us orders federal agencies to remove deepseek

Felix Abagnale 20 Oct 2026 3 min read

big tech

the us department of homeland security published an order on deepseek. federal civilian agencies must remove deepseek products from their it systems and cannot buy new ones. the deepseek ban also covers third-party products and services built on deepseek's models. it applies only to federal civilian agencies. state governments, private companies and individual users are not covered.

what the order covers

dhs issued it under the federal acquisition supply chain security act of 2018 (fascsa), according to the federal register notice. the order has two parts: removal and procurement. agencies have to remove deepseek products that are already installed, and new contracts cannot include them. it names hangzhou deepseek artificial intelligence co. and its affiliates.

the third-party clause is the wide part. the notice covers products and services developed using deepseek implementations. that means a us vendor's product falls under the order if it runs on deepseek's models. because deepseek publishes open weights, that can include tools built and hosted entirely outside china.

dhs also made public a july memo. in it, the department told agencies to monitor traffic in and out of government websites for risks linked to deepseek.

why dhs says deepseek is a risk

the notice itself does not list the evidence. it says only that dhs reviewed a recommendation from the federal acquisition security council.

per reuters, dhs says deepseek's models pose "multiple threat vectors". the department warns that sensitive data could end up in china if government employees or vendors use them. that data could give beijing intelligence on us citizens and government agencies. it also says the models may contain malware or other security vulnerabilities. cnbc reports that the agency cited open-source research that found several critical vulnerabilities in deepseek's ios app.

the bans before this one

deepseek was already blocked in parts of the us government. the commerce department and the navy banned it on government devices in 2025. the defense information systems agency moved to block access in january 2025, according to bloomberg, as cited by cnbc. cnbc also reports that congress restricted deepseek in intelligence community systems in december 2025. new york, tennessee and virginia have state-level bans.

the order came two days after deepseek shipped v4.1-flash-1017 on oct 17.

deepseek moves onto huawei chips

the same day the order went public, the information reported that deepseek is helping huawei design its next ascend chips. deepseek reportedly sends huawei the compute problems it hits when it runs large models. huawei works on memory bandwidth and the interconnect between chips. per huawei central, huawei plans to unveil the ascend 960dt training chip in q1 2027. zerohedge, citing the same report, says huawei will give deepseek most of the first 960dt supply.

the report fits what deepseek has done in public. on sep 30 it said on wechat that it built ascend programming tools with huawei, reuters reported. the two companies also tuned a supernode of 128 ascend 950 chips. deepseek pushes tilelang, an open-source language for ai chips, as a simpler option than cuda from nvidia. in september bloomberg reported that deepseek plans to order at least 160,000 ascend 950dt chips for a data center in inner mongolia.

both sides close the door

three days before the dhs order, china moved from the other side. on oct 16 its commerce and science ministries proposed export controls on the country's most powerful ai models. the draft sets three thresholds: 2t parameters, 10^26 training flops or a global top-10 rank. a model past any of them would need a license before its weights leave china. api and app access would stay outside the controls. the same list covers chip design know-how at 14nm and below.

the two rules target the same thing from opposite directions. beijing wants a say over which flagship weights go abroad, while its labs keep selling access through their own apis. washington now cuts its civilian agencies off from both routes: deepseek's own products, and any us-hosted service built on its weights. so for a vendor selling ai to the us government, it now matters where a model was trained, not only what its license allows. deepseek, in turn, is building on a chinese stack: huawei's ascend chips and its own tools in place of cuda.