Could this possibly not be about Chinese model builders stealing some of the spotlight from their U.S. competitors?
+Comment
Anthropic has encouraged the White House to strengthen the so-called AI diffusion regulations even more — these measures, slated to go into effect mid-May, aim to limit or prohibit the export of advanced GPUs and accelerators from the U.S. and a limited number of allied countries, potentially impacting companies like Nvidia.
On Wednesday, the chatbot company based in San Francisco stated in a briefing document that these impending export restrictions will not be sufficient to curb the smuggling of chips that are supporting China’s ongoing progress in artificial intelligence.
In general, Anthropic seeks stricter regulations not just for the legitimate sale of AI-specific processors – especially those produced by Nvidia – to different countries, but also aims to rigorously clamp down on smugglers who use nations where these chips remain legal as intermediaries to illegally import them into China, bypassing U.S. export restrictions.
This follows Western AI firms being unnerved when China’s DeepSeek unveiled a free large language model (LLM) at the end of January, which outperformed their models on certain benchmarks with less resource-intensive training compared to their U.S. counterparts.
The export regulations under discussion were introduced by the Biden administration in mid-January, aiming to tackle the threats from overseas AI infrastructures through a framework of three increasingly stringent levels. These categories are outlined below; China falls into the third category.
Nations fortunate enough to be placed in the top tier would maintain unrestricted access to U.S.-designed accelerators, whereas countries in the secondary group would face stringent limitations on imports. Meanwhile, those landing in the lowest category would essentially be prohibited from purchasing these chips—unless they can obtain special permission from the United States government.
Nevertheless, as we highlighted at the beginning of the year, the Trump administration would have the ultimate say on proceeding with this. Anthropocentric is now pushing for the Trump administration to take an even more stringent approach.
“As the leader in cutting-edge semiconductor technology, the US benefits from export controls that leverage the pattern of computational power increasing twofold every couple of years; hence, as American chip technology keeps progressing, China’s development is hindered,” according to an argument made by Anthropic in their document submitted formally [PDF] to the U.S. Department of Commerce.
The widening efficiency disparity suggests that by 2027, nations utilizing outdated chips might experience AI training expenses up to tenfold greater compared to those employing advanced U.S.-made technology.
Anthropic desires changes to the current tier system. Specifically, they hope to witness the Trump administration reducing the quantity of computational resources that tier-2 countries can purchase without prior approval from Uncle Sam.
“At present, nations in tier two have the ability to purchase up to 1,700 Nvidia H100 high-performance chips without requiring governmental approval — an amount valued at approximately $40 million worth of technology. This situation could potentially serve as a loophole for illicit activities,” according to the statement from the LLM laboratory.
This proposal supports boosting funding for tier-two nations provided they possess “strong” data center security measures and implement “bilateral accords that deter illicit trafficking and harmonize technological safeguards.”
Ultimately, Anthropic recommends that the Trump administration boost funding for export enforcement. They argue that increasing support for the Bureau of Industry and Security would greatly enhance control effectiveness.
Anthem’s position sharply diverges from that of leading American semiconductor companies like Nvidia, which has suffered significant financial losses amounting to billions due to U.S. export restrictions.
On Wednesday as well, Nvidia CEO Jensen Huang took the opportunity to urge the Trump administration to ease limitations on selling artificial intelligence hardware abroad.
“We must hasten the global dissemination of American AI technology,” Huang stated during a press conference. “It’s crucial for the administration’s policies and incentives to back this initiative.”
NVIDIA is undoubtedly the biggest provider of AI infrastructure globally, with its GPUs driving most of the current training and inference operations. Consequently, NVIDIA has the most to lose if the Trump administration decides to make stricter regulations on AI dissemination, as proposed by Anthropic.
Huang’s attempts to persuade Trump to relax regulations, which allegedly involved participating in a $1 million-per-person feast with the President at Mar-a-Lago, seem not to be unfolding as the GPU boss had anticipated.
Earlier this month, the semiconductor giant recorded a $5.5 billion expense related to recently enforced export restrictions on the sale of AI accelerators to China, which includes Hong Kong, as well as other countries deemed risky. If the Trump administration proceeds with these regulations, revenues lost due to U.S. export limits could significantly increase.
Anthropic’s approach to U.S. trade policy could appear as an unusual and possibly perilous move considering Nvidia’s prominent role in the AI sector. It would indeed be unfortunate if anything were to affect their supply allocation, wouldn’t it?
Nevertheless, Anthropic depends much less on Nvidia’s hardware compared to many of its rivals. This Californian research facility has strong ties with Amazon Web Services—one of its key supporters—and extensively utilizes AWS’s bespoke Trainium AI accelerators for developing their Claude series of models. Interestingly, Amazon reportedly faced an internal shortage of GPUs in 2024.
Towards the end of last year, AWS announced Project Rainier, which features an AI supercomputer equipped with hundreds of thousands of Trainium2 accelerators. This system is scheduled to go live this year and will offer Anthropic five times the computational power required for training their most recent batch of artificial intelligence models.
Thwarting Chinese advancements in artificial intelligence might be even more crucial for Anthropic than avoiding conflict with Huang. The unveiling of China’s DeepSeek R1 earlier this year sent shockwaves through the tech community. It was reported that the foundational model underlying their large language model needed considerably fewer computational resources compared to those used by rival companies in the West.
Following that, Chinese cloud service providers and model developers have kept releasing increasingly competitive models like Alibaba’s Qwen 2.5 Max, QwQ, along with an entire lineup of Qwen 3 series this week.
However, this isn’t limited to China alone. The U.S. AI export regulations are expected to hinder the distribution of AI accelerators across many parts of the developed world. This situation implies that companies like Anthropic and OpenAI may encounter reduced competition from abroad and have a lower chance of experiencing setbacks due to rival countries’ advancements.
It appears that Anthropic might achieve its desire. According to reports from Reuters earlier today, sources close to the situation have indicated that the White House is considering eliminating the three-level classification system and negotiating individually with foreign nations for a portion of U.S. accelerator facilities. This move could potentially turn these accelerators into additional tools within President Trump’s diplomatic toolkit.
Nvidia has communicated that China can compete effectively against the U.S. without having to smuggle its GPUs.
“A representative from Nvidia informed The Register that China, which houses half of the global AI research community, boasts highly skilled AI professionals across all levels of the technology stack,” in reaction to Anthropic’s observation.
America can’t use regulators to secure an advantage in AI. U.S. companies ought to concentrate on innovation and meet the challenges head-on instead of spinning yarns about bulky, delicate electronic components being concealed in ‘pregnancy bumpers’ or transported alongside living lobsters.
Now, this refers to Anthropic’s filing with the Department of Commerce where they mentioned:
We believe Nvidia is protesting excessively. In response, a representative from Anthropic countered: