Grace Blackwell · 72 GPUs · liquid cooled
The first rack-scale Grace Blackwell system: 72 GPUs in one NVLink domain for trillion-parameter models, with strong double precision for science.
GB200 NVL72 introduced the idea of the rack as the unit of AI compute. Thirty-six GB200 superchips, each pairing a Grace CPU with two Blackwell GPUs, are joined by NVLink Switch trays into one 72-GPU domain with 130 TB/s of bandwidth.
NVIDIA quotes up to 30 times faster real-time inference for trillion-parameter models, 4 times faster training and 25 times better energy efficiency than H100. It also keeps substantial double-precision performance, 2,880 TFLOPS of FP64 per rack, which makes it a strong choice where the same capacity serves AI and scientific computing.
As GB300 becomes the default for new AI-only builds, GB200 remains a mature, well-understood platform, and a good fit for research and sovereign programmes that value FP64 and proven software.
Models too big for one server stay inside the rack's NVLink fabric, avoiding the network as the bottleneck for tensor and expert parallelism.
Unlike Blackwell Ultra, GB200 keeps strong FP64, so simulation and AI users can share the same rack-scale capacity.
Liquid cooling and Grace CPUs give much better performance per watt than an equivalent Hopper estate, which NVIDIA puts at 25 times for inference.
Frameworks, schedulers and inference servers have been tuned for GB200 for longer, which lowers the risk for first-time rack-scale buyers.
NVIDIA's published figures for the rack.
| Configuration | 36 Grace CPUs, 72 Blackwell GPUs (36 GB200 superchips) |
|---|---|
| GPU memory | 13.4 TB HBM3e, 576 TB/s |
| CPU | 2,592 Arm Neoverse V2 cores, 17 TB LPDDR5X at 14 TB/s |
| NVLink | 130 TB/s across the rack |
| NVFP4 Tensor Core | 1,440 | 720 PFLOPS |
| FP8 / FP16 Tensor Core | 720 / 360 PFLOPS (sparse) |
| FP64 | 2,880 TFLOPS |
| Cooling | Fully liquid cooled |
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 a GB200 NVL72 rack needs.
| Power | Around 120 to 130 kW per rack (published estimate) |
|---|---|
| Cooling | Direct liquid cooling with coolant distribution and facility water |
| Floor and access | Rack weight, door heights and lift capacity checked for rack-level delivery |
| Network | 800 Gb/s-class fabric to other racks and storage |
One accountable UK team from sizing to day-two operations
Power, cooling loop and access planned with your estates team or colocation provider.
Rack-level delivery, coolant fill, power-on and burn-in.
NVLink, fabric, FP64 and AI benchmarks recorded against agreed targets.
Monitoring, firmware and software life cycle, and SysOps247 if you want us to run it.
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.
For AI-only work, GB300 is the newer and faster option. For mixed AI and FP64 science, or where a proven platform matters most, GB200 is still an excellent choice.
Yes. Both connect at 800 Gb/s-class speeds to InfiniBand or Spectrum-X, so they can share storage and a scheduler, although each rack remains its own NVLink domain.
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.