Private AI infrastructure, from requirement to production.
PureTensor designs, builds, commissions and supports AI infrastructure inside your own estate, or runs the workload on the Blackwell-generation platform we operate in the United Kingdom. The same engineers, either way.
What we can do with infrastructure: yours or ours.
We are not doctrinaire about where the machine sits. The workload, the data and the regulatory boundary decide that, and you do not have to decide before you talk to us. We can run phase one on our platform, measure it, and size your own estate from the evidence.
Build in your estate
We design, procure, build and commission private AI infrastructure inside your environment: your racks, your network, your security boundary. Then we train your team to run it.
For organisations that need
- Sensitive or regulated data kept inside their own boundary
- Physical control and no external egress
- Integration with existing identity, storage and networks
- Sustained, predictable workloads
- Long-term ownership of the asset
Runs in your estate. Your data never reaches PureTensor infrastructure.
How we buildRun on ours
Run inference, evaluation, fine-tuning and batch workloads on the Blackwell-generation platform we own and operate in the United Kingdom, as a managed service, without buying hardware first.
For organisations that need
- Pilots and proofs of value
- Benchmarks on real hardware before a purchase
- UK-resident processing
- Interim capacity while their own system is built
- Workloads that do not justify an estate of their own
Runs on hardware PureTensor Ltd owns in the United Kingdom.
The platformThe same engineers design, deploy and support both.
From an empty rack to a production AI system.
We work as a forward-deployed engineering team. The people who size the system are the people who rack it, commission it, deploy the models on it and teach your staff to run it. Engage us for the whole arc, or for the stages you need.
01 · Plan
Discover
Workloads, data, users and constraints: what has to run, how fast, and inside which boundary. Where it helps, we run your actual workload on our platform and measure it.
Output: Requirements and workload profile
- 01
DiscoverPlan
Workloads, data, users and constraints: what has to run, how fast, and inside which boundary. Where it helps, we run your actual workload on our platform and measure it.
- 02
DesignPlan
GPU, CPU and server sizing; rack, power and cooling; network fabric; storage; virtualisation and orchestration; identity, access and the management plane; the security boundary.
- 03
ProcurePlan
Specified by us and supplied by leading server OEMs through their established channels. You own the hardware directly and hold the warranty; we own the architecture.
- 04
BuildBuild
Racking, cabling, firmware, hypervisor, Kubernetes, distributed storage and the NVIDIA stack, built to the design and documented as built.
- 05
CommissionBuild
Burn-in, failure testing and acceptance against criteria agreed in advance. Recovery is restored from backup and proven before any production workload goes near it.
- 06
DeployBuild
Model serving, retrieval, agents, embeddings, speech and vision pipelines, integrated with the systems your people already use.
- 07
OperateTransfer
We run the system through stabilisation, with monitoring, alerting and incident response in place from the first day.
- 08
TrainTransfer
Hands-on operator training on your system, not a generic course: day-to-day operation, failure handling, recovery.
- 09
HandoverTransfer
Diagrams, runbooks, recovery procedures, credentials and access, transferred and rehearsed with your team.
- 10
SupportSustain
Remote or on-site specialist engineering for upgrades, incidents and expansion, for as long as you want it.
Not ready to buy?
- Discover
- Run on ours
- Measure
- Decide
Start on our platform. Benchmark the real workload, then choose: build your own estate, stay on ours, or run both.
Services shaped around the problem, not the product.
Each service stands on its own or forms one stage of a larger engagement. The label on each shows where it runs.
AI Infrastructure Discovery
Workload profiling, capacity modelling and a costed architecture, with the option to benchmark on our hardware before anything is bought.
Private AI Infrastructure
Complete AI environments designed for your estate: compute, fabric, storage, virtualisation, orchestration and the NVIDIA stack.
Deployment & Commissioning
Build, burn-in and acceptance testing against agreed criteria. Systems are proven, recovery included, before they are handed over.
AI Workload Engineering
Model serving, retrieval, agents, embeddings, speech and vision, and persistent organisational memory, integrated with the systems your people use.
Migration & Modernisation
Moving AI workloads off hyperscale cloud or ageing on-premises estates onto infrastructure you control, without a flag day.
UK-Hosted AI Compute
Inference, evaluation, fine-tuning, batch processing, GPU workstations and object storage on our UK platform, delivered as a managed service.
Operations, Training & Handover
Stabilisation, operator training, runbooks and a rehearsed handover, so your own team runs the system day to day.
Specialist Support
Remote or on-site engineering for incidents, upgrades and expansion, from the people who built the system.
Built to operate without us.
Dependency is not our business model. A system only its supplier can run is a liability on your risk register, not an asset.
Every serious deployment ships with what your staff need to run it day to day: architecture diagrams, as-built documentation, runbooks, recovery procedures that have been rehearsed rather than merely written, and training on the system itself.
We stay available for the work that genuinely needs a specialist: upgrades, incidents, expansion and new workloads, on terms you choose.
