We do not design from diagrams alone.
PureTensor Ltd operates its own Blackwell-generation compute estate in the United Kingdom, because the only credible way to design AI infrastructure for others is to run the same class of system yourself.
We designed, own and operate every layer of it: the silicon, the fabric, the storage, the orchestration and the operations practice on top. No layer is borrowed from a hyperscaler, and no abstraction hides an operator we cannot name. The estate does three jobs.
01
Reference deployment
Proof that we operate the class of infrastructure we design for customer estates: the same fabric, storage, orchestration and recovery path, run in production every day.
02
Engineering laboratory
Where we measure models, serving stacks and failure modes before we recommend them. The results are published by our US research affiliate.
03
Hosting platform
Available as a managed service when you do not need, or do not yet want, an estate of your own.
Run on ours
Our reference platform is also available as a managed service. We operate the infrastructure; you consume the service, in an isolated environment, on hardware PureTensor Ltd owns in the United Kingdom. Use it for:
- Private inference endpoints
- Model evaluation and benchmarking
- Fine-tuning inside your isolation boundary
- Batch and document processing
- GPU workstations for design and engineering
- S3-compatible object storage
- An interim environment while your own system is built
Moving to your own estate later is a planned path, not a migration crisis: the architecture you run on here is the class of system we would build for you. How we build in your estate.
The platform as built
- Isolated per clientClient environments
- KubernetesOrchestration
- Compute
- Storage
- 200G RDMA fabricDedicated interconnect.
- FoundationUnited Kingdom
Current-generation NVIDIA Blackwell GPUs on AMD Zen 5 server platforms, with terabytes of system memory. Sized for private inference, fine-tuning, and GPU desktop workloads — and owned outright, so capacity commitments are commitments.
A 200G RDMA network fabric connects compute and storage tiers, keeping model loading, checkpointing, and data movement off the bottleneck list.
Hundreds of terabytes of erasure-coded distributed storage for durability, with PCIe Gen5 NVMe working tiers for active workloads. S3-compatible object access; geographic replication to our Iceland disaster-recovery site.
Declarative workload deployment across the estate. Client environments are isolated at this layer: separate namespaces, separate storage boundaries, separate audit trails.
The platform is monitored and operated by the engineers named on this site. Every inference request is logged; every operational action is attributable. Autonomous detection and remediation run under human authority, not instead of it.
The measured engineering behind this platform — benchmarks, failure analyses, system records — is published by our US research affiliate at puretensor.ai. What runs here is what is measured there. How we secure and operate it is set out in our operating doctrine, measured daily in a public ledger.