Washington Moves to Purge Chinese Models
On October 19, the Department of Homeland Security (DHS) made public an order barring federal civilian agencies from procuring products from Chinese AI lab DeepSeek and requiring such bodies to strip its technology, including third-party tools “developed using DeepSeek implementations,” from their IT systems. Homeland Security Secretary Markwayne Mullin signed the order on July 24 under the Federal Acquisition Supply Chain Security Act of 2018 (FASCSA), making it the first civilian-wide exclusion of an AI company under that law. The directive does not apply to the Pentagon, which already operates under a statutory DeepSeek ban, and does not reach other popular models from the People’s Republic of China (PRC).
Chinese models’ share of tokens used by American firms rose from under 2% to over 40% from April 2025 to April 2026, because of their release as free, open-weight products. A report from the U.S.-China Economic and Security Review Commission pointed to an investor estimate that roughly 80% of U.S. AI startups use them to develop derivative models. Washington’s concerns have grown in parallel. In July, after two Chinese models launched within days of each other, it was reported that the administration was reviving plans to restrict Chinese models through Entity List designations, security advisories, and procurement rules.
FASCSA created the Federal Acquisition Security Council (FASC) to recommend excluding or removing high-risk products from federal procurement and systems, with the DHS Secretary issuing the resulting orders for civilian agencies. The DeepSeek directive followed a FASC recommendation, citing the models’ security vulnerabilities and risk of transferring “sensitive data to the PRC if used by U.S. government employees or vendors.” A clause directing the removal of “DeepSeek, DeepSeek-affiliated entities, and third-party products and services developed using DeepSeek implementations” from agency systems may have a wide reach and difficult implementation. Given that DeepSeek releases its models under the permissive MIT license, the order could cover vendor tools running DeepSeek weights on U.S. servers—where no data reaches China—or models fine-tuned on or distilled from DeepSeek outputs, which leave no detectable trace.
The DHS directive fills the last major federal gap for DeepSeek specifically. The Defense Information Systems Agency began blocking DeepSeek on Pentagon networks in January 2025, and the Navy, NASA, and the Commerce Department issued their own internal prohibitions soon after. Congress then codified the restrictions into the Fiscal Year 2026 National Defense Authorization Act (NDAA), enacted in December:
Beyond federal regulation, at least 17 states have blocked DeepSeek on government devices. In January, Texas Governor Greg Abbott went further by adding 26 Chinese firms to the state’s prohibited technology list, though the list applies only to state-owned devices and networks.
Outside the defense package, lawmakers have introduced a series of standalone measures that would extend the restrictions beyond DeepSeek and, in some cases, beyond government systems. None has reached the floor of either chamber:
As with last year’s DeepSeek ban, the FY2027 NDAA is the most realistic and imminent vehicle for broader restrictions. The House passed its version in July, while the Senate bill remains stalled after Democrats blocked a procedural vote. Among nearly two hundred China-related measures in the House-passed bill and the version advanced by the Senate Armed Services Committee (SASC), many focus on the AI threat. The upper chamber’s most significant provisions codify and expand the bans outlined in the agency orders:
Restricting Chinese models inside the government has become a reliable point of bipartisan agreement, yet the dispute over whether restrictions should spread to the private sector has split the legislature, the administration, and the AI industry.
If the Senate’s Sections 1651 and 609 survive NDAA negotiations, they could codify and expand the restrictions already imposed by the Pentagon and the intelligence community. The executive branch, however, does not need to wait for Congress. The FASC can recommend further FASCSA orders against Alibaba, Moonshot, or other Chinese labs at any time, and the DeepSeek directive provides a template for reaching third-party products built on their models. Since July, the administration has also weighed Entity List designations, security advisories, Commerce Department supply-chain rules targeting Chinese open-source models, and an executive order making companies liable for hosting them. Treasury Secretary Scott Bessent has separately threatened sanctions against labs found to have distilled American models.
Enforcement may prove the larger obstacle as Beijing’s state-backed open strategy has made Chinese base models a default starting point for many developers. Published weights cannot be recalled, Chinese models often sit invisibly beneath U.S.-branded products, and no audit tool can reliably trace whether a model was fine-tuned on or distilled from a Chinese one, leaving any lineage-based rule dependent on vendor self-certification. Beijing, however, may ultimately constrain the supply itself. On October 16, after months of talks with its leading labs, China proposed export controls requiring licenses for overseas transfers of large models. Whether through statute, executive order, or Beijing’s own export rules, the open pipeline that made Chinese models ubiquitous in American AI is unlikely to keep flowing on its current terms.
The content in this article represents the views and opinions of the individual author and does not necessarily represent the views or opinions of the Mackinder Forum.