ADC · November 2026
From one model
to a portfolio
Open-weight models on your own infrastructure
Sound familiar?
What if our AI provider is switched off overnight?
We want to be less dependent on the US.
This data can’t leave our network, so it gets no AI at all.
Our AI costs go up with every new use case.
Our pilot runs on one desk machine, for one person.
Our developers aren’t allowed to use Copilot.
What we hear from government and high-security organisations
Most of these come down to the same question. Who is in control of your AI?
Every organisation has dozens of tasks where AI could help
Today: one model for everything, or no AI at all
The future is a portfolio: the right model for each task
One frontier model provider via API
Data too sensitive
Not good enough, or too costly
Not implemented yet
Frontier model
Complex reasoning
Model A
Bulk, non-sensitive work
Model B
Sensitive data
Model C
High-volume, narrow tasks
Model D
Fast classification
The why, the proof and the how
What changed, and why it matters
The proof: our own test
What it means for you
Two terms to know
The trained model is published, so you can run it on your own hardware. Nothing goes back to the lab that made it.
For a single, narrow task, a small model is often just as good as the biggest ones.
Open weights
Download the model, run it anywhere
Small language models
Small enough for a single GPU
The capability race
Most AI solutions don’t need the very best model. More than fifteen open models from 2026 score in this range.
Closed models keep raising the bar
Open models follow close behind
~4 months
Why run your own models
Sovereignty
- Your data.
- Your model.
Nobody abroad can switch it off or change the terms
Business case
With the right model for each task
Sovereignty is no longer hypothetical
The model can be switched off
A US export directive took Claude Fable 5 offline worldwide, EU included, for 19 days.
Weeks later, GPT-5.6 launched to government-approved partners only.
US law follows the data
Microsoft France, under oath: it “cannot guarantee” French citizens’ data won’t go to US authorities.
An EU region doesn’t change who the operator answers to.
Your account can be cut off
After US sanctions, the International Criminal Court’s chief prosecutor lost his Microsoft email.
Sanctions follow the provider, wherever the servers are.
Sovereignty is a scale, not a binary
1Chinese API
Data to China
2US API
Data to the US
3US cloud, EU region
EU data, US law
4EU provider API
EU data, EU law
5Your EU cloud
Your model, rented GPUs
6Your own GPUs
Everything yours
Closer to you, more sovereign
Many organisations today Nothing to switch off from abroadThe model you pick decides the hardware you need
Up to ~30B
For example gpt-oss 20B, Qwen3.8 27B
- Runs on
- One GPU, even a desk machine
- Good for
- Fine-tuned, narrow tasks such as triage
Around 120B
For example gpt-oss 120B
- Runs on
- One 80 GB data-centre GPU, or a desk machine with 128 GB
- Good for
- Drafting and rewriting on sensitive data
500B and up
For example DeepSeek V4.1 (550B), Kimi K3.1
- Runs on
- A server with eight data-centre GPUs
- Good for
- Close to frontier quality, but a big investment
We tried it on a real task
We ran eight model set-ups on support tickets, sorting each one by intent, team and priority. It’s much the same job as triaging notifications.
Fine-tuned: trained further on our own labelled tickets
Fine-tuning teaches a small model your task
You can only do this when you have the weights, which is a big reason open weights matter.
Open-weight model
A generalist that knows language, but not your work
Your labelled examples
Past cases and how your team handled them
A specialist on your task
Small enough for one GPU, inside your own network
What you need
Hundreds to thousands of labelled examples from your history, and one GPU to train on
When it’s worth it
Narrow, frequent tasks with sensitive data, where the same question comes up thousands of times
The catch
When the task changes, you have to train the model again
What we learned
A small test on 100 tickets, so the pattern matters more than the exact numbers.
No trade-off on quality
Out of the box, the best open model did the job about as well as the best closed model we tried.
Your own data is the edge
Trained on our own past tickets, a small open model beat every model we tried, the frontier included.
Fast, on hardware you can own
The trained model ran on one GPU inside our own network and answered several times faster than the frontier.
Remember today: one model, or no AI at all
Now we know how to split a portfolio
Tasks Today Where we are headingOne frontier model provider via API
For every task that uses AI today
Frontier model via API
Complex reasoning
No AI at all
Too sensitive, or no model good enough
Six places to start
Most organisations start with a chat assistant. Here it runs on your own network: staff ask questions, draft text and chat with their own documents, and nothing leaves the building.
For trend detection, the model looks for patterns in sensitive data and writes up what it finds, so the insight can be shared without the data.
For notification triage, the model classifies incoming notifications in the case management system, merges duplicates and sends each one to the right team. Anything it isn’t sure about goes to a person.
For classifying sensitive data, the model labels each document by how sensitive it is, with a confidence score, so people only need to check the unclear ones.
For reports, the model turns sensitive documents into draft reports for external use, and translates them where needed. A separate check flags anything that shouldn’t leave.
For coding, an agent works across all your repositories on an open-weight model inside your own network, so source code never leaves the building.
Internal chat assistant
Trend detection
Notification triage
Classifying sensitive data
Reports and translation
Code assistant
How we get you from no models to a portfolio
Phase 1Prove it
One sensitive, internal task
Work you can’t send to a cloud model today, where a person still checks the result
One open-weight model
Off the shelf and running inside your network. Nothing leaves, nothing trains on your data
One GPU, inside your walls
Or a trusted European host. Enough to test with real users before investing more
Phase 2Build it out
Three to five use cases
Chosen from the work that was ruled out for AI before, by value and by sensitivity
Specialist models added
Fine-tuned small models for volume, decision models that hand unclear cases to a person
A small GPU cluster
Load-balanced and sized on what phase 1 showed: users, peaks and response times
Phase 3Make it standard
AI as a normal option
Every new task gets a model choice, from frontier API to fully offline, based on its data
A portfolio, kept current
New open releases tested and swapped in as they appear, without lock-in
Capacity grows with use
Capacity planning and failover, with offline setups where the data demands it
Three things to remember
Sensitive data is no longer a reason to skip AI
Open models are only months behind the best closed ones, and they run entirely inside your network
A small model trained on your data can beat the frontier
In our test, a small model trained on our own tickets beat every model we tried, on a single GPU
Sovereignty is a choice per task, not all or nothing
How sensitive the data is decides where each task runs, from a frontier API to fully offline
Get in touch
Your AI, on your own terms
Which task would you start with?