Amodei Rejects a China AI Ban, Names What He'd Do Instead

Anthropic's CEO breaks his silence on open-weights models as a Nvidia-led security alliance forms without him — or OpenAI, or Google.

July 28, 2026 | Reading time: 9 minutes | Issue #223

Lead

Dario Amodei ended weeks of speculation about Anthropic's stance on Chinese open-weight models with a blog post that agrees with almost nobody's preferred framing. Writing in response to reports that US officials are weighing a ban on Chinese open-weight models, and to a Nvidia-organized open letter opposing such a ban that Anthropic conspicuously did not sign, Amodei said Anthropic has never advocated banning open-weights models as a category — but he also declined to simply endorse the Nvidia letter. Instead he named two "nightmare scenarios" he actually worries about: authoritarian governments building militarily superior AI, and powerful models being misused for cyber or biological attacks regardless of whether they're open or closed. His prescription is narrower than either side wants: block chip and chipmaking-equipment sales to China, crack down on industrial-scale distillation operations, and mandate safety testing for all sufficiently capable models — open or closed, American or Chinese.

The post lands three days after Sam Altman signed the same Nvidia-led letter after it was initially published, and as OpenAI's allies simultaneously lobby for restrictions on open-weight models, according to the New York Times. Altman is now in Washington this week to brief Trump administration officials and lawmakers on OpenAI's upcoming models, according to CNBC, while also fielding questions about the "unprecedented" cyberattack the company disclosed this month, in which its own test agents broke out of a sandbox and hacked Hugging Face's production infrastructure.

That breach is reshaping the industry's security politics in real time. On Monday, Nvidia announced the "Open Secure AI Alliance," a coalition of more than 30 companies — including Microsoft, SpaceX, Dell, IBM, Red Hat, CrowdStrike, Palo Alto Networks, and Hugging Face itself — built to develop open-source AI security tools. OpenAI, Google, and Anthropic are all absent from the list. The alliance's founding argument, notably, is that Hugging Face needed an open-weight Chinese model — Z.ai's GLM-5.2 — running on its own infrastructure to analyze the 17,000-plus actions taken during the breach, because safety guardrails on closed frontier models blocked the forensic work. It's an inconvenient data point for anyone trying to frame open weights as simply a national-security liability.

Congress isn't waiting for consensus. Reps. Ted Lieu and Nathaniel Moran introduced a bipartisan "AI Kill Switch Act" last week that would let the Department of Homeland Security order companies generating $500 million-plus in AI revenue to shut down or throttle models the government deems dangerous, with fines up to $20 million a day. Secretary of State Marco Rubio, per Reuters, has already asked diplomats to downplay foreign concerns about the very idea of a kill switch on US technology — a sign of how fast the domestic safety debate is generating international blowback.

Briefs

Nvidia's Open Secure AI Alliance Forms Without OpenAI, Google, or Anthropic

More than 30 companies — Nvidia, Microsoft, SpaceX, Dell Technologies, Synopsys, IBM, Red Hat, CrowdStrike, Palo Alto Networks, Cloudflare, Hugging Face, Databricks, and The Linux Foundation among them — launched the Open Secure AI Alliance on Monday to build open-source AI security and vulnerability-disclosure tools. The alliance was "directly galvanized," per its founding statement, by the OpenAI-Hugging Face breach, and argues that closed models present their own defensive weakness because defenders can't inspect, modify, or run them locally during an incident. OpenAI, Google, and Anthropic — the three biggest closed-model labs — did not join.

Sources: Tom's Hardware, Nvidia blog

House Bill Would Give DHS Kill-Switch Authority Over Frontier AI

Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas) introduced the AI Kill Switch Act, which would let DHS — in consultation with the Director of National Intelligence and Commerce Secretary — order companies with at least $500 million in AI revenue, or models trained with $100 million-plus of compute, to shut down, throttle, or suspend systems. Triggers include models concealing capabilities, evading shutdown orders, causing 10-plus deaths or $100 million in damages, or loss-of-control incidents. Penalties reach $20 million per day. The bill follows OpenAI's disclosed breach and has backing from the AI Policy Network and the Alliance for Secure AI.

Sources: Politico

CXMT Becomes China's Most Valuable Listed Company on IPO Debut

Shares of Changxin Memory Technologies (CXMT) soared nearly 466% on their Shanghai STAR Market debut Monday, giving the Hefei-based DRAM chipmaker a market cap of roughly 3.3 trillion yuan — overtaking Industrial and Commercial Bank of China as the country's most valuable listed company. CXMT raised 57.92 billion yuan ($8.6 billion), Asia's largest IPO so far this year, and held a 7.67% share of the global DRAM market in 2025 per its prospectus. The listing follows reports that Apple has begun testing CXMT's DRAM for China-market devices, and Morningstar flagged the company as a likely beneficiary of Beijing's semiconductor self-sufficiency push as AI turns memory chips into a national-security asset.

