The US Department of Homeland Security has ordered civilian federal agencies to remove all DeepSeek products from their IT systems. The directive also covers third-party software built on DeepSeek open-source models and blocks future purchases of such tools. Beijing called the move discrimination and threatened countermeasures, which matters because two large enterprise software markets are now closing to each other on security grounds.
Scope of the US ban and Beijing response
The ban applies to civilian federal agencies and requires removal of DeepSeek code from government IT infrastructure. Washington cited concern that sensitive data could reach Beijing through the Chinese vendor. China Foreign Ministry said the US stretches the concept of national security too far. The Commerce Ministry demanded the restriction be lifted and warned of resolute countermeasures, creating a formal basis for retaliation.
The restriction is structured to prevent circumvention through repackaged technology. It does not stop at branded DeepSeek applications and reaches software from other vendors that incorporates its open-source models. Future procurement of such products is prohibited for the covered agencies. That design closes a common loophole where a restricted model returns inside another supplier package. For vendors, the origin of model weights and code becomes a compliance attribute.
China regulates foreign AI across the whole market, not only in government systems. Since 2023, AI services need review by the Cyberspace Administration and must meet political guidelines. ChatGPT, Claude and Gemini have no approval there and remain unavailable. State bodies and state-owned companies are expected to replace foreign hardware and software with domestic products over time. Reports also describe iPhone restrictions for officials, while Beijing last week proposed its own AI export and licensing controls.
What the split means for AI buyers
Companies operating in both jurisdictions face parallel vendor lists and duplicated validation work. A model approved for internal use in one market may be unusable in government-linked projects in the other. Procurement teams will need to document model provenance, including open-source components and downstream integrations. Small firms with one standard stack feel this first, while large groups can maintain separate approved toolsets at higher administrative cost.
Security review becomes a permanent part of AI selection alongside accuracy and price. The US case rests partly on the 2017 National Intelligence Law, which obliges Chinese firms and citizens to support intelligence work. US agencies and AI firms have also accused Chinese companies of copying leading US models at scale, a claim Beijing denies. Added to this is content behavior: Chinese models avoid certain topics and repeat state positions with less critical distance than Western systems.
The marker to watch is whether Beijing converts its protest into formal restrictions on US AI products or services. A licensing rule, procurement directive or export control naming American models would confirm a tit-for-tat cycle. If no measure follows within the next policy round, the statement remains diplomatic positioning. Either outcome tells procurement planners how fragmented model supply will stay.
