Anthropic's Amodei Denies Backing Open-Weight AI Ban, Points to Chip Controls and Distillation Curbs Instead

How Did Amodei Clarify Anthropic's Position — and Why Was It Accused of Backing a Ban?

Anthropic CEO Dario Amodei posted on Monday, July 27, to clarify that "Anthropic has never advocated for banning open-weight models," adding that a blanket ban would not solve the national-security concerns his company has raised, according to an iThome report. Inside's report on the same July 27 post notes it appeared on Anthropic's official newsroom and was framed as a direct response to the open-weights controversy, opening with the line: "Anthropic has never advocated for banning open-weight models."

The accusation stemmed from what Anthropic did not do: per iThome, when Microsoft, Meta, OpenAI, and Nvidia jointly opposed restrictions on Chinese open-weight AI models, Anthropic did not join the coalition, which led outside observers to suspect the company supported a ban to protect its own closed-model business.

Where Anthropic Splits From Microsoft, Meta, OpenAI, and Nvidia on Open Weights

The rift traces to a specific document. Inside reports that on July 24, an open letter titled "Open Weights and American AI Leadership," led by Nvidia CEO Jensen Huang, went live and saw its signatory count double from 25 companies to more than 50 within a single day; OpenAI and Google both signed on afterward. Anthropic never signed, and Inside states this absence is precisely why critics suspected the company of wanting a ban to shield its own commercial position.

iThome's account of the same episode confirms the composition of the opposing camp: Microsoft, Meta, OpenAI, and Nvidia publicly opposed restricting Chinese open-weight models, while Anthropic stayed out of that joint stance entirely.

What Amodei Proposes Instead of a Ban: Chips, Smuggling, and Distillation

According to iThome, Amodei said his actual concern is that authoritarian governments could develop AI models that surpass the US for military use, surveillance, and repression, or that powerful models could be misused for cyberattacks or biological attacks. To address this, he called for restricting China's access to advanced chips and chipmaking equipment, cracking down on smuggling and evasion of export controls, and curbing industrial-scale model distillation.

Inside's report frames this as a reaffirmation of a standing position: Amodei "reiterated the three measures Anthropic has consistently supported" — not selling advanced chips and chipmaking equipment to China while cracking down on smuggling, curbing industrial-scale distillation, and requiring mandatory safety testing for all capability-threshold models regardless of whether they are open- or closed-source.

Does Amodei Think Open Weights Help or Hurt AI Safety?

Amodei's position is not a blanket defense of open weights. Per iThome, he does not agree that open weights necessarily help safety and defense, or that broadly available AI capability necessarily favors defenders over attackers — he argues this should instead be determined empirically through pre-release testing, not assumed.

At the same time, Inside reports Amodei directly addressing the chip-and-smuggling measure by saying that whether a model is open- or closed-source "doesn't matter" for that specific risk; the most dangerous scenario, in his view, is a secretly trained model handed only to the People's Liberation Army and state security departments. Yet he also concedes, per Inside, that open weights do carry higher risk overall, because guardrails are hard to apply, usage cannot be monitored, and once weights are released they cannot be recalled — a point he backs by citing a report from the UK AI Security Institute.

What Would Amodei's Universal Safety-Testing Regime Cover?

iThome reports that Amodei's proposed testing regime would apply to any model that crosses a defined capability threshold, regardless of whether it is open or closed, and regardless of the country it comes from. Such models would need to undergo pre-release testing for cyberattack risk, biosecurity risk, and model-alignment risk before release. Lower-capability models from startups and academic labs would be exempt from this requirement.

China's Open-Weight Models Are Catching Up — the Backdrop to Amodei's Stance

The timing of Amodei's clarification coincides with fast-moving releases from Chinese labs. Inside reports that DeepSeek released a stable version of V4 on July 24, and Beijing-based Moonshot AI released the full weights of Kimi K3 on July 27. Kimi K3 uses a mixture-of-experts architecture with 2.8 trillion total parameters but activates only 104 billion parameters per token, supports a 1-million-token context length, and takes native multimodal input — making it, per Inside, the largest publicly downloadable open-weight model at the time.

Moonshot AI itself acknowledged, according to Inside, that K3's overall performance still trails Claude Fable 5 and GPT 5.6 Sol, but said the gap has narrowed to single-digit benchmark scores.

What This Means

The two reports, read together, show Amodei drawing a narrower line than either side of the open-weights debate assumes: he rejects a blanket ban (iThome, Inside) but also rejects the premise, held by the Nvidia-led signatory bloc of 50+ companies, that open weights should simply be left unrestricted (Inside). His own test — mandatory pre-release safety checks for any sufficiently capable model, open or closed, from any country (iThome) — sits alongside his admission that open weights are harder to govern once released, per the UK AI Security Institute citation (Inside). That admission arrives in the same week Moonshot AI shipped a 2.8-trillion-parameter open-weight model whose benchmark gap against Claude Fable 5 and GPT 5.6 Sol has narrowed to single digits (Inside) — the concrete case Amodei's chip-and-distillation proposals are aimed at pre-empting.