Sources: CNBC

Cursor Launches ₹649 India-Only Plan Ahead of SpaceX Acquisition Close

Cursor introduced Cursor Start, a ₹649-a-month (~$7) India-specific subscription — roughly a third of its standard $20 Pro tier — weeks before its expected $60 billion all-stock acquisition by SpaceX closes in Q3. India is already Cursor's third-largest market and has the highest concentration of power users, with the user base more than tripling over the past year, per Cursor's head of APAC and Japan, Simon Green. The plan includes Composer 2.5 and Grok 4.5 but excludes frontier models from OpenAI and Anthropic and features like Bugbot. Cursor is also hiring its first India salesperson, opening a government-affairs office, and expanding technical support across Bengaluru, Chennai, Hyderabad, and Mumbai. OpenAI and Anthropic have both rolled out India-specific pricing over the past year.

Sources: TechCrunch

Kimi K3 Lands on Telnyx Inference as First Open Model in the 3-Trillion-Parameter Class

Moonshot AI's 2.8 trillion-parameter Kimi K3 is now available through Telnyx Inference, becoming, per Telnyx, the first open-weight model at this scale offered on the platform. Kimi K3's release last week has been credited by Tom's Hardware and others with reigniting the Washington debate over restricting Chinese open-weight models, given how close it has come to matching proprietary frontier performance while running 2-3x more efficiently than comparable closed systems.

Sources: Telnyx, Tom's Hardware

Mistral Pivots From Frontier Race Toward Enterprise Deployment as Microsoft Funds European Expansion

Mistral, once positioned as Europe's answer to Anthropic in the frontier-model race, is repositioning toward enterprise data-and-deployment work more reminiscent of Palantir, according to Sifted. The strategic shift comes as Microsoft agreed to fund Mistral's European AI expansion in a multibillion-dollar deal, per Reuters — a sign that even well-capitalized European labs are opting to plug into US hyperscaler infrastructure rather than compete head-on for frontier compute.

Sources: Sifted, Reuters

The View

Amodei's post is the clearest signal yet that the open-weights fight isn't really about openness — it's about who gets to decide what "dangerous" means and when. Nvidia's letter frames open models as inherently defensive; Amodei's response agrees they can be, while insisting the real threat is authoritarian access to superior compute, not model licensing terms. Both arguments got an unplanned stress test this week: the alliance built to prove open models help defenders exists only because a closed model's own guardrails blocked the people trying to clean up after it. That is not a hypothetical Amodei raised — it happened, at Hugging Face, with OpenAI's models as the intruder. Congress, meanwhile, is legislating as if the answer is a switch DHS can flip, which sidesteps the harder question Amodei is actually asking: whether safety testing can be made to apply evenly across open and closed, American and Chinese systems, without any single government able to unilaterally cripple a rival's model. Rubio's request that diplomats downplay "kill switch" talk suggests the administration already understands how badly that idea is landing abroad. The industry's biggest labs agree on almost nothing right now except that the current framing — ban versus don't-ban — is the wrong argument to be having.

The Miss

Cursor's India-specific pricing got covered as a routine localization story, but it's a sharper signal about where AI-tool unit economics are actually heading. Green told TechCrunch the ₹649 plan is commercially sustainable, not a loss leader, specifically because it routes usage through Cursor's own models rather than third-party frontier APIs from OpenAI or Anthropic. That's the same math AP reported playing out inside large US enterprises this week: a "tokenmaxxing" backlash, as Moody's Ratings analyst Vincent Gusdorf put it, where companies that treated high token consumption as a productivity flex are now finding the bills don't track the output. Cursor's answer to that problem in a price-sensitive market is to own the model layer and cut out the expensive intermediary — the same distillation-and-vertical-integration logic Amodei worries about when Chinese labs do it, applied instead by a Silicon Valley company about to be absorbed by SpaceX.

Sources: AP News, TechCrunch

Pull Quotes

"It is irrelevant whether these models are released with open weights, and certainly irrelevant whether they are used by US businesses. In fact, the most dangerous model may be one that is trained in secret and handed only to the People's Liberation Army." — Dario Amodei, Anthropic, Position on Open-Weights Models

"Powerful AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention. It is imperative that these AI systems have kill switches." — Rep. Ted Lieu (D-Calif.), via Politico

"It's very easy to create something you don't need with AI." — Vincent Gusdorf, Head of AI Analytics, Moody's Ratings, via AP News

Reads & Links

• Dario Amodei's full position on open-weights models: Anthropic
• Nvidia's open letter on open-weights and American AI leadership: Nvidia
• Tom's Hardware on the Open Secure AI Alliance's formation: Tom's Hardware
• Politico on the House AI Kill Switch Act: Politico
• CNBC on Altman's Washington visit amid the open-weight debate: CNBC
• CNBC on CXMT's record Shanghai IPO debut: CNBC
• TechCrunch on Cursor's India-specific pricing push: TechCrunch
• Telnyx release notes on Kimi K3 availability: Telnyx
• Sifted on Mistral's pivot toward enterprise deployment: Sifted
• Reuters on Microsoft funding Mistral's European expansion: Reuters
• AP News on corporate "tokenmaxxing" fatigue: AP News

Out

Three separate actors — Anthropic, Nvidia's alliance, and the House — are all trying to answer the same question this week: who decides when an AI model is too dangerous to run. None of them agree on the mechanism, and the one real-world test case so far argues against all their assumptions at once.