In short
- Your AI runs on dedicated hardware reserved for you, not on a shared service.
- The models are private and chosen for the job. Your documents never train them.
- Your offices reach it over private encrypted links.
- We test newer, cheaper and lighter kinds of hardware in the Labs, and tell you plainly what is ready.
In short
When you ask a question about your documents, a model has to read the question and the relevant papers and write an answer. That takes a lot of calculation, done by specialised chips. This page explains, in plain language, what those chips are, how we use them and what we are still testing.
The kinds of chips behind AI
Not all chips are the same. Here are the families people talk about, in simple terms, and where we stand with each.
Graphics cards (GPU)The all-rounder of AI: fast, flexible and widely supported. Powerful, but costly and power-hungry.
Where we standIn use today
FPGAChips that can be reprogrammed to fit a particular model. Potentially cheaper to run and much lighter on energy.
Where we standTesting
ASICChips built for one task. Very efficient, but not flexible.
Where we standUnder evaluation
Small AI accelerators (edge TPU / NPU)Compact, low-power modules that plug into small computers such as a Raspberry Pi and turn them into capable little neural machines.
Where we standPrototype
Why we do not publish a parts list.The right chip depends on the job, and the market changes every few months. What matters to you is the result: fast answers, private by design.
Private models
A model is the part that understands your question and writes the answer. We use private models, reserved for each customer, and match the size to the task:
- A small, fast model that is always ready for everyday questions and searches.
- A larger model that we start when a harder job needs it, such as summarising a long contract.
Your documents are used to answer you. They are never used to train a model.
How your offices connect
Your files stay in your own systems. Each office connects to the AI over a private encrypted link, and the AI builds an index it can search. Access rules sit in front of everything, so a person only gets results from files they could already open.
A small box in the office
One of our experiments is a small, low-power computer in the office. It joins the office network to the private AI and keeps frequent searches quick. It is still a prototype, and you can follow its progress on the Labs page.
What you can ask us
If you need more detail for a security review or a data protection assessment, ask. We are happy to explain how your deployment is set up, where it runs and who can reach it, under an agreement where that is appropriate.
Next stepCurious how this would look for your offices? Talk to us.