US approved DeepSeek for its trade blacklist, then sat on it for eight months

A blacklist that was approved but never published

Reuters reported that the US has held off adding DeepSeek, memory-chip maker CXMT and more than 100 other Chinese companies to the Commerce Department's Entity List, the trade blacklist that bars American firms from selling to a named party without a license.

The striking part isn't the decision to add them. That part is done: an interagency committee approved the firms for listing last year. The striking part is the eight months of silence that followed. The Entity List, maintained by the Bureau of Industry and Security, has not been updated since October 2025. Jeffrey Kessler, the under secretary of commerce for industry and security, has sought since late 2025 to avoid new listings for fear of worsening the relationship with Beijing while tariff and export-control talks are still live.

So the security review and the diplomacy point in opposite directions: the firms are deemed risks on paper, and left untouched in practice.

The Anthropic allegation that set the thread off

Folded into the same reporting is a separate claim that did most of the work in the HN discussion: Anthropic said it identified a campaign by DeepSeek and two other Chinese AI labs to illicitly extract capabilities from its Claude platform to improve their own models. OpenAI separately warned lawmakers that DeepSeek was targeting its models too.

That word — "illicitly" — is where the thread turned.

"'illicitly' implies a law that is being violated. What law?" — an engineer questioning the legal basis

The skepticism wasn't just semantic. A commenter who had actually poked at the models offered a reproduction that cuts against the idea of a one-way heist:

"I made Qwen respond it was made by Google with a simple Chinese greeting. But also, I made Sonnet introduce itself as made by OpenAI… Found it like a month ago and it kept working." — a user who ran the prompt himself

The implication: cross-contamination of training data and self-identification quirks run in every direction, not just from West to East. A freelance product engineer added that the Musk-versus-Altman litigation surfaced evidence the same distillation is "regularly done by all labs."

The security case, restated without the politics

Strip away the country-versus-country heat and a sharper reading of the listing fight remains — about who benefits from each outcome. The clearest version came from a commenter who laid out both sides as competing capital interests:

"The reason why some Capital Interests want to blacklist DeepSeek in the US is so that you are forced to buy Claude/GPT/Gemini… The reason why some Capital Interests don't want to ban DeepSeek is so companies that utilize AI have more options, and running your own DeepSeek cluster acts as an independent cost comparison for enterprise inference contracts." — a commenter mapping the incentives on both sides

That framing explains the eight-month pause better than any single motive. A self-hosted open-weight model is also a benchmark — a way for an enterprise to price-check what the closed labs charge for inference. Blacklisting it removes that reference point as much as it removes a security risk.

The same commenter floated the part that should worry anyone with a GPU cluster, regardless of nationality: "any institution with enough compute might be watched in a similar way chem labs are observed by the DEA." That's the supply-chain endgame of treating model weights as a controlled substance — the controls land on compute, not on a flag.

What actually changed: nothing, on paper

The firms remain approved for listing and unlisted. The security review concluded; the enforcement didn't follow. For a US company building on AI, the practical takeaway is narrow and concrete: DeepSeek's weights stay legally reachable for now, but the entire stack sits one signature away from a status change that was decided months ago and is being held in reserve as a bargaining chip.