Grace Blackwell Ultra · 72 GPUs · one NVLink domain
72 Blackwell Ultra GPUs and 36 Grace CPUs joined into a single liquid-cooled NVLink domain: one rack that behaves like one enormous GPU.
In an ordinary GPU server, eight GPUs share fast NVLink and everything else talks over the network. GB300 NVL72 removes that boundary. Its NVLink Switch spine joins all 72 Blackwell Ultra GPUs in the rack into one domain with 130 TB/s of bandwidth, so a model can be spread across the whole rack as if it were a single accelerator.
That matters most for the hardest AI work: very large mixture-of-experts models, long-context reasoning and serving thousands of users at low latency. NVIDIA quotes up to 10 times better responsiveness per user and 5 times more throughput per megawatt than Hopper, adding up to a 50-fold increase in AI factory output.
It is also a facilities project. The rack is fully liquid cooled and draws well over 100 kW, so the hall, the coolant loop and the power feeds have to be ready before it arrives. Planning that is where we start.
Every GPU can reach every other at full NVLink speed. Tensor and expert parallelism stay inside the rack instead of crossing the network, which is where large models usually lose time.
With 2 times the attention performance of Blackwell and 1.5 times the dense FP4 throughput, GB300 is built for reasoning models that generate long chains of thought for every request.
20 TB of HBM3e on the GPUs plus 17 TB of LPDDR5X on the Grace CPUs, coherently connected, so huge key-value caches and models stay in memory.
We plan the coolant distribution unit, secondary loop, water quality, leak detection and maintenance regime with your facility team, not around them.
NVIDIA's published figures for the rack.
| Configuration | 72 NVIDIA Blackwell Ultra GPUs, 36 NVIDIA Grace CPUs (2,592 Arm Neoverse V2 cores) |
|---|---|
| GPU memory | 20 TB HBM3e, up to 576 TB/s |
| CPU memory | 17 TB LPDDR5X, 14 TB/s |
| NVLink | 130 TB/s across the rack |
| FP4 Tensor Core | 1,440 | 1,080 PFLOPS |
| FP8 / FP16 Tensor Core | 720 / 360 PFLOPS (sparse) |
| Networking | ConnectX-8 SuperNICs, 800 Gb/s per GPU, on Quantum-X800 InfiniBand or Spectrum-X Ethernet |
| Cooling | Fully liquid cooled, rack-scale |
| Software | NVIDIA AI Enterprise, Dynamo and NIM for inference, Mission Control for operations |
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.
GB300 NVL72 is a data-centre engineering project. These are the questions we answer with you first.
| Power | Well over 100 kW per rack; we confirm the exact figure for your configuration |
|---|---|
| Cooling | Direct liquid cooling with a coolant distribution unit and facility water supply |
| Floor | A fully populated rack weighs well over a tonne: floor loading and access routes are checked |
| Network | Spine capacity for 72 × 800 Gb/s per rack, plus storage and management networks |
| Hosting | Few UK halls can take this density today; we can shortlist colocation sites that can |
One accountable UK team from sizing to day-two operations
Power, cooling loop, floor and fibre designed with your facilities team or colocation provider.
Rack-level delivery, positioning, coolant commissioning and power-on, with the vendor where required.
NVLink, fabric and thermal acceptance tests before production workloads.
Mission Control, monitoring, firmware life cycle and SysOps247 cover once it is live.
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.
It is 18 compute trays and 9 NVLink switch trays in one rack. To your software, the 72 GPUs form one NVLink domain, so frameworks can treat the rack as one very large accelerator.
Only if the room can supply well over 100 kW to one rack and has liquid-cooling infrastructure. Most customers place it in a specialist colocation hall. Our site survey settles the question early.
DGX GB300 is NVIDIA's own version of the rack with NVIDIA's software and support package. Partner GB300 NVL72 systems use the same architecture from other server makers. We can supply either and explain the trade-offs.
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.