hall.
Using these numbers
Treat everything here as a planning starting point, not a specification. Rack power depends on configuration, workload, and what else shares the cabinet — a partially populated rack does not draw its rated figure, and a fully loaded one under sustained training draws close to it.
The numbers move. Platforms in the announced and roadmap rows have shifted more than once between announcement and shipping, and secondary coverage lags. This page is reviewed quarterly and after each major vendor keynote; confirm anything load-bearing against the vendor’s current specification.