While leaders from Meta, IBM, and Microsoft were signing manifestos championing open-weight AI models, Anthropic CEO Dario Amodei broke ranks with a statement that doesn’t call for a total ban but introduces a significant geopolitical nuance: uncontrolled openness is a strategic risk when models can be repurposed beyond recall. At a time when Chinese lab Kimi has just released its K3 model under a permissive license, Amodei’s stance isn’t merely a technical disagreement—it’s a warning about who truly controls open weights in AI and how that reshapes global digital sovereignty.

The Nuance Dividing Silicon Valley

According to Chinese media outlet 36Kr in an article published at 9:55 AM Beijing time on July 28, Amodei stated that Anthropic has never advocated for an outright ban on open-weight models. The phrase, seemingly conciliatory, is actually a maneuver to distance himself from both open-source maximalists and advocates of absolute control. The Anthropic CEO added a nuance that dismantles the central argument of the pro-openness coalition, expressing disagreement with the notion that open models are necessarily easier to protect.

Amodei’s position rests on a concrete risk he himself articulated: once open weights are published, developers can hardly retract or control their use. This claim, which may seem like a technical truism, carries profound implications when applied to models with advanced capabilities. It’s not about prohibition, but about recognizing that full openness transfers control from the creator to the end user—and that user could be in any country, including those Washington considers strategic adversaries.

Politico, which covered the same statement on July 27, ran a blunt headline: “Anthropic CEO: Don’t ban cheap AI — but clamp down on China.” According to the U.S. outlet, Amodei continued to distance himself from the unified praise other American tech leaders heaped on cheap open-weight models. The word “cheap” is no accident: the debate isn’t just about safety, but about who can afford to develop frontier AI and who can simply download it.

The Chinese Context No One Mentions Out Loud

Amodei’s statement cannot be understood without the context of China’s open-source AI offensive. In the weeks before his remarks, Chinese lab Kimi released its K3 model under a permissive license allowing commercial use, modification, and redistribution with virtually no restrictions. This is no isolated case: DeepSeek, Qwen, GLM, and Hunyuan have pursued similar strategies, releasing increasingly powerful models under licenses that let any developer in any country download, run, and modify them without oversight.

While U.S. labs debate whether to open their weights, Chinese labs have already opened theirs—and with models that compete in performance with the best the West has to offer. Amodei’s position, though it doesn’t explicitly mention China in his remarks covered by 36Kr, directly targets this imbalance: if the U.S. imposes restrictions on its own labs while Chinese labs release models without controls, the result won’t be safer AI, but a strategic advantage for Beijing.

Amodei’s proposal, per 36Kr, is that risks should be evaluated through rigorous testing, not a priori judgments. That means not labeling all open models as dangerous, but establishing a case-by-case evaluation process. This approach, technically reasonable on its face, is actually a way to slow model releases without an explicit ban that would pit Anthropic against the rest of the industry.

The Internal Fracture in the U.S. Tech Bloc

Amodei’s stance reveals a fracture that had remained latent in Silicon Valley. On one side, Meta, IBM, and Microsoft have publicly defended open-weight models as a way to democratize AI access and prevent power from concentrating in a few labs. On the other, Anthropic and OpenAI have shown growing caution, albeit with different nuances: OpenAI has opted for a closed model from the start, while Anthropic tries to navigate between both extremes.

The problem is that this division isn’t just technical or philosophical—it carries direct regulatory consequences. The U.S. administration, seeking to balance fostering innovation with containing China, finds itself without clear consensus from industry players themselves. If major labs can’t agree on whether to open their models, any regulation will be seen as arbitrary or, worse, as a concession to particular interests.

Amodei’s position, rejecting both total prohibition and uncontrolled openness, tries to occupy a middle ground that is politically useful but technically complex. How do you evaluate a model’s risk through rigorous testing before publication? Who conducts those tests? By what criteria? And crucially, what happens if tests conclude the model is safe, but once released, it reveals unforeseen emergent capabilities?

The Dilemma of Digital Sovereignty in the Age of Open Weights

Amodei’s statement, though framed in technical terms, points to a fundamental issue that this publication has long highlighted: AI is not just a technology, but a vector of sovereignty. Whoever controls the weights of the most advanced models ultimately controls the ability to influence entire sectors—from defense to healthcare, from education to surveillance.

If open-weight models become the global standard, any country with sufficient computing power can download, modify, and deploy frontier AI without developing its own technology from scratch. This is exactly what Chinese labs are doing: releasing models for others to use, thereby building an ecosystem of technological dependency that benefits Beijing.

Amodei’s warning, as covered by Politico, to clamp down on China is therefore not a declaration of trade war, but recognition that full openness of weights disproportionately benefits those without internal restrictions on AI development. While U.S. labs get tangled in debates about safety and ethics, Chinese labs move forward with a clear strategy: release models, gain global market share, and establish de facto standards.

A Forward-Looking Reflection: Control as the New Geopolitical Frontier

Dario Amodei’s statement will go down in AI history as the moment Silicon Valley stopped speaking with one voice. But its true importance lies not in internal disagreement, but in what it reveals about the future of global technology governance.

The dilemma facing the U.S. is the same facing Europe, and soon every country with AI ambitions: how to balance the openness that fosters innovation with the control that protects national security? Amodei’s answer—case-by-case evaluation, with neither total bans nor blind openness—is intellectually honest, but it poses a practical problem: Chinese models are already on the market, already being downloaded, already in use. The debate over whether to open U.S. weights arrives when competitors have already opened theirs.

21st-century digital sovereignty won’t be defined at UN summits or trade treaties, but in model repositories and licenses that permit or prevent cross-border use. Anthropic has pointed out the crack; the rest of the industry will have to decide whether to widen it or seal it. Meanwhile, in Beijing, engineers at Kimi, DeepSeek, and Qwen keep releasing models, waiting for no one in the West to resolve their internal dilemmas. Control over weights is no longer a technical debate—it is the new frontier of digital geopolitics.