GPU models, and the systems built around them.
We quote the accelerator that fits your model size, user count and facility, not the one with the biggest number. Availability and pricing are confirmed per order.
| GPU | Memory | Best for | Form factor |
|---|---|---|---|
| NVIDIA B300 | 288 GB HBM3e | The largest models, long-context and reasoning workloads, fewest GPUs per model | SXM (HGX systems) |
| NVIDIA B200 | 192 GB HBM3e | High-throughput serving and training of large models, lowest cost per token at scale | SXM (HGX systems) |
| NVIDIA H200 | 141 GB HBM3e, 4.8 TB/s | Large-model inference and fine-tuning, long context windows, high concurrency | SXM or PCIe (NVL) |
| NVIDIA H100 | 80 GB HBM3 (SXM), 94 GB HBM3 (NVL) | Proven training and inference for mid-size to large models | SXM or PCIe (NVL) |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB GDDR7 with ECC, about 1.6 TB/s | Private inference up to 70B-class models, large-scene rendering, mixed AI and visual workloads; can be split into up to four isolated instances | Server Edition (passive) and workstation editions |
| NVIDIA RTX PRO 5000 Blackwell | 48 GB GDDR7, about 1.3 TB/s | Developer and studio workstations, compact servers, model prototyping | PCIe workstation card |
| NVIDIA L40S | 48 GB GDDR6 | Cost-efficient inference, render farms and virtual workstations in standard servers | PCIe |
| NVIDIA GeForce RTX 5090 | 32 GB GDDR7 | Individual artists and developers, rendering and models up to about 32B | Desktop and workstation card |
Memory and bandwidth figures come from the manufacturer's published specifications and vary by configuration. NVIDIA, GeForce, RTX and related product names are trademarks of NVIDIA Corporation. GeForce products are intended for desktop and workstation use and are not licensed for data centre deployment.
Solutions
Frontier-scale platform
B300 or B200 systems for the largest models, long context and the lowest cost per token at scale.
Large-model platform
H200 systems for long-context, high-concurrency language model serving.
Training and fine-tuning
H100, H200 or B200 nodes with high-speed networking and storage, sized to your data.
Private inference server
RTX PRO 6000 or L40S in air-cooled standard servers, for teams serving models on their own premises.
Rendering and visual computing
RTX PRO workstations and L40S servers for studios, architecture and engineering firms, render farms and virtual workstations.
Developer workstations
RTX PRO 5000, RTX PRO 6000 or RTX 5090 workstations for model development, simulation and visualisation.
Which GPU should you ask about?
A starting point before we size your workload properly.
Memory is the limit
If your model or context window does not fit, start with the H200, or the B200 and B300 for the largest models. More memory per GPU means fewer GPUs per model.
You want standard servers
The RTX PRO 6000 Blackwell Server Edition runs in standard PCIe servers and suits private inference and mixed workloads.
You are training or fine-tuning
H100 and H200 systems with high-speed GPU interconnect are the usual choice, and we size networking and storage with them.
You render or design
RTX PRO cards give you large memory, ECC and certified drivers for professional software. The RTX 5090 suits individual artists on smaller scenes.
You are building and testing
RTX PRO workstations give developers local capacity before you commit to a rack.
Not sure which GPU fits?
Send your model, expected users and site. We reply with a recommendation and lead times.