DeepSeek after the US ban: what UK organisations should check

I first wrote this in February 2025, a few weeks after DeepSeek's R1 landed. It's been rewritten in October 2026, after the US barred DeepSeek from its federal agencies. The original argument is still here, with a frank look at what held up.

In February 2025 I called DeepSeek a seismic jolt. A small Chinese lab had come close to the best US models for a fraction of the cost, then published how it did it. I said I was eating my words, because I hadn't thought that was possible.

Twenty months on, the jolt has a second half. On 20 October 2026 the US Department of Homeland Security published an order removing DeepSeek from federal civilian agencies. So the question has moved on. It's no longer "can they build it?" It's "can you trust what's underneath the product you're using?"

What the US has actually done

The order tells federal civilian agencies to stop buying DeepSeek products and to take out what they already run. Think Treasury, State and Justice. Agencies can ask for an exception, but the default is removal.

Two details matter more than the headline:

• It covers more than DeepSeek's own app. It reaches affiliated products and third-party tools built on DeepSeek technology. If a supplier's product runs a DeepSeek model underneath, that product is in scope too.
• It builds on earlier moves. Congress had already restricted DeepSeek in intelligence community systems in December 2025. This order extends that to the civilian side of government.

DHS gives three reasons. DeepSeek's servers sit in China, where the state has broad powers over private infrastructure. Prompts and account data may travel there. And security researchers have found real problems before, including an exposed database in January 2025 and flaws in the iOS app a month later.

What it doesn't do

It's worth being precise, because a lot of the coverage isn't.

It doesn't apply to US businesses, the public, or state and local government. It doesn't cover Department of Defense national security systems, which have their own rules. And it doesn't apply in the UK at all.

It also isn't a verdict on the engineering. Nothing in the order says the models are bad. It says the US government won't accept the risk of where they come from and where their data might go.

The part every buyer should notice

The most useful line in the whole order is the one about third-party products. It names a problem most organisations haven't looked at yet.

When you buy software with "AI" in it, you're usually buying someone else's model wrapped in someone else's product. A chatbot, a document summariser or a coding assistant might run on OpenAI, Anthropic, Google, Meta, Mistral or DeepSeek. You often can't tell from the outside. DeepSeek models are even offered through Microsoft Azure and AWS, so "we're on a US cloud" doesn't answer the question.

That's a supply chain problem, the same kind we already manage for open-source libraries and hosting. The difference is that very few procurement forms ask about it yet.

Open weights don't settle it

In 2025 I was thrilled by how open DeepSeek was. You could read the method, watch R1 reason step by step and run your own copy. I still think that openness moved the whole field forward.

But I underplayed something. Running open weights on your own servers solves where the data goes. It doesn't tell you what went into the training, or how the model behaves on topics its makers care about. And this year OpenAI and Anthropic have both said they caught Chinese labs, DeepSeek among them, trying to copy their models' outputs at scale.

Openness tells you how something was built. Trust is a separate question, and it's the one buyers are now being asked to answer.

What I got right, and what I didn't

Right: efficiency would spread. The tricks DeepSeek showed off, like mixture of experts and cheaper reasoning training, are now standard. Prices for capable models kept falling, for everyone.

Right: export controls wouldn't stop progress on their own. DeepSeek has kept shipping open-weight models since. Chip restrictions slowed things down. They didn't freeze anyone out.

Wrong, or at least naive: "everyone is suddenly in the game." Technically, yes. In practice the deciding factor stopped being who could build a good model. It became who you're willing to hand your data to. That's a policy and procurement question, not a benchmark one.

Still open: the UK. In 2025 I hoped the AI Opportunities Action Plan would let smaller British teams build on these efficiency gains. As far as I can see, what the UK hasn't done is set out a clear public position on Chinese models in the public sector. The US just did. I'd expect UK departments to start asking their suppliers the same questions, with or without a formal rule.

If you run a UK organisation, here's what I'd do

You don't need to panic, and you don't need to rip anything out this week. You do need to know what you're running.

1. List every AI feature you use, including the ones inside other products. Ask each supplier which model it runs on and where that model is hosted.
2. Separate the app from the model. Typing into DeepSeek's own app sends data to China. A DeepSeek model hosted in a UK data centre doesn't. They're very different risks.
3. Write the question into procurement. "Which models does this product use, who made them, and where do prompts go?" It belongs in every AI supplier questionnaire.
4. Match the model to the data. Public or made-up data is a fine place to try any model. Personal, health or government data needs a model you can account for.
5. If you sell to UK government, get ahead of it. Have the answer ready before a buyer asks. "We don't know what's under our product" won't survive a security review for long.

Where this leaves us

I still think February 2025 was a turning point. Good AI got cheaper and more open, and that hasn't reversed. What's changed is that knowing where your model came from is now part of the job.

At Contic we build AI into products for founders, health brands and government teams. When we choose a model for a client, we write down which one, where it runs and what data it sees. If you're not sure what's under your own stack, talk to us. We'll help you find out.