Every handover includes
- Architecture and network diagrams
- As-built documentation
- Operational runbooks
- Recovery procedures, rehearsed
- Acceptance test records
- Operator training
- Credential and access transfer
- Specialist support on request
Five rules, applied to our own estate first.
Trust in an AI system is earned in operations, not in a policy document. We run our own production estate under a written security doctrine, measure it every day, and build the same controls into the systems we deliver.
Rule I: Attack yourself continuously.
NCSC CAF 4.0On our estate
Independent model seats review our repositories adversarially every day and open evidence-bearing pull requests; a human re-runs every claim before merge. Exposure is probed from outside our network, never from inside it, and monitors are audited for the ability to go red.
In what we build for you
Continuous adversarial validation of the system we build, plus red teaming of the AI itself: prompt injection, memory poisoning, tool abuse and agent authority boundaries.
Rule II: Self-heal with bounded authority.
NCSC CAF 4.0On our estate
A deterministic scan-and-fix loop with models only at the edges, promoted from shadow to suggest to canary to auto. Outcomes are graded resolved, recurred or indeterminate, and never green by default.
In what we build for you
Automated remediation with a written authority boundary, and a human approval gate for everything outside it.
Rule III: Three copies, two media, one out of reach.
Cyber Governance CodeUK GDPR Art. 32On our estate
Nightly backups to erasure-coded storage, mirrored to Iceland over a send-only to receive-only topology, so the recovery site can never push damage upstream. Restores are tested weekly by automation and quarterly by a manual drill.
In what we build for you
3-2-1-1-0 backup architecture with immutable copies, an isolated recovery path and recovery objectives that are tested, not assumed.
Rule IV: Trustless by construction.
Cyber EssentialsUK GDPR Art. 32On our estate
Credential rotation with proof of completeness, same-day privilege reduction, a private management mesh, and trust tiers graded by exposure.
In what we build for you
Workload identity, short-lived credentials, four-eyes approval on destructive actions, and zero-trust backup: the identity that writes a backup cannot delete it.
Rule V: Agents run the estate. Agents never touch the vault.
NCSC CAF 4.0On our estate
Audited daily: every model-driven seat is probed from inside its own sandbox to prove no agent credential can read, list or delete disaster-recovery copies. The measurement goes red the day a path appears.
In what we build for you
An operations design in which no agent credential has a path to disaster-recovery storage. It is the control that turns an AI incident into an inconvenience instead of an extinction event.
The live doctrine ledger
Every rule is measured on our production estate and published daily, red included. A ledger that cannot go red is not a ledger. It is published by Varangian, the security practice of PureTensor Ltd.
Every design we propose has a counterpart we run.
We operate our own Blackwell-generation estate in the United Kingdom. It is our reference deployment, our engineering laboratory and, where it suits the workload, a hosting platform for clients. The architecture below is the class of system we design for customer estates, and the platform you run on if you choose ours.
Compute & AI Services
Current-generation NVIDIA Blackwell inference on AMD Zen 5 platforms. Every environment is isolated, auditable and under the client's control.
- GPU node
- GPU node
- GPU node
- Ceph pool
- Ceph pool
- Ceph pool
- Ceph pool
Platform & Storage Services
Highly available erasure-coded storage pools on a dedicated storage fabric. Documented deletion procedures for regulatory compliance.
Offices in Marylebone, London. Infrastructure owned and operated in the Thames Valley data corridor, governed by UK law.
Hardware is a capital decision. Make it on evidence.
The Blackwell-generation workstation GPU at the heart of our platform has risen from $8,565 at launch (March 2025) to $16,000 (August 2026): an 87% increase for identical silicon. US hyperscale cloud prices the same card at rates that pass its full purchase price in roughly seven months of continuous use.
At these prices, sizing errors are expensive in both directions. Because we operate the class of hardware we specify, you can run your real workload on our platform, measure throughput and memory, and buy what the numbers support, or conclude that you do not need to buy yet.
For work that stays on our platform, owned hardware means our cost base does not move when hyperscaler pricing does.
+87%List price, Mar 2025 to Aug 2026. Identical silicon.
- Mar 2025
- $8,565
- Jun 2026
- $13,250
- Aug 2026
- $16,000
Control the infrastructure that runs your AI.
Sovereignty is a property of a design, not a slogan, and different deployments have different residency. We state which one you are buying:
Your estate
No-egress deployment
Models, data and inference stay inside your own facility and security boundary. Nothing reaches PureTensor infrastructure; our engineers work under your access controls.
PureTensor UK
UK-hosted deployment
Processing and storage on hardware PureTensor Ltd owns and operates in the United Kingdom, contracted under the law of England and Wales.
UK + Iceland
Disaster recovery
Where agreed, encrypted backup copies go to our recovery site in Iceland, an EEA state covered by UK adequacy regulations. Never to US-controlled infrastructure.
External APIs
Authorised external models
Some workflows call frontier models from external providers. We do this only where you have authorised it, and we name the provider and the region in writing.
