Your data stays your data

Most AI vendors ask you to trust a privacy policy. We would rather show you the architecture. Nearly everything we build is self-hosted, which means the models run on computing hardware under your control and your documents, drawings, quotations, customer records and support history never leave that boundary. This page sets out exactly how that works.

Every Ascentis system is deployed in one of three ways. The software and the assurances are identical across all three; what changes is who owns the metal and where it sits.

Ascentis Private Compute is not a GPU cloud and we do not rent capacity. We invest in the machine, install and operate it, and run the system we built for you on it. It never takes on workloads we did not build.

How isolated is it? That is a contract decision, and we would rather be straight about it than overpromise. Where the work is time-sensitive, used by many people at once, or falls under defence or regulated requirements, the machine is single-tenant — yours alone. Where the workload is light and periodic — a market intelligence digest that runs for three hours on a Wednesday night, say — it makes no commercial sense for a machine to idle for the other 165 hours, so the hardware may be shared across a small number of projects we have built and operate, with logical isolation between them. Which of the two applies to you is agreed before we start and written into the contract. If you want the detail on what owning the hardware yourself actually involves, our 2026 build, benchmark and cost guide shows the real numbers.

Residency is a property of the arrangement, not an add-on you pay extra for. A UK client’s machine sits in the United Kingdom. A French client’s machine sits in France. There is no cross-border replication, no failover to another region, and no analytics pipeline quietly shipping telemetry somewhere else.

This matters more than it used to. Under UK GDPR and the EU GDPR you remain the controller of your data and you are accountable for where it goes. Under the EU AI Act, transparency and human oversight obligations sit with the deployer. An architecture where your content never leaves your own hardware, in your own country, is the simplest answer to both.

• We will not train on your data. Your documents and conversations are never used to train, fine-tune or benchmark a model — not for you, not for another client, not for us.
• We will not send your content to a third-party AI service. Self-hosted models run locally. If a project genuinely benefits from a frontier API, we tell you before we design it, you approve it in writing, and we scope exactly what leaves the building.
• We will not resell your capacity or run a stranger’s workload on your hardware. Where a machine is shared, it is shared only across projects we built and operate ourselves. It is never sold on to third parties.
• We will not run other people’s workloads on it. Your hardware exists to run your system.
• We will not name you without permission. Every case study on this site is anonymised by agreement. The problems, the solutions and the figures are real; the client names are withheld.

We are precise about the difference between holding a certification and aligning to a standard, because plenty of vendors are not.

Our systems are designed to be checkable. Answers cite their sources, so a claim can always be traced back to the document it came from. Where output is customer-facing, a human approves it before it goes out. Where a system makes a recommendation, the reasoning is visible rather than buried in a score.

This is not decoration. It is the difference between an AI system your team trusts after six months and one they quietly stop using.

An AI system that holds you hostage is not an asset, it is a liability with a subscription. So the exit is written into how we build:

• The model weights, configuration and prompts are yours to keep.
• Your data is yours, in open formats, exportable on request.
• The stack is open-source. There is no proprietary runtime you cannot run without us.
• If you are on Ascentis Private Compute and you want to bring it in house, we will talk about the machine.

Can Ascentis staff read our data?

Access is limited to the engineers working on your system, for the purpose of building and supporting it. On an air-gapped deployment, we have no route in at all unless you give us one. Ask us for the specifics of your deployment and we will put them in writing.

Do you train on our data?

No. Not for you, not for another client, not for a general model. Your content is used to answer your questions and nothing else.

We are in a regulated sector. Is this viable?

It is usually the reason clients come to us. Regulated, NDA-heavy and defence-adjacent work is precisely where a self-hosted system beats a cloud subscription, because you can say exactly where the data is and prove it. We are Cyber Essentials certified and align to JSP 936 and DEFCON 658.

What happens if the machine fails?

For systems we operate, hardware failure is our problem, not yours — that is a large part of what Ascentis Private Compute is for. For systems on your own hardware, we agree the support arrangement up front so nobody is guessing on the day it matters.

Is the machine dedicated to us alone?

It depends on the workload and it is agreed in the contract. Single-tenant is standard where the work is time-sensitive, used by many people at once, or subject to defence or regulatory requirements. For light periodic workloads the hardware may be shared across a small number of projects we have built and operate, with logical isolation between them. Capacity is never resold to third parties.

Can we audit it?

Yes. The stack is open-source and inspectable, answers are source-cited, and we will walk your IT or compliance people through the architecture. We would rather have that conversation early than in a procurement review.

Related reading: Off-the-shelf AI vs bespoke AI · Self-hosted AI vs cloud AI · AI ROI calculator