
Season’s greetings from Berkeley Square
M3CA / Sovereign AI
Private AI in the UK, built around your company data
M3CA Sovereign AI gives your company a UK private-compute environment isolated from the public internet. Use approved company knowledge for retrieval or specialist model training, with access, evaluation and data transfer agreed before work begins.
Live now
Taking clients today
400TB
Private storage capacity
40Gbps
Internal transfer speed
UK
Your data stays in the UK
What Sovereign AI does
01
Private LLM sandbox
Compile, fine-tune and train a language model on your own documents, so it answers in your terminology, policies and tone, and nobody else’s.
02
Secure SME hosting
An isolated environment for each client to host applications, data and models, kept separate from other tenants and from the public internet.
03
Build your own data centre
For organisations that need their own facility, we design and specify a private data centre around your data, workloads and compliance needs.
How it stays secure
Not on the public internet. Not shared.

Isolated from the public internet
The environment is not exposed to the public internet. This reduces exposure, but does not remove risks from people, transferred files or configuration; access and operating controls still matter.
Controlled access
Access is agreed per client and per person, with data brought in and out through controlled transfer agreed at the start.
UK data residency
Your data, training material and models stay in the UK.
Certified practice
Our security certifications are shared with prospective clients as part of due diligence.
Who it is for
01
Regulated SMEs
Firms that want AI but cannot send client data to public AI services.
02
Legal and finance
Confidential matters, contracts and records that must stay private.
03
Healthcare administration
Administrative and operational data kept in a closed environment.
04
Manufacturing and engineering
Drawings, quotes and process knowledge that define your competitive edge.
How it works with Company AI
01
Discovery and deployment choice
We agree the task, users, data classification and access needs. We compare retrieval, specialist fine-tuning and other deployment options before recommending a private workload.
02
Controlled data preparation
We agree the transfer route, document ownership, retention and permitted uses. Approved material is organised, with representative evaluation examples kept separate from training data.
03
Build and evaluate in the sandbox
The agreed model or knowledge workflow runs inside the isolated environment. We test source accuracy, company terminology, failure handling and human review against your acceptance criteria.
04
Operate and hand over
The engagement defines access, update procedures, recovery responsibilities and exit arrangements. Your team receives operating guidance and practical training through Company AI.
Relevant client work
Private by design, specific by task.
Names remain withheld until clients give written permission. The full references and technical evidence can be shared privately during due diligence.
Read the AI case studiesA UK manufacturer
A private company knowledge assistant made historic quotes and specifications searchable with cited source documents. The client name remains withheld pending written permission.
A UK care group
A specialist classifier routed non-clinical administrative emails for human review. It did not make clinical decisions, and the client name remains withheld pending written permission.
M3CA private compute
The live 400TB, 40Gbps environment provides the isolated infrastructure behind this offer. Technical scope and security evidence are shared during due diligence.
Does private AI mean training a model from scratch?
No. Retrieval can ground an existing model in approved company documents without retraining it. Fine-tuning can suit stable specialist tasks. Training a foundation model is a separate programme requiring a justified dataset, compute plan and evaluation scope.
Can staff use it from a public website?
The isolated environment is not a public web endpoint. We agree how authorised staff reach the workload and move approved data during discovery. A public application integration requires a separate connectivity and security design.
Do 400TB and 40Gbps guarantee model performance?
No. These describe infrastructure storage capacity and internal transfer speed, not model accuracy, GPU capacity, training time or capacity reserved for your company. Workload sizing and acceptance tests are agreed separately.
What evidence should we review before commissioning?
Request the proposed architecture, access and transfer controls, applicable certification evidence, retention and deletion arrangements, recovery plan and model test results. Requirements and responsibilities are documented for your engagement.
Who owns the data and the trained model?
You do. Your data and the model trained on it are yours, and are not used for any other client.
How do files get in and out without internet access?
Through a controlled transfer method agreed with you at the start, so every movement of data is deliberate and recorded.
What happens if we leave?
We return your data and model in an agreed format, then securely delete them from our environment.
How does this differ from Company AI?
Company AI covers strategy, integration, training and fine-tuning. Sovereign AI is the private infrastructure where that work can run. Most clients use both.
How much does it cost?
Every engagement is scoped to your data and workloads, so we quote on request.
Start a conversation
Your AI, your data, your data centre.
Tell us about your data and what you want the model to do. We reply personally with scope and a quote.
Book a 20-minute fit callWarm wishes for Christmas and the new year, from all of us in Berkeley Square.