Why the question matters
Microsoft France's chief legal officer testified under oath that he cannot guarantee European customer data will never be handed to US authorities under the CLOUD Act, regardless of where the data is stored.
The United Kingdom placed AWS, Microsoft, Google Cloud and Oracle under direct Bank of England, PRA and FCA supervision. Dependence on hyperscale cloud is now, formally, a UK financial-stability concern.
Where you need jurisdictional certainty, we provide it architecturally, in your estate or on ours, rather than by contract alone. PureTensor Ltd is registered in England and Wales and operates its own hardware in the United Kingdom.
Private by architecture. Sovereign where required.
Four ways to start.
Most engagements begin small and bounded. Every one is scoped in writing, with named engineers and acceptance criteria agreed before work starts.
Discovery & Design
A bounded engagement that ends in a costed, defensible architecture and, where useful, benchmark results from your workload on our hardware.
What's included
- —Workload and data profiling
- —Capacity and sizing model
- —Architecture and bill of works
- —Procurement specification
Build & Commission
We build the designed system in your estate, prove it against acceptance criteria, deploy the workloads and hand it to a trained team.
What's included
- —Build and configuration
- —Burn-in and acceptance testing
- —Workload deployment
- —Training, documentation and handover
UK-Hosted AI Compute
Your workloads on our UK platform. We operate the infrastructure; you consume the service.
What's included
- —Isolated client environment
- —Model serving and private endpoints
- —Monitoring and audit logging
- —Named engineers
Operations & Support
Specialist engineering for systems we built or systems you already run: stabilisation, incidents, upgrades and expansion.
What's included
- —Remote and on-site support
- —Incident response
- —Upgrade and expansion planning
- —Periodic health and recovery reviews
Not sure where to start? Tell us about the workload.
Start a conversationNo commitment required. Talk directly to our engineering team.
Verified written intelligence, delivered.
Argus reads open sources on the companies and events you track, checks every claim against its sources, and delivers written briefs on a schedule. It is a delivered service: you receive the brief, and nothing of your own goes in. Argus works from public sources only.
The engineering is public.
PureTensor, Inc. — our US research affiliate — publishes the engineering records behind the platform: measured systems, documented failures, reproducible results. The infrastructure we run here is the infrastructure measured there.
Engineering Records
System records for the inference runtime, the self-healing operations layer, and the memory substrate — published as measured engineering, not marketing.
puretensor.ai →Threat Intelligence, Delivered
Continuous geopolitical and cyber intelligence briefings for UK organisations — delivered as a managed service on the PureTensor Argus platform, operated under UK jurisdiction.
Argus system record →Infrastructure With Conviction
Most infrastructure today is built for abstraction, not accountability. It promises scale while diluting ownership, performance while obscuring control, and convenience while increasing dependence. PureTensor takes the opposite view. We believe serious organisations require systems they can understand, trust, and direct. That means infrastructure that is not merely available, but intelligible; not merely powerful, but governed; not merely modern, but properly run. We build systems their owners can run without us, and we run our own to the same standard.
Heimir HelgasonLinkedIn ↗
Founder & CEO
Designed and built PureTensor's AI platform from bare metal. 200G RDMA fabric, erasure-coded distributed storage, NVIDIA Blackwell inference. Background in algorithmic trading, cross-border capital markets, and entrepreneurship. Deep expertise in autonomous agent systems, sovereign infrastructure design, and large-scale model deployment.
Ahmed W. KhalilLinkedIn ↗
Strategic Advisor
CFA charterholder and former infrastructure finance lawyer. Career spanning top-tier international law, institutional capital allocation, and European private equity. Advises on capital strategy, investor relations, and market expansion. Fluent in Arabic, English, German, and French.
Alan B. Apter
Advisory Board Member
Investment banker with over 40 years advising multinationals and corporate boards. Morgan Stanley, Merrill Lynch, Renaissance Capital, Eaglestone Group. Led the first NYSE-listed IPO from post-Soviet Russia. Columbia Law School. Founder of Bretalon Ltd.
PureTensor Ltd and PureTensor, Inc. (Delaware, US) are independent companies under common founder ownership. The Ltd builds and operates AI infrastructure in the United Kingdom, including its own compute estate; the Inc is a US research laboratory.
We are building the team: platform engineering, field deployment and commissioning, and operations. Introduce yourself via the contact form.
Jurisdictional Clarity
UK-registered company. Residency stated per deployment: your estate, our UK platform, or an agreed recovery copy. Registrable facts, not marketing claims.
Technical Accountability
The engineers who design your system build it, commission it and support it. Named people, direct access, no ticket queues.
Security by Design
Encryption at rest and in transit, least-privilege access, full audit logging, and recovery copies no agent credential can reach. Applied to our own estate first.
Start a Conversation
Tell us about your project and we'll configure the right environment.
2 Portman Street, Marylebone, London W1
We respond within one business day.
Direct access to the engineering team. No ticket queues. No sales pipeline.