NVIDIA Vera Rubin is NVIDIA's AI platform for 2026. Its flagship rack, the Vera Rubin NVL72, brings significant demands for UK data centres, including over 200kW of power per rack, 100% liquid cooling and facility water capable of reaching temperatures of up to 45°C.
This guide is for UK infrastructure leads and AI teams planning where to put Vera Rubin. It covers the published power and cooling numbers, how the platform compares with Blackwell, and what to ask a colocation provider before you sign.
Key takeaways
- The NVIDIA Vera Rubin NVL72 rack holds 72 Rubin GPUs and 36 Vera CPUs, and is fully liquid-cooled with no fans in the compute trays.
- A single Vera Rubin NVL72 rack can draw over 200kW, compared with roughly 120kW for a GB200 NVL72 rack.
- NVIDIA designs Vera Rubin for coolant inlet temperatures up to 45°C, letting dry coolers reject heat without chillers.
- Rubin Ultra in the Kyber rack is expected in 2027 at around 600kW per rack, so a hall built for Vera Rubin today should have a route to much higher densities.
What is NVIDIA Vera Rubin?
NVIDIA Vera Rubin is a rack-scale AI computing platform that pairs NVIDIA's Rubin GPU with its Vera CPU. It succeeds the Grace Blackwell generation and is built for training and running very large AI models. The name is shared with the American astronomer Vera Rubin and the Vera C. Rubin Observatory, which are unrelated to the hardware.
The main product is the Vera Rubin NVL72, a single rack of 72 Rubin GPUs and 36 Vera CPUs linked by sixth-generation NVLink. According to The Register's CES 2026 coverage, each rack carries 20.7TB of HBM4 GPU memory and 54TB of LPDDR5X across 18 compute blades and nine NVSwitch blades. Each Rubin GPU has 288GB of HBM4 and is rated by NVIDIA at 50 petaflops of NVFP4 inference. Even the smaller HGX Rubin NVL8 eight-GPU server now requires liquid cooling, so there is no air-cooled way into the Rubin generation.
Vera Rubin NVL72 specifications for data centre planning
The Vera Rubin NVL72 specifications that matter to a data centre are power and cooling, not petaflops. The figures below are the ones published by NVIDIA or reported by the technical press as of November 2026.
| Specification | Vera Rubin NVL72 | What it means for the facility |
|---|---|---|
| GPUs and CPUs | 72 Rubin GPUs, 36 Vera CPUs | One rack is one 72-GPU NVLink domain |
| Rack power | Can exceed 200kW | Several times a typical enterprise rack |
| Cooling | Fully liquid-cooled, fanless trays | Facility water to every rack |
| Coolant inlet temperature | Up to 45°C | Dry coolers can replace chillers |
| Power architecture | 800VDC, fed from a power sidecar rack | Extra rack positions for power equipment |
| GPU memory | 20.7TB HBM4 per rack | No facility impact, but shows the scale |
NVIDIA revised several numbers between announcement and shipping. ServeTheHome reported in October 2026 that shipping Rubin GPUs run HBM4 bandwidth of 19.2TB/s rather than the 22TB/s first quoted, and that NVLink bandwidth for a full NVL72 rack is 216TB/s rather than 260TB/s. Plan around shipping specifications, not launch slides.
How much power does a Vera Rubin rack need?
A Vera Rubin NVL72 rack can draw more than 200kW. Tom's Hardware, reporting from NVIDIA's Engineering SuperLab in July 2026, wrote that a single rack "can easily consume over 200 kW". NVIDIA has not put a single-rack power figure on its own specification page, so treat 200kW as a floor for planning, not a ceiling.
That is a big step up from Blackwell. NVIDIA's guide for the DGX GB200 rack puts it at around 120kW, as we explain in what 120kW per rack looks like. At 400V three-phase and unity power factor, a 200kW rack requires roughly 290A before allowances. For comparison, a 120kW rack requires around 173A.
Vera Rubin also changes how power reaches the rack. NVIDIA is moving to 800VDC, and its current approach uses a power "sidecar": a separate rack of power equipment that takes 415V or 480V AC from the facility, converts it once to 800VDC and feeds the compute racks beside it. In practice, that means you need floor positions and power capacity for the sidecars as well as the compute racks.
What cooling does Vera Rubin need?
Vera Rubin needs direct liquid cooling to every compute tray, with no air-cooled fallback. The compute trays have no fans, and coolant enters through large pipework at the top of the rack and leaves at the bottom. Earlier GB200 and GB300 NVL72 racks used hybrid cooling, with some heat still removed by air.
Vera Rubin uses cold-plate direct liquid cooling. Cold plates sit directly on the GPUs and CPUs to remove heat. Manifolds distribute coolant throughout the rack, while a coolant distribution unit (CDU) separates the rack cooling loop from the facility water loop. This is the same architecture we run as direct-to-chip cooling for GPU estates, and our guide to what a coolant distribution unit does covers the basics.
CDU capacity has to be sized at hall level, not rack by rack. Supermicro's shipping Vera Rubin packages, as reported by StorageReview in September 2026, pair NVL72 racks with in-row CDUs rated at 1.8MW each, deployed with N+1 redundancy.
Immersion cooling is a strong fit for many dense GPU servers, and we offer single-phase immersion for those estates. Vera Rubin NVL72, however, is designed as a cold-plate rack with quick-connect manifolds, so it belongs in a direct-to-chip environment.
