Washington Orders Federal Agencies to Remove Chinese DeepSeek Models

The United States has taken a notable step in its technological competition with China, after the Department of Homeland Security issued a directive requiring federal civilian agencies to stop purchasing any services or products developed by the Chinese AI startup DeepSeek, and to remove any already installed from government information technology systems. The decision is not limited to the company's own applications, but extends to solutions and services built by third parties on top of its technologies and models, while allowing for exemption requests in specific cases.

What Is the Stated Rationale Behind the Decision?

The Department of Homeland Security justified the directive by citing concerns over how the company handles user data. The text of the order warned that sensitive information belonging to US government users could become accessible to Chinese authorities under the laws in force in China. This reframes the matter from a commercial debate about model quality into an issue of national security and data protection.

Notably, this is not the first restrictive measure. It was preceded by legislation passed in December 2025 that limited the ban to national security systems operated by intelligence agencies. The new directive, however, broadens the scope to include federal civilian agencies and covers both future procurement and existing systems.

The Chinese Response

The decision did not pass without comment from Beijing. Chinese Foreign Ministry spokeswoman Mao Ning stated that her country opposes "the overstretching of the concept of national security and the politicization of technology and trade issues," affirming that China would defend the legitimate interests of its companies. At the same time, media reports indicate that Beijing is considering export restrictions on some 23 AI-related technologies, including chip designs and certain advanced models, on the grounds of addressing security and misuse risks.

Why Has DeepSeek Become the Eye of the Storm?

The Chinese company rose to prominence in January 2025 when it released its R1 reasoning model, which caused widespread disruption across the AI industry and contributed to a sharp sell-off in Nvidia shares. The company adopted a dual strategy that gave it rapid visibility:

• Open-weight models that allow local deployment and customization.
• Operating costs and pricing far lower than those of US competitors.
• Competitive performance in reasoning and coding tasks against models from companies such as OpenAI and Meta.

This combination made it an attractive alternative for developers and organizations worldwide, but it also multiplied the regulatory and governmental scrutiny directed at it.

The "Model Distillation" Issue

Alongside data concerns, the company faces technical allegations of a different kind. A report released in September by US parties accused DeepSeek and five other Chinese companies — including Alibaba and Moonshot AI — of using "model distillation" on a large scale to extract knowledge from American AI systems.

Model distillation is a training technique that leverages the outputs of an existing, advanced model to train another, smaller or less costly model, theoretically allowing some of the original model's capabilities to be transferred without bearing the cost of training it from scratch. OpenAI informed US lawmakers that it had detected suspected attempts by DeepSeek to distill its models, while Anthropic announced that it had identified unauthorized attempts by DeepSeek and other Chinese labs to collect outputs from the Claude model.

What Does This Mean for Organizations?

For organizations outside the US government framework, the decision imposes no direct obligation, but it sends a clear signal that a model's "origin" and the location where data is processed have become core factors in any technology adoption decision. This underscores the importance of several practical measures:

• Inventorying the models and services used internally, including those embedded within third-party tools.
• Reviewing data retention policies, server locations, and the legal frameworks governing them.
• Classifying sensitive data and preventing it from being passed to cloud models that have not been security-assessed.
• Maintaining a technical fallback that allows the model provider to be replaced without rebuilding the entire stack.

The bottom line is that the AI market is heading toward a more clearly defined geographic and regulatory split, where comparisons between models are no longer based on performance and price alone, but on compliance, digital sovereignty, and the full chain of trust behind every model.

✦ بقلم فريق دروب أيديا