Hopper · PCIe · 141 GB HBM3e
PCIe Hopper GPUs with 141 GB of HBM3e each, bridged in groups of two or four: high-memory inference that drops into existing air-cooled racks.
Sometimes the right AI factory is the one you can install next month in the racks you already have. H200 NVL brings Hopper's large, fast HBM3e memory to standard PCIe servers: 141 GB per GPU at 4.8 TB/s, in a dual-slot, air-cooled card that runs at up to 600 W.
Two or four cards can be joined with NVLink bridges at 900 GB/s per GPU, so a model larger than one card's memory can be split across a bridged group. NVIDIA rates H200 NVL at up to 1.7 times the LLM inference performance of H100 NVL, and NVIDIA AI Enterprise is included with the GPU.
It is also a sensible platform where software is validated on Hopper or the facility rules out Blackwell's power density. Note that H200 NVL supply is heavily allocated, and lead times are long.
141 GB per GPU holds large models and long contexts that would not fit on 48 GB or 80 GB cards.
NVLink bridges at 900 GB/s per GPU let a group of cards serve a model together without going through PCIe.
NVIDIA-Certified and MGX servers with up to eight H200 NVL GPUs, in normal air-cooled racks.
An NVIDIA AI Enterprise subscription comes with the GPU, so NIM microservices and supported frameworks are ready to use.
NVIDIA's published H200 NVL figures and a typical server.
| GPU memory | 141 GB HBM3e, 4.8 TB/s |
|---|---|
| FP8 Tensor Core | 3,341 TFLOPS (sparse) |
| FP16 / BF16 Tensor Core | 1,671 TFLOPS (sparse) |
| FP64 Tensor Core | 60 TFLOPS |
| Power | Up to 600 W (configurable) |
| Interconnect | 2- or 4-way NVLink bridge at 900 GB/s per GPU; PCIe Gen5 at 128 GB/s |
| Partitioning | Up to 7 MIG instances of 16.5 GB |
| Server | Up to 8 GPUs per NVIDIA-Certified or MGX server |
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.
Usually the easiest platform to host after RTX PRO.
| Power | Up to 600 W per GPU plus host; an 8-GPU server sits well below HGX power levels |
|---|---|
| Cooling | Air cooled, dual-slot cards |
| Rack | Standard server racks |
One accountable UK team from sizing to day-two operations
We check which models and context lengths fit per GPU and per bridged group.
Bridges, airflow and power validated in a certified server.
NIM, vLLM or TensorRT-LLM configured and benchmarked on your models.
UK support, with NVIDIA AI Enterprise support behind 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 new builds where you can host Blackwell, usually not. It makes sense where air cooling, power limits or a Hopper-validated software stack rule Blackwell out.
H200 NVL is allocation-constrained and lead times are long, currently longer than for RTX PRO cards. We confirm on a written quote after checking with suppliers.
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.