Why does 45°C water matter for UK data centres?
A 45°C coolant inlet temperature means a Vera Rubin hall can reject heat with dry coolers and no mechanical chillers. Tom's Hardware reported that if the coolant inlet temperature is 45°C or less, a heat exchanger can cool the whole system without chillers. Removing chillers cuts the energy spent on cooling, which improves power usage effectiveness (PUE).
The UK climate suits this design. The highest temperature ever recorded in the UK is 40.3°C, at Coningsby in Lincolnshire, verified by the Met Office in 2022. For almost all of the year, outside air is well below 45°C, giving dry coolers plenty of margin. Designers still need to size for the hottest days, because a dry cooler needs some temperature difference between the water and the air to work.
How does Vera Rubin compare with Blackwell and Rubin Ultra?
Vera Rubin takes rack power from around 120kW on Blackwell to over 200kW and goes fully liquid-cooled, and Rubin Ultra will push densities much higher again from 2027.
| Platform | Rack power | Cooling | Availability |
|---|---|---|---|
| GB200 NVL72 (Blackwell) | Around 120kW | Hybrid liquid and air | Shipping |
| Vera Rubin NVL72 | Over 200kW | 100% liquid, up to 45°C inlet | Ramping through late 2026 |
| Rubin Ultra (Kyber rack) | Around 600kW | New Kyber rack and cooling system | Expected 2027 |
The Rubin Ultra figure comes from The Register's reporting on NVIDIA's roadmap. For anyone signing a five-year colocation contract, the real question is whether the hall that takes Vera Rubin today can grow towards Kyber densities later.
Is Vera Rubin available yet?

Vera Rubin is shipping, with volume ramping through the final quarter of 2026. NVIDIA announced at CES in January 2026 that the platform was in full production, with systems expected in the second half of the year. ServeTheHome reported at the end of October 2026 that Rubin was mid-ramp and that hyperscalers were making some of the first NVL72 systems generally available, with much more to deploy "over the next few quarters".
For UK buyers, the hardware lead time may now be shorter than the time it takes to secure power. The UK government's AI Growth Zones policy paper sets out grid connection reforms aimed at cutting time to power for AI data centres by up to five years, which shows how long new connections can take.
What should you ask a UK colocation provider before deploying Vera Rubin?
Ask for specific numbers, in writing, on power, cooling and growth. A provider that cannot answer these questions clearly is not ready for Vera Rubin.
- Confirm the sustained power per rack, not the peak, and whether it is available now or depends on a future grid upgrade.
- Check the facility water temperature the loop is designed for. "Liquid cooling ready" does not always mean ready for 45°C.
- Ask how CDUs are sized and made redundant, for example N+1 at row or hall level.
- Ask about 800VDC sidecar racks and how many extra rack positions they need.
- Check floor loading and pipework routes for liquid-filled racks and manifolds.
- Ask about the upgrade path towards Rubin Ultra densities in 2027 and beyond.
- Check connectivity, because each NVL72 rack has 32.4TB/s of scale-out networking bandwidth and needs dense fibre to match.
Example: a UK AI start-up planning four Vera Rubin NVL72 racks should budget for at least 800kW of IT load for the compute racks alone, plus power sidecars, storage and networking. That fits within a single 1.8MW CDU on paper, but N+1 redundancy means at least two.
Can we host NVIDIA Vera Rubin?
Our AI-ready data centres are engineered for sustained densities of 50kW to 200kW per rack, with direct-to-chip and immersion cooling run by resident on-site engineers. Whether a specific Vera Rubin deployment fits depends on the rack count, the site and the power profile.
Frequently asked questions
Is NVIDIA Vera Rubin the same as the Vera Rubin Observatory?
No. NVIDIA Vera Rubin is an AI computing platform made up of the Rubin GPU and Vera CPU. The Vera C. Rubin Observatory is an astronomical observatory, and Vera Rubin was an American astronomer known for her work on dark matter. They share a name but have no technical connection.
What is the difference between Vera Rubin NVL144 and NVL72?
Vera Rubin NVL144 and NVL72 refer to the same rack. At GTC 2025, NVIDIA said it would count each GPU die rather than each package, which produced the NVL144 name. NVIDIA later dropped that plan, so the shipping rack is called Vera Rubin NVL72, with 72 Rubin GPU packages.
Can NVIDIA Vera Rubin be air-cooled?
No. The Vera Rubin NVL72 compute trays are fully liquid-cooled and have no fans, and even the smaller HGX Rubin NVL8 eight-GPU server requires liquid cooling. Any data centre hosting Vera Rubin needs direct liquid cooling with facility water delivered to the rack.
How much does it cost to host a Vera Rubin rack in the UK?

The cost of hosting a Vera Rubin rack depends on the power you reserve, the cooling infrastructure required, the contract length and the site. Reserved power is a major factor, so a rack drawing over 200kW is priced very differently from a standard rack.
Planning your Vera Rubin deployment
Hosting NVIDIA Vera Rubin comes down to three questions: can the site deliver well over 200kW per rack, can it run direct liquid cooling at up to 45°C, and can it grow towards Rubin Ultra densities? Put those questions to any provider early, because securing power can take longer than securing the hardware.
If you are sizing a Vera Rubin cluster for the UK, request a power assessment and our engineers will review your rack count, power profile and cooling needs with you.





