Current-generation data centre accelerators for training and production inference.
Blackwell accelerators are the mainstream choice for new AI infrastructure. They pair very large memory with high bandwidth, so bigger models fit on fewer GPUs.
Who it is for
Typical workloads
Previous: nvidia-hopper
Next: nvidia-blackwell-ultra
Related families
The mainstream high-end training GPU. Sold as part of an eight-GPU server rather than as a card you fit yourself.
Memory
180 GB
Deployment
Rack server
From £32,500
ex VAT
Available to order
The established rack-scale Blackwell system. It is the practical choice where GB300 allocation is tight and the workload is large-scale training or trillion-parameter inference.
Memory
13824 GB
Deployment
Datacentre
From £2,250,000
ex VAT
Allocation only
The standard building block for a modern eight-GPU AI server. It goes into a certified chassis, not an existing server.
Memory
1440 GB
Deployment
Rack server
From £245,000
ex VAT
Available to order
Everything here is priced and confirmed by a person before anything is agreed.