Hardware, sized to your business case.
We start from what you need AI to do and the budget you have, then size the system, source it through authorized distribution channels and commission it so it is productive on handover. One contract covers delivery, installation and warranty.
Why own the hardware.
Renting tokens is the fastest way to start. Owning makes sense once the volume and the data justify it.
Lower cost at steady volume
When usage is high and constant, owned capacity usually costs less per token than renting. We model the break-even for you.
Control of sensitive data
Prompts, documents and outputs never leave systems you operate, which simplifies risk and compliance reviews.
Capacity you can allocate
Share one cluster across teams and projects, and choose which models run on it.
Not sure whether to buy or rent? See ways to buy.
Data centre GPU servers
Air- and liquid-cooled multi-GPU servers for training, fine-tuning and high-volume inference.
Workstations
Desktop and rack workstations for model development, simulation and visual computing teams.
Edge and embedded
Compact systems for factories, vehicles, hospitals and sites with limited power and no data centre.
Networking and storage
High-speed fabrics and parallel storage sized so the GPUs are never waiting for data.
Typical engagements
| Scale | Typical use | What we deliver |
|---|---|---|
| Single node | Pilot, department assistant, private fine-tuning | One GPU server or workstation, installed, configured and benchmarked on your models |
| Rack | Production inference for one organization | Servers, switching and storage in one rack, with power and cooling plan and serving software |
| Multi-rack pod | Enterprise AI platform | Pod design, phased procurement, facility coordination, commissioning and operator training |
From sizing to lifecycle.
Every engagement covers the same four stages, so nothing is left for you to discover after the hardware arrives.
Sizing
From tokens per day, model size and latency targets to GPU count, memory, interconnect, storage and power draw.
Sourcing
Systems and components procured through authorized distribution channels, with lead times confirmed before you commit.
Integration
Rack, power and cooling planning with your facilities team, then installation, burn-in and the software stack.
Lifecycle
Warranty handling, spares, firmware and driver updates, and a path to expand the cluster as demand grows.
Is your facility ready?
The four questions that decide whether a system can be installed on your timeline. We help you answer them.
Power
Available kilowatts per rack, redundancy and the electrical work needed to connect.
Cooling
Whether air is enough or the room needs liquid cooling, and what that changes.
Space and weight
Rack positions, floor loading and access for delivery.
Network
Uplinks, switching and the security zone the system will sit in.
Reference configurations.
Typical building blocks we quote. Final configuration, memory, networking and storage are specified per order.
| Configuration | GPUs | Total GPU memory | Typical use |
|---|---|---|---|
| HGX H200 server | 8 x H200 SXM | 1,128 GB HBM3e | Large-model inference, fine-tuning |
| HGX B200 server | 8 x B200 SXM | 1,440 GB HBM3e | High-throughput serving and training |
| HGX B300 server | 8 x B300 SXM | About 2.1 TB HBM3e | Frontier-scale models, long context |
| RTX PRO server | Up to 8 x RTX PRO 6000 Blackwell | Up to 768 GB GDDR7 | Private inference, rendering, virtual workstations |
| Developer workstation | 1 to 2 x RTX PRO 5000 or 6000 | 48 to 192 GB GDDR7 | Model development, visualisation |
Planning your facility.
Power and cooling decide what you can install and when. These are indicative maximums for planning, and we confirm figures against the vendor datasheet for your exact configuration.
| System | Indicative maximum power | Cooling |
|---|---|---|
| 8-GPU Hopper server (H100, H200) | Up to about 10 kW per server | Air-cooled |
| 8-GPU Blackwell server (B200) | Up to about 14 kW per server | Air-cooled or direct liquid cooling |
| RTX PRO 6000 Server Edition | Up to 600 W per GPU, plus CPUs and fans | Air-cooled, passive cards need server airflow |
A single 8-GPU server can draw more than a conventional rack of enterprise servers. Most existing colocation contracts and room circuits need upgrading before installation, which is why we start with the facility.
What every quote includes.
So you can compare us with any other supplier on equal terms.
Bill of materials
Every component with manufacturer part numbers, quantities and software licences.
Price and validity
Itemised pricing in your currency, with the period the price holds.
Lead time and delivery
Confirmed lead times from our distributors, delivery terms and who carries the freight and import duties.
Warranty and support
Manufacturer warranty terms and the support options available for your site.
Facility requirements
Power, cooling, weight, rack space and network needs, written down before you commit.
Installation and acceptance
The installation plan and the tests that must pass before handover.
Describe the system you have in mind.
Even a rough workload is enough for a first configuration and lead-time estimate.