Rubin · HBM4 · NVLink 6
72 Rubin GPUs with HBM4 and 36 Vera CPUs in one liquid-cooled rack. If you are building capacity for 2027 and beyond, the facility decisions start now.
Vera Rubin is the generation after Blackwell. The NVL72 rack keeps the familiar shape, 72 GPUs and 36 CPUs in one NVLink domain, but almost every number moves: 288 GB of HBM4 per GPU at up to around 22 TB/s, NVLink 6 at 3.6 TB/s per GPU and 260 TB/s across the rack, and around 50 PFLOPS of NVFP4 inference per GPU, roughly five times Blackwell.
The catch is power. Published estimates put a Vera Rubin NVL72 rack at roughly 190 to 230 kW, against around 120 to 130 kW for GB200 NVL72 and about 40 kW for a Hopper-era rack. Very few UK halls can supply that today.
So the most valuable work for a Rubin-era factory happens long before delivery: choosing a site, securing power, designing the cooling loop and getting into the supply queue. That is the service this page describes.
Around five times Blackwell's inference performance per GPU means far more tokens from each rack, if the site can power it.
Up to around 22 TB/s per GPU, roughly 2.75 times Blackwell, which speeds up exactly the memory-bound decode phase that dominates serving.
ConnectX-9 SuperNICs bring 1,600 Gb/s per GPU, so the fabric must be planned for double today's port speeds.
At around 200 kW per rack, grid connection, distribution and heat rejection decide when you can go live, not the GPUs.
Published platform figures. Final specifications are confirmed by NVIDIA and server makers as systems ship.
| Configuration | 72 Rubin GPUs, 36 Vera CPUs (88 custom Arm cores each) |
|---|---|
| GPU memory | 288 GB HBM4 per GPU, up to around 22 TB/s |
| NVLink | NVLink 6, 3.6 TB/s per GPU, 260 TB/s per rack |
| Inference | Around 50 PFLOPS NVFP4 per GPU |
| Networking | ConnectX-9 SuperNICs, 1,600 Gb/s per GPU, Spectrum-X or InfiniBand |
| Cooling | Fully liquid cooled |
| Availability | In production in 2026, with partner systems from the second half of the year |
Figures are NVIDIA's published specifications for the platform. Where NVIDIA quotes two Tensor Core figures, the first assumes sparsity and the second is dense. The exact build (CPUs, memory, storage, cooling, networking) is agreed with you and confirmed on a written quote.
What to start planning now.
| Power | Roughly 190 to 230 kW per rack; grid capacity and lead time for new supplies checked first |
|---|---|
| Cooling | Direct liquid cooling with facility water sized for the full heat load |
| Site | Purpose-built or upgraded halls; we can shortlist suitable UK sites |
| Supply | Early engagement on allocation, as for every new NVIDIA generation |
One accountable UK team from sizing to day-two operations
Own hall, upgrade or colocation: we compare options against power, cooling and timescale.
Power, cooling and network requirements written up for your engineers or colocation provider.
Early allocation discussions and a phased plan, often starting on Blackwell and adding Rubin later.
Blackwell capacity now, designed to share fabric and storage with Rubin racks later.
NVIDIA data-centre GPUs are allocated by NVIDIA and the server makers, and lead times are measured in months rather than weeks. Complete, tested systems usually get allocation sooner than loose parts, so we quote these platforms as built systems. Shipments outside the UK, and many advanced GPUs within it, need an end-user and end-use check under UK and US export rules before we can confirm supply.
Trade members get our engineers' time on sourcing and design questions, early stock and waitlist priority, and free technical support on every order.
Usually not. Waiting delays the value of AI work for a year or more, and the facility for Rubin takes time to prepare anyway. A common plan is Blackwell now, in a site designed to take Rubin racks later.
We can start the conversation, site planning and supply discussions now. Firm pricing and dates depend on allocation and are confirmed on a written quote.
Tell us what you want it to produce, how many people will use it and where it would live. An engineer will come back with a sized recommendation and an honest view on whether owning makes sense for you.