Earlier this month, facing backlash for his removal of export controls on Nvidia’s H20 AI chip, President Donald Trump attempted to reassure critics by declaring the H20 “obsolete.” That prophecy quickly proved to be self-fulfilling.
Chinese regulators urged tech firms not to purchase the chip, and demand plummeted. Soon after, Nvidia halted production of the H20, which it had designed specifically for the Chinese market. Trump’s deal with Jensen Huang, reached after months of lobbying from the Nvidia CEO and which included a 15% cut for the U.S. government on H20 sales to China, appeared to be a flop.
But no sooner had the H20 died than reports emerged that Nvidia planned to offer a new chip, this time a downgraded version of its newest Blackwell architecture. Speaking to reporters in Taipei on Friday, Huang said, “I’m offering a new product to China for … AI data centers, the follow-on to H20.” If selling the H20 was enough to draw condemnation from both parties in Congress, a deal on the B30A — as Reuters reported Nvidia’s next China-focused chip is tentatively being called — could prompt an uproar.
“Allowing the export of B30As would be a drastic and unprecedented weakening of export controls,” said Konstantin Pilz, a researcher at RAND’s Technology and Security Policy Center. Pilz’s view aligns with that of many analysts, who told Domino Theory that the B30A is no follow-on to the H20, but a completely different animal, capable of swallowing America’s computing advantage whole.
Trump first signalled that he would be open to allowing Nvidia to sell a less-powerful Blackwell chip to China on August 11. “Take 30% to 50% off of it,” Trump said at the time, adding that it “will be an unenhanced version of the big one.”
The “big one” is Nvidia’s B300, which uses a dual-die architecture — where the integrated circuit is spread across two wafers of silicon — to achieve a world-leading 15 quadrillion floating point operations per second. Last week, Reuters reported that the B30A will use a single-die design to produce half the raw computing power of the B300. According to analysis by Lennart Heim, another researcher at RAND, the power of the B30A would exceed Biden-era export controls by more than 12 times.
Because the B30A is reported to be a single-die version of the B300, Saif Khan, former director of Technology and National Security in the Biden White House, suspects that China could achieve the performance of a B300-powered data center simply by buying twice as many B30As.
“When it comes to AI computing power, what really matters is not the performance of a single chip, but rather how well you can aggregate large amounts of computing power,” Khan said.
The U.S. has long enjoyed a significant lead over China in total compute. A recent estimate found that the U.S. controls 75% of AI compute, with China trailing at 15%.
During the Biden administration, Commerce Department export controls focused on maintaining this lead. Unlike other inputs to AI development, such as research talent and energy supply, raw compute relies on a supply chain controlled almost entirely by a few companies. Nvidia dominates the market for designing the most advanced chips; TSMC dominates the market for manufacturing them; and Dutch company ASML has a monopoly on the extreme ultraviolet lithography equipment essential to TSMC’s processes. That concentration might be bad for competition, but it’s great for regulators: sanctioning a single company can choke the entire supply chain.
Daniel Nystedt, a research analyst at TriOrient Investments, said that this strategy was effective. “People love to complain about the government,” he wrote in an email to Domino Theory. “But in this case, the technical guidelines/rules the Commerce Department created were robust — well thought out, well done.”
Facing limited access to chips, China has invested heavily in other areas. Supercomputers require tremendous amounts of energy to run, human ingenuity to design and data to feed. On these metrics, China has an advantage, said Janet Egan, senior fellow at the Center for New American Security. China is adding to its energy grid at a much faster rate than the U.S. Its weaker copyright protections allow AI companies to freely scrape the internet for training data. Its government is taking advantage of the Trump administration’s recent crackdown on international students coming to the U.S.
According to Egan, those factors only make America’s lead in computing power even more crucial. “Semiconductors and AI chips are the most defensible advantage that the U.S. has,” Egan said.
Trump’s team had pioneered the use of chip-focused sanctions and export controls during his first term, an approach it seemed poised to continue this time around. But then in July, Trump announced the deal to allow Nvidia to sell its H20.
The move drew swift rebuke. Democratic Senators Mark Warner and Chuck Schumer cosigned a letter to Trump on August 15 saying that the Nvidia deal “blatantly violates the purpose of export control laws.” Three days later, John Moolenaar, Republican chairman of the House Select Committee on China, wrote to Commerce Secretary Howard Lutnick: “We can’t let the CCP use American chips to train AI models that will power its military, censor its people, and undercut American innovation.”
The war in Ukraine, where attack drones stalk their targets using image-recognition algorithms, has only heightened the fear that China will use advanced chips to bolster its military. “Supercomputers with the latest AI chips can and will be used to develop and improve weapons systems, and AI-enhanced weapons will be a massive game changer on the battlefield,” Nystedt said.
Military applications for computing power go beyond AI. In 1997, following reports that China had used U.S. supercomputers to design a next-generation fighter jet, Congress passed a provision of the National Defense Authorization Act barring advanced supercomputer exports to China. That was back when chatbots were still the stuff of science fiction movies. But similar fears persist today.
Pilz, the researcher at Rand, said that because Nvidia’s chips are general purpose, they can be used for both military and civilian applications. “Given China’s military-industrial cooperation,” he said, “accelerating civilian research also often indirectly helps military research.”
Computing power is especially useful in the development of weapons like missiles and drones, which are designed using advanced computational methods, said Iain Boyd, director of the Center for National Security Initiatives at the University of Colorado.
The faster the computer, the more efficient the weapons development. “The end result is a more accurate design, achieved in less time and at a reduced cost,” Boyd said.
After leaving Taipei on Friday, Huang flew back to Washington, where he mingled on Monday with South Korean President Lee Jae Myung and other business leaders at a Korea-U.S. business roundtable. On Wednesday, Huang will address investors on Nvidia’s quarterly earnings call. Last quarter, the company said it expected to take an $8 billion dollar hit from restrictions on access to the Chinese market.
U.S. export controls are one barrier to Nvidia’s interest in China. The Chinese government is another. In June, one of Lutnick’s deputies told Congress that the chip production capacity of Huawei, the only Chinese firm capable of competing with Nvidia, “will be at or below 200,000” for 2025. China has encouraged its tech firms to buy up Huawei’s chips in an effort to boost that number, which lags significantly behind the U.S. Analysts say that China’s industrial policy is as much a consideration as national security when discouraging procurement of the H20.
Even so, Pilz says that China isn’t yet in a position to turn down all of Nvidia’s high-end chips. “I expect the CCP will try to nudge/force companies to rely on inferior Chinese chips,” he said. “This is likely a strategic error. Chinese companies know of the value of the B30A, so they will push back.”
Lutnick, for his part, has argued that letting Nvidia sell its chips to China will keep the country reliant on U.S. tech. “You want to sell the Chinese enough that their developers get addicted to the American technology stack,” Lutnick said. “That’s the thinking.” Reports later emerged that China found Lutnick’s remark especially insulting — memory of China’s addiction to British opium in the 19th century may explain why.
But Janet Egan says that the country that should be most worried about the B30A is the U.S., not China. “Once you export a physical chip,” she said, “it’s gone forever.”
Leave a Reply