E-Paper
Anthropic, Nvidia and Big AI: Great rift on open models, chips and distillation
Beneath the open-weight AI debate lies a far bigger struggle over sovereignty, geopolitical influence and most importantly, frontier AI leadership.
Days after a number of artificial intelligence (AI) companies came together in support for open-weight models, Anthropic has weighed in on why they didn’t join the group of signatories that includes Google, Microsoft, Meta, Nvidia, Perplexity, SpaceX and later OpenAI as well. Anthropic CEO Dario Amodei clarified in an open letter that “Anthropic has never advocated for a ban on open-weight models”, and called out accusations that the AI company wanted to protect its business of proprietary AI models. He suggested a three point template, which on many issues, walks a different path.
“Open-weight models that don’t have dangerous capabilities are a public good: they don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers,” Amodei said in the letter. He warned that AI models from Chinese frontier labs are bad for the AI industry in the US, and believes the open weights debate shouldn’t be intertwined by the threat he perceives from Chinese AI.
“My primary concern is the risk that authoritarian governments — not solely the Chinese Communist Party (CCP), although the CCP is clearly the most capable threat — build AI models that are more powerful than those built by the US, and use them to achieve permanent military superiority or perpetrate incredibly deep repression of their own people,” he said.
Amodei also warned that powerful AI models may be misused to carry out cyberattacks or biological attacks, and may have serious alignment problems.
Three point plan, and a different view
Amodei’s three point plan to address concerns about AI from other countries and fears of powerful AI models being misused by entities including governments and enterprises inevitably posing global risk, seems to differ in opinion with the signatories of the letter.
First, Anthropic wishes for a ban on selling powerful AI chips or chipmaking equipment to Chinese AI companies. He also calls for cracking down on smuggling of hardware. “China has limited domestic production capacity, and therefore, due to the scaling laws, cannot build more powerful models than the US without US chips. This is the most efficient and direct way to block threat #1, and by hampering the training of models that are out of reach of US law,” Amodei said.
AI chip exports, including those from Nvidia, are tightly regulated by U.S. policy, featuring conditional approvals for specific markets like China. The US government permits restricted sales of Nvidia’s H200 AI chips to approved buyers in China under specific conditions, including a 25% surcharge. Earlier this month, some shipments went out, to fulfil orders placed by Chinese companies including Tencent and Bytedance. Jeffrey Kessler, the US under secretary of commerce for industry and security, specified at the House Foreign Affairs Committee hearing these are a “very small quantity of chips, so it’s trivial”.
The newer, cutting-edge Blackwell processors remain entirely blocked from export to mainland China.
Secondly, Anthropic calls for a ban on “industrial-scale distillation operations”, with the advantage for distillers being this is a much more compute-efficient method to training models, than training one from scratch. “It allows China to build much better models than its number of chips would ordinarily enable, and thus partially evade chip bans. Distillation does not allow the CCP to obtain equivalent or superior AI capabilities to the US, but it can bring the Chinese frontier to within a few months of the US frontier,” Amodei said.
Recent leaps by Chinese AI companies is where this fear stems from. According to the 2026 Stanford HAI AI Index Report, the performance gap between top US and Chinese AI models has shrunk to a razor-thin 2.7%, with the US leading in total top-tier model quantity and private investment, while China leads in research volume, patents, and citations.
There may be partial truth in Amodei’s claims that Chinese AI will struggle to develop without chips from US chipmakers. For context, Zhipu has clarified that the GLM-5.2 training happened solely on Huawei’s AI chips, and no Nvidia hardware was used. The release of GLM-5.2 by Z.ai represents a critical milestone in the decoupling of global tech supply chains too. There is a geopolitical play too: By releasing GLM-5.2 under an unrestricted MIT license, China is positioning itself as a reliable, accessible alternative for AI capabilities, in a stark contrast to increasingly closed-door policies of US frontier labs.
Third, Anthropic says all sufficiently capable models, open and closed, should go through mandatory safety testing. “ Whether open models do or don’t pose an increased risk, and whether that risk can be mitigated, is something that should emerge from testing, rather than be decided in advance — and there may be promising methods for improving the safety of open-weight models, including recent research from Anthropic on modular training strategies,” Amodei said.
This month, Sir Demis Hassabis, Nobel Laureate as well as co-founder and CEO at Google DeepMind, proposed a new standards body for regulating frontier AI and AGI, or artificial general intelligence. Hassabis specifically noted the Financial Industry Regulatory Authority (FINRA), suggesting the AI regulator should also have a similar structure including a board that includes independent leading technical experts and open-source representatives.
Differences of opinion
In the letter, signatory AI companies and platforms had urged policymakers to not impose what they illustrated as “premature restrictions on open models that stifle competition or drive innovation overseas”. In the original letter first shared by Nvidia CEO Jensen Huang, there seems to be some confusion about the differences between open weights and open source.
The Nvidia led broader industry coalition argues that open-weight AI operates like open-source software — transparency inherently makes it more secure. By letting thousands of researchers inspect the models, vulnerabilities are found and patched faster. They argue this broad access disproportionately helps defenders, who need advanced AI to combat cyberattacks from bad actors.
In contrast, Anthropic rejects that argument arguing that open-weight models are inherently riskier for dangerous capabilities (like cyber and biological attacks) because, once released, guardrails can be stripped away and the model cannot be recalled. Amodei worries about a massive structural asymmetry — especially in biology — where an attacker with a powerful AI could weaponise a pathogen quickly, while defenders would still take years to develop a vaccine or cure.
A growing false equivalence between open weights and open source is becoming clearer, even though are fundamentally different. True open-source software allows anyone to inspect the underlying source code line-by-line. In AI, ‘open weight’ merely means you are given the precompiled, final neural network to run locally. It does not mean you have access to the training data or the exact training recipe. To claim that open weights provide the same level of transparency as open-source software is structurally false. You can run the model, but it is still effectively a black box.
The signatories to the letter released by AI companies last week seemed to support the method of distillation, not too long ago, the entire narrative was around ‘distillation is stealing’. They want to treat distillation as a standard, legitimate technique for improving and validating models, comparing it to the long open-source tradition of building upon existing tech. They warn policymakers against sweeping restrictions on it, arguing that only direct misappropriation should be targeted.
Anthropic views industrial-scale distillation as a critical national security loophole. Because authoritarian governments (specifically China) face US export bans on advanced chips, Anthropic argues that they use distillation to efficiently bypass those limits, bringing their AI capabilities dangerously close to the US frontier. Anthropic explicitly calls for a crackdown on the practice.
Ultimately, a visible divide between the Nvidia led coalition, and Anthropic captures defining tension of modern AI policy, which needs to balance economic advantages of open innovation, against unprecedented security risks of frontier models, often proprietary. Eventually, policymakers will likely try to thread the needle — protecting an open-weight ecosystem for everyday AI, while simultaneously taking cognisance of Anthropic’s pitch for strict safety testing and hardware controls of frontier capabilities.
