The US government has decided that the cheapest way to deal with a Chinese AI model it does not trust is to remove it from the building. The Department of Homeland Security (DHS) has banned DeepSeek across federal agencies, and the ban reaches software already installed, not just future purchases.
What the ban actually covers
DHS says the worry is simple. Chinese law may allow the authorities there to obtain confidential data belonging to American government employees, and a model used inside federal systems is a convenient place for that data to end up.
The scope is wider than a procurement freeze:
- New purchases. Agencies may not buy DeepSeek technology.
- Existing systems. Installations already in place are covered, which is why the headline talk is of removal rather than prevention.
- Third-party products. Software and services from outside companies that use DeepSeek technology are also caught, so a vendor's AI feature can become a compliance problem for the agency that bought the product.
That last item is the one procurement teams will remember. Checking what you bought is one task. Checking what your vendors quietly built it on is another.
There is a pressure valve. Government bodies can ask for exceptions to the new requirements. The report does not say who grants them, on what grounds, or how many are expected.
Building on an earlier law
This is an extension, not a first strike. In December 2025 a law restricted the use of DeepSeek in national security systems tied to American intelligence. The new measure carries the same logic over to federal civilian agencies, the ordinary machinery of government that was previously outside the restriction.
The report does not give the date of the DHS decision, the legal instrument behind it, or a deadline for removing existing installations. Those details will determine how disruptive the ban is in practice.
Why DeepSeek, and why now
DeepSeek became a geopolitical name in January 2025 with the release of its reasoning model R1. It competed with American systems on the strength of low cost and open weights. Open weights are the model's parameters, which anyone can download and run locally. That openness is a large part of why the model spread, and it makes the ban an interesting problem.
The stated concern is access to data by Chinese authorities. A model downloaded and run on an agency's own hardware sits in a rather different position from a service that sends queries to servers elsewhere. The report does not say whether the ban distinguishes between the two. It only says that existing systems and vendor products are covered, which suggests the rule is about the technology's origin, not just where the data travels.
The distillation accusation
The ban also arrives with another grievance attached. In September, the US accused DeepSeek and five other Chinese AI companies, among them Alibaba and Moonshot AI, of using model distillation to extract information from American systems. Distillation means training one model on the answers of another. It is a standard technique, and it becomes a dispute when the model supplying the answers belongs to someone else and did not agree to it.
OpenAI and Anthropic have both said they suspect Chinese developers of unauthorized distillation of their models. So the American case against DeepSeek now has two strands. One is about what Chinese authorities could take from US users. The other is about what Chinese developers allegedly took from US labs.
Beijing's response
China objected. Foreign ministry spokesperson Mao Ning said her country opposes a stretched reading of national security and the politicization of technology and trade issues. She added that China intends to protect the legitimate interests of its companies.
That is a familiar diplomatic script, and it should be read as one. It does not respond to the specific claim about Chinese law and data access, and the report does not suggest it was meant to.
A separate report says China is weighing export restrictions on 23 AI-related technologies. They reportedly include chip development and some advanced models. That is unconfirmed in the material available here, and no decision is described. If it happens, it would turn a one-directional American restriction into a two-way one.
What to take from it
The DeepSeek ban is less a story about one model than about how AI has moved into the category of strategic goods, alongside chips and telecom equipment. Once that happens, the question for an agency stops being "is this model good and cheap" and becomes "where did it come from, and who else is inside it."
For federal IT staff, the practical work is an inventory: where DeepSeek is installed, and which vendor products embed it without saying so. For everyone else, the lesson is that a model's origin can matter as much as its benchmark scores, and that open weights, for all their convenience, do not come with a passport.
The exceptions process is the part to watch. A ban with a sensible waiver is policy. A ban with a generous waiver is a suggestion with paperwork.