Amodei Breaks Silence: Anthropic Never Called for a Ban on Open Models, but Wants to Restrict Chinese Compute

Dario Amodei published a public post on Anthropic's website on Monday, July 27, stating in bold letters that the company has never called for a ban on open-weight models. The exact phrase is: Anthropic has never advocated for a ban on open-weights models. This gesture was made after a public coalition formed by Nvidia, Microsoft, Meta, OpenAI, and Google targeted Anthropic as a supposed vector for restrictive regulation in the U.S. Amodei separates the defense of open models, which he classifies as a public good when they do not have dangerous capabilities, from what he described as his central concern: the compute and weights currently available freely in China.

The Immediate Trigger

The text comes three days after Moonshot AI published the weights of Kimi K3 on HuggingFace, a model with 2.8 trillion parameters in a mixture of experts architecture, the largest open weight in history. Kimi K3 leads six out of seven domains in the Frontend Code Arena and scores 88.3 on the Terminal Bench 2.1, a result that brings the Chinese open model closer to the latest closed generations from Amodei's own lab. Amodei does not mention Moonshot by name but describes the scenario as one in which an increasing amount of compute becomes available below the marginal cost for actors outside the American regulatory framework.

Amodei's response is twofold. He states that Anthropic does not work towards banning the open modality and does not support the moral argument of existential risk against any public distribution of weights. He advocates for three specific pillars of public policy that, in his view, address the real problem: keeping advanced chips out of authoritarian hands, cutting industrial-scale distillation of American models by external actors, and requiring mandatory safety testing for any sufficiently capable model, whether open or closed.

What Competitors Wrote and What Amodei is Accepting

The collective letter signed on July 24 by Jensen Huang (Nvidia), Yann LeCun (Meta), Satya Nadella (Microsoft), Sam Altman (OpenAI), and Sundar Pichai (Google) urged the U.S. government not to impose categorical restrictions on open models, arguing that national competitiveness and supply diversity require plurality. Amodei accepts the premise and publicly adopts the same thesis. The divergence then shifts from being about open models themselves to being about mandatory testing and control of compute exports. In this framing, Anthropic remains aligned with the Department of Commerce, and none of the five labs in the letter formally disagree.

According to the text itself, what Amodei wants from American policy is the second stage of the AI Diffusion rule, which treats compute as a dual-natured commodity and creates a specific licensing category for exportable open weights. Translated for the corporate buyer, this means that if the proposal is adopted, Brazilian, Indian, or European companies that currently fine-tune American open models may need to acquire an authorization number for commercial scale usage.

How the Corporate Buyer Should Read This

The underlying thesis is more significant for the 2027 roadmap than for the decision of this week. If Chinese compute continues to grow at the current rate, the cost of frontier tokens in open models will keep falling, and the proposal for proprietary fine-tuning on free weights is likely to establish itself as a real alternative to consumption via API from American labs. Kimi K3 is already being offered in corporate contracts in Europe through its partner DeepInfra and in India via Yotta.

In Japan, NTT Data announced in July that it will support internal deployments of Kimi K3 in its portfolio of architectures for banks and insurers, a decision that sets a strong precedent for clients like MUFG and Nomura. In Germany, Deutsche Telekom included Chinese and American open weights on equal footing in its Sovereign AI Cloud catalog. In Brazil, TOTVS, Stefanini, and CI&T are already running internal POCs with Llama 4, Kimi K2, and now K3 in client environments subject to LGPD, betting that proprietary fine-tuning will be the most defensible route to meet data residency requirements.

The Point That Amodei Does Not Close

Mandatory testing for frontier open and closed models is the most challenging part of the proposal. In a closed model, the original developer conducts the test and signs the attestation. In an open model, who is responsible for testing the version running within the client? Amodei writes that Anthropic supports governance at the initial training level, not at the level of each fork. However, the text does not resolve what happens when an autonomous agent with custom fine-tuning of Kimi K3 causes harm in a regulated jurisdiction. Therefore, the discussion that the post opened is just the first round.