White House officials spent recent weeks debating sanctions, a blacklist naming Chinese open-model makers, and a rule barring US cloud providers from serving those firms, then shelved the plan after opposition from major tech companies, The Decoder’s Maximilian Schreiner reported, citing the New York Times. The dispute splits the industry along a line that tracks who sells closed model access and who does not.

Three officials anchored the internal debate, according to the account. Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent took part. So did Susie Wiles, who holds the title of White House chief of staff. The paper’s reporting rests on interviews with more than six people, some who currently work in government or industry and some who used to. Five of those people described a specific menu of measures under review: sanctions, a trade blacklist naming Chinese open-model makers, and a rule that would have stopped US cloud providers from doing any business at all with those firms. The account also mentions internal discussion over which models federal agencies may use.

The debate followed a run of competitive Chinese releases. Kimi K3, built by Moonshot AI, matches leading US models on several benchmarks, according to The Decoder. Michael Kratsios holds a White House post: he directs the Office of Science and Technology Policy, known as OSTP. On X, he accused Moonshot of distilling outputs from Anthropic’s model Fable and called it theft of US technology. Bessent, that same day, raised the possibility of sanctions, per the account.

Read the alignment through incentives rather than stated principle, and the split stops looking arbitrary. OpenAI and Anthropic sell access to closed frontier models. Free, competitive Chinese open weights make that access easier to substitute, so tighter restrictions protect a business model both companies are preparing to take public. Nvidia, plus Microsoft, Google, and Meta, carry the opposite exposure: Nvidia sells chips that run whatever model a customer picks, and Meta and Google ship their own open-weight releases, so a cloud ban on Chinese open models sets a precedent that could reach their products too. That reading is our own inference from the two sides’ commercial positions, not a claim either company made in the reporting.

Nvidia CEO Jensen Huang led the public pushback. He posted on X for the first time on July 24 to defend open models, and the company built an alliance around open security tooling that grew past 230 members from a small founding group, per the report. One characterization of the standoff, given to the New York Times, called it a “cage fight over the future of the AI industry.” Money and market power were driving it, that assessment held, with security concerns merely “tacked on.” Christopher Padilla, who once served as a Commerce Department official, was the source of that characterization.

A cloud ban would not undo what has already happened. Open-weight models such as Kimi K3 are already downloaded, mirrored on third-party hosts, and fine-tuned by developers worldwide, and no US rule can retract copies that have already left Chinese servers. What a ban on US cloud providers can do is block commercial hosting inside American infrastructure, cutting off a distribution channel without touching the model’s existence anywhere outside it.

Facing that pushback, the administration shifted its stated focus toward making US models more competitive rather than restricting rivals directly, per the report. Industry officials do not expect firm measures until after Xi Jinping arrives in September, the report adds, though it does not name an exact decision date.

OpenAI told the New York Times that open models and closed models each serve different purposes, framing Chinese open-weight progress as an argument for a national policy framework rather than a case against openness. Anthropic declined to comment on the deliberations, though CEO Dario Amodei has previously written about his position on open-weight risk. Separately, the paper is pursuing its own copyright claim against OpenAI and Microsoft, a fact The Decoder notes without linking it to the sanctions debate.

This fight is not the only place the administration is drawing lines around AI this month. The White House has also been convening labs over a voluntary cybersecurity review framework, and AI Insiders has separately covered a distinct sanctions threat tied to allegations of model theft. Different threads, same officials, same few weeks.

Any operator running Chinese open-weight models on US cloud infrastructure should treat a hosting restriction as a live possibility before Xi’s September visit, not a settled no. The safer near-term hedge is confirming now which cloud regions and providers would still serve those weights if Washington narrows the rule to commercial hosting rather than possession.

This account is based on reporting by Maximilian Schreiner for The Decoder, published August 4, 2026, which cites the New York Times.