Anthropic’s CEO says Nvidia’s H200 too powerful for China • The Register
Anthropic CEO Dario Amodei isn’t happy about the US allowing Nvidia to sell GPUs to Chinese companies, and likened the decision to giving nuclear weapons to an adversary.
“The CEOs of these [Chinese] companies say it’s the embargo on US chips that’s holding them back,” he said in an interview with Bloomberg at the World Economic Forum in Davos, Switzerland this week. “I think it’s a big mistake to ship these chips.”
Amodei’s comments come just over a month after the Trump administration announced it would allow shipments of Nvidia H200 accelerators to Chinese customers so long as Uncle Sam gets a 25 percent cut of the revenues. It’s now up to Chinese authorities to allow local buyers to acquire the GPUs.
Anthropic wants stricter controls on AI exports, a hawkish position that contrasts sharply with chipmakers like AMD and Nvidia, which have warned that closing the door on China, which is home to roughly half of the world’s AI researchers, would result in a technological decoupling.
“The US currently leads in advanced semiconductor technology and export controls capitalize on the trend of computing power doubling every two years, so while US chip technology continues advancing, China’s progress is slowed,” Anthropic argued in a statement last northern spring.
Amodei’s thoughts on the matter don’t appear to have changed much in the months since then. “We are many years ahead of China in terms of our ability to make chips,” he said this week.
Access to these chips would put Chinese model devs, like DeepSeek, in a better position to compete with the west, particularly for enterprise adoption.
Many of the most capable Chinese models are open weights, which is to say, anyone can download and run them so long as they have the hardware. Open weights models running entirely on prem offer enterprises assurances that their data won’t “accidentally” find its way into a training dataset.
American model builders have mostly locked their models behind APIs with vague promises that they won’t use customers’ data for further training, while simultaneously wrestling copyright holders in court over alleged copyright infringement.
Amodei’s argument about the threat posed by Chinese model builders is overblown, but that that could change if China were allowed to import more sophisticated silicon from the US.
“I think they never really caught up that much,” he said. “There was a lot of excitement around DeepSeek, but the truth was… those models are very optimized for the benchmark.”
In other words, Chinese developers make sure their models look good on paper, but may not perform well when applied to real-world tasks.
The fact remains that enterprises don’t have many options outside of China if they want a frontier-class open weights model.
Despite this, Amodei sees OpenAI and Google as Anthropic’s main competitors. “When competing for contracts, we see Google and we see OpenAI. Every once in a while, we see a couple other US players. I have almost never lost a contract to a Chinese model.”
The key word there is “almost,” as it suggests Chinese developers are finding some buyers.
Anthropic hasn’t released any of its flagship models to the public as open weight models, but has engaged in some collaborative efforts with select customers – compute partner Amazon Web Services being the most prominent. ®
