Washington has ordered DeepSeek out of federal systems, including software built on it. Nobody has said what "built on" means.

The US Department of Homeland Security has ordered all civilian federal agencies to stop buying products from the Chinese AI company DeepSeek and to remove them from their systems, according to a notice in the Federal Register published on 20 October. The removal covers DeepSeek’s own products, those of its affiliates and “third-party products and services developed using DeepSeek implementations”. The order uses the Federal Acquisition Supply Chain Security Act of 2018, a procurement law built for exactly this kind of exclusion. Two days later, TechCrunch reported, citing Bloomberg, that DeepSeek had raised about $15 billion.

What the DHS order requires of agencies and contractors

The order has two parts: no new DeepSeek purchases, and removal of what is already installed. It applies to civilian agencies only. Defence and intelligence have their own issuing officials under the same law. The order does not restrict private companies or consumers in the US.

Contractors are where it reaches further. According to The Hacker News, the standard federal contract clause for these orders means suppliers may not provide or use covered products while performing a federal contract unless they get a waiver. They must check for new orders at least every three months, report covered products to the contracting officer within three business days of finding them, and send mitigation plans within ten. The clause also flows down to subcontractors.

Agencies can ask for exceptions. BleepingComputer reports that they have 60 days to do so, with the Cybersecurity and Infrastructure Security Agency (CISA) handling the requests. The Federal Register notice itself does not give that deadline.

One detail says something about how Washington now handles AI policy. The order was signed on 24 July and published on 20 October, 87 days later. The law lets the issuing official decide when to release an order, so for nearly three months agencies were bound by a decision the public could not read.

Why the phrase “developed using DeepSeek implementations” matters

The order never defines it, and almost everything turns on that phrase. DeepSeek’s chatbot and API are easy to identify. Its open-weight models are much harder to trace. They run on hardware outside China, are available through American cloud platforms, and have been fine-tuned and distilled into many other models. A product built on a fine-tuned DeepSeek derivative may not say so anywhere in its documentation.

The notice gives no reasons either. BleepingComputer points to the familiar concerns: that Chinese authorities could reach DeepSeek’s infrastructure and user data, the possibility of malicious code in the models, and a 2025 security analysis of DeepSeek’s iPhone app that found unencrypted data transmission. DeepSeek’s own privacy policy says it stores user data in China. Those concerns apply strongly to the hosted service and much less to a model running on your own servers, where the weights cannot send data anywhere by themselves. The order does not distinguish between the two, and I think that is a mistake. A rule that treats a downloaded model file the same way as an app that sends data to Hangzhou will be hard to apply and easy to stretch.

The stretching has already started in the background. In September, as TechCrunch noted, a report linked to a CISA advisory accused DeepSeek and five other Chinese AI companies, Alibaba and Moonshot AI among them, of large-scale distillation of American models. Those two companies make the Qwen and Kimi models that many European firms run. Moonshot’s Kimi K3 was the strongest open model in the world when its weights came out in July. If a future order names them, the same undefined phrase will apply to a much larger share of the open-model world.

Why the ban did not slow DeepSeek down

Two days after the order, DeepSeek reportedly closed one of the largest funding rounds in Chinese history, about 100 billion yuan, with the battery maker CATL and Tencent named as potential investors. In June it had raised more than $7 billion at a $50 billion valuation, according to TechCrunch. US federal agencies were never a meaningful customer for a Chinese AI lab, so the order costs DeepSeek little directly. Its value to Washington is the signal it sends to everyone who sells to the government.

China responded in its usual terms. The foreign ministry’s Mao Ning said Beijing opposes “overstretching the concept of national security”. China is also preparing export controls on its own AI technology, including models above two trillion parameters. Both governments are now treating models as strategic goods, a few weeks after they agreed to a channel for AI incidents that left the technology race untouched.

How Europe and Denmark have handled DeepSeek

Europe has acted through data protection law and device rules rather than procurement bans. Italy’s data protection authority blocked DeepSeek from processing Italian users’ data in January 2025, and the ban is still in force, according to a regulatory tracker kept by the MIAI institute in Grenoble. Berlin’s data protection commissioner asked Apple and Google to remove the app from German app stores, without success. The Netherlands and Belgium banned it on government devices, and the Danish Parliament did the same on its work devices in early 2025.

None of those measures touch the open weights. A Danish company can legally download a DeepSeek model, run it in a European data centre and build a product on it, and under GDPR that setup avoids the transfer to China that worried the regulators in the first place. The EU AI Act does not ban any model by country of origin. Europe’s rules ask where the data goes. Washington’s rule asks who made the software.

What a European business should check now

If you sell software or services to the US federal government, directly or as a subcontractor, find out now whether any DeepSeek model sits anywhere in your products, including inside components from your own suppliers. Once you know, you have three business days to report, and the clause expects you to have made a reasonable inquiry into your supply chain in the first place.

Make a list of the AI models in your products and their origins. Many companies cannot say which base model sits under a feature they bought from a vendor. Ask your suppliers in writing, and add the question to new contracts. The same list will help with AI Act documentation and with customer security reviews, which will start asking the same question.

If you run DeepSeek models yourself, separate the two cases. Using DeepSeek’s hosted chatbot or API sends your data to China, which is why Italy’s regulator blocked it. Keep personal and confidential data out of it. Running open weights on your own infrastructure keeps data at home, and is a reasonable choice on cost and quality. But expect American customers, and possibly US-owned acquirers, to treat it as a risk regardless.

Keep a second model ready for any feature built on a Chinese model, so you can switch within weeks if an American customer or a future order requires it. The same advice applied after the Fable 5 shutdown for American models. Governments on both sides now decide which AI you can use for which customers, and a product with one model and no plan B carries that risk in full.