OpenAI Urges Trump Administration to Bar DeepSeek, Turning Rivalry Into Federal Policy

OpenAI urged Donald Trump‘s administration in a March 13 policy submission to bar Chinese AI models, including DeepSeek, from government use, turning competition between model providers into a formal U.S. technology-policy recommendation.

A Proposal With No Immediate Legal Force

The company made the request in a March 13 policy submission, a concrete intervention in the administration’s AI agenda rather than a product release or an informal warning. Officials were asked to establish a government-use restriction and specifically include DeepSeek among the Chinese models excluded from federal systems.

The OpenAI submission did not itself create a ban, change federal procurement rules or establish that the administration had accepted its recommendation. Any prohibition would require an action by the White House, an agency, Congress, or a combination of those institutions, each with its own timeline, legal authority and implementation burden. The Department of Defense and the General Services Administration, which sets baseline procurement standards for civilian departments, have not publicly indicated whether they will adopt the recommendation or on what timeline.

That distinction matters enormously for the practical outcome. A company can recommend a policy that affects national-security procurement, but agencies still need statutory authority, implementation guidance and a defensible mechanism for identifying which systems fall under the restriction. Without those elements, the March 13 filing is a preference, not a rule.

What the submission does is force a choice into explicit language. Rather than allowing a general debate about whether U.S. model developers can compete with Chinese rivals to drift through policy circles, it asks the government to decide, which models may federal users, contractors and integrated systems actually access?

What “Chinese AI Model” Actually Means, And Why That’s A Hard Question

Large language models, the underlying technology at the center of this dispute, are statistical systems trained on text datasets to predict and generate language. They power chatbots, code assistants, document summarizers and decision-support tools now embedded across both commercial and government software.

DeepSeek attracted international attention earlier this year after releasing models that matched frontier U.S. performance at dramatically lower reported training costs, a development that rattled assumptions about the resource advantages American labs held. An enforceable rule would need to specify whether it covers application programming interfaces, downloadable model weights, cloud-hosted tools, fine-tuned derivatives and software that embeds a model inside a larger service. Each of those pathways represents a distinct technical and legal question.

The rule would also need a workable definition of what makes a model “Chinese.” Policymakers could base that on a developer’s country of incorporation, ultimate ownership structure, training compute sourced from Chinese infrastructure, or the entity selling commercial access.

Those tests will not always produce the same list of restricted products. A model trained partly on Chinese cloud infrastructure but sold through a U.S. reseller occupies ambiguous ground under any of those frameworks.

The uncertainty compounds when contractors enter the picture. A federal vendor might use one model directly, route inference through a third-party platform running another, or switch providers after a procurement contract is signed. No existing federal acquisition regulation cleanly addresses mid-contract model substitution.

Also Read: OpenAI and Anthropic compete for business users

Recent TechCrunch reporting described enterprise customers shifting among AI labs as new models arrive. A government restriction would concern a distinct buyer category, but public-sector rules consistently shape the technical assurances and vendor disclosure practices that companies prepare for all customers, including commercial ones.

The company did not disclose a financial impact from the proposal, and the available materials show no response from DeepSeek or the White House. That leaves the request as an explicit policy preference, not a settled rule governing federal AI adoption.

Security Claims Now Face An Implementation Test

The submission frames access to Chinese-developed models as a government-security issue rather than solely a commercial one. That framing asks officials to evaluate model provenance and control, who built it, where weights are stored, what telemetry the system can send, alongside the conventional procurement variables of performance, price and deployment capability.

Those are genuinely different due-diligence tasks, and most federal agencies have not yet developed standard processes for the former. Officials would need to define which specific risks arise from using particular models, data exfiltration, adversarial manipulation, supply-chain compromise, and then determine whether a targeted procurement restriction addresses those risks without sweeping in unrelated software that happens to share an architecture or a training dataset source. Competitive effects on U.S.-based providers alone would not settle the national-security case.

The next meaningful development will be an official response clarifying whether the Trump administration intends to pursue the proposal and, if so, which entities and technical pathways it would cover. Until that response arrives, the March 13 filing stands as a specific request to convert concerns about Chinese AI into a binding federal access rule, with all the definitional, legal and enforcement work that conversion still requires.

Read Next: How U.S. AI rules could reshape model procurement