Own your compute outright
Deploy proprietary, enterprise-grade AI infrastructure. Secure supply, maximize ROI over time, and build your bespoke sovereign cloud environment.
Request Purchase QuoteACQUISITION SNAPSHOT
Own the stack.
CONTROL
Dedicated hardware
HORIZON
Long-term ROI
NETWORK
Private fabric
SUPPORT
Optional SLA
Build a sovereign AI foundation around your workload.
Loading purchasable GPU models
The NVIDIA stack, generation by generation
We supply NVIDIA accelerators for AI and high-performance computing across four architecture generations. Availability and pricing are quoted per deal — tell us the workload and we come back with what can actually be delivered.
Architecture hierarchy
Ampere
The workhorse generation still deployed at scale for training and inference, and the cheapest route to dense capacity.
- A100 40GB
- A100 80GB
- A30
- A40
Ada Lovelace
Inference, rendering and visualisation. Strong price-per-token where HBM bandwidth is not the constraint.
- L40S
- L4
- RTX 6000 Ada
Hopper
The current industry standard for large-model training. Transformer Engine, HBM3 and 8-way NVLink inside the node.
- H100 80GB SXM
- H100 80GB PCIe
- H200 141GB
Blackwell
LatestThe newest generation, and where frontier training capacity is heading. HBM3e memory and NVLink 5 fabric.
- B200 192GB
- B300 288GB
- GB200 NVL72
Data center accelerators
| GPU | Architecture | Memory | Interconnect | Best for |
|---|---|---|---|---|
| A100 | Ampere | 40 / 80 GB HBM2e | NVLink 3 · MIG up to 7 instances | Cost-efficient training and mixed inference fleets |
| H100 | Hopper | 80 GB HBM3 | 8-way NVLink 4 · MIG up to 7 instances | Large-model training at production scale |
| H200 | Hopper | 141 GB HBM3e | 8-way NVLink 4 · MIG up to 7 instances | Memory-bound inference and long-context serving |
| B200 | Blackwell | 192 GB HBM3e | NVLink 5 · 1.8 TB/s per GPU | Frontier training where interconnect is the bottleneck |
| B300 | Blackwell | 288 GB HBM3e | NVLink 5 · 1.8 TB/s per GPU | The largest models — maximum memory per accelerator |
Manufacturer specifications, shown for comparison. They are not a statement of stock — every configuration is confirmed on quote.
Partitioning and scale
GPU-to-GPU fabric
NVLink
Direct high-bandwidth links between accelerators, so a multi-GPU job behaves closer to one large device. H100 supports 8-way NVLink inside the node; Blackwell introduces NVLink 5 at 1.8 TB/s per GPU.
Multi-Instance GPU
MIG
Partitions a single data-center GPU into isolated instances with their own memory and compute, so smaller workloads can share a card without contending for it. Available on A100, H100 and H200.
Gaming and workstation
- GeForce RTX 5090Current market leader on rasterisation and ray tracing
- GeForce RTX 5080Same generation, lower power and cost
- GeForce RTX 4090Previous generation, still widely deployed
Renting vs Buying
Strategic considerations for your infrastructure scaling.
| PARAMETER | Rent (Cloud) | Buy (Owned) |
|---|---|---|
| Upfront Cost | Low (OpEx) | High (CapEx) |
| Ownership | None | Full Asset Equity |
| Warranty & Maintenance | Managed by Provider | Manufacturer + Optional SLA |
| Best For | Variable workloads, fast scaling | Sustained baselines, sovereign AI |
Buying questions
Where is the hardware delivered?
To your facility, or racked at one of ours if you prefer colocation. Both options are quoted separately so you can compare.
What warranty applies?
Manufacturer warranty applies as standard. An optional SLA adds on-site support and spares — this is the main practical difference from renting, where maintenance sits with the provider.
What are the lead times?
Lead time depends on configuration and current supply. It is stated on the quote before you commit, not after.
Can you help finance the purchase?
Yes. Leasing and staged-payment structures are available so a cluster does not have to land as a single capital event. See the Financing page.
Request a purchase quote
Tell us the configuration and we'll come back with pricing and lead times.