The U.S. government escalated its long-running effort to block China’s access to cutting-edge AI hardware.
What’s new: The White House announced that future shipments of Nvidia H20s, AMD MI308s, or equivalent chips to China would require a license. Concurrently, the United States Congress launched an investigation into whether chip vendor Nvidia violated earlier export rules.
How it works: Nvidia launched the H20 in late 2023 to comply with a 2022 U.S. ban on China-bound shipments of Nvidia’s H100 and H200 processors. The H20 uses the same architecture as the H200, but it’s an order of magnitude slower with less memory and memory bandwidth.
- Nvidia estimated that the new restrictions will cost the company $5.5 billion in revenue. AMD similarly expects to lose $800 million.
- Congressional leaders opened an investigation into whether Nvidia assisted DeepSeek with developing AI models, a potential violation of U.S. trade restrictions.
- The action spurred China’s biggest chip maker to accelerate production of its own AI chips. Huawei plans to begin mass shipments of its Ascend 910C AI chip, which is purportedly equivalent to Nvidia’s H100, in May, Reuters reported. The company expects to mass produce its Ascend 920, a potential substitute for the H20, in the second half of this year, according to DigiTimes Asia.
Behind the news: The U.S. government’s many moves to restrict shipments of advanced processors to China have sought to protect the nation’s lead in AI, but they have not prevented Chinese developers from closing the gap. In 2020, the U.S. required chip makers that use U.S. technology — which includes both domestic chip designers like Nvidia and makers of advanced fabrication equipment like the Netherlands’ ASML — to seek permission before doing business with Chinese tech giant Huawei. Last December, the U.S. published sweeping limits on sales of processors that involve U.S. technology, as well as the technology itself, to Chinese businesses.
Yes, but: Export restrictions may have slowed China’s production of advanced chips, but they have also incentivized China to invest in establishing leadership in AI. In January, the Chinese AI developer DeepSeek surprised U.S. policymakers and AI leaders with the release of DeepSeek-R1, which performs comparably to OpenAI’s o1, but whose weights are freely available and trained using less computation.
Why it matters: The first wave of restrictions on sales of advanced chips to China did little harm to U.S. chipmakers, largely because demand outstripped supply. But later restrictions have had a greater impact on their sales. The new limits could cost Nvidia and AMD significant revenue and likely will degrade their competitiveness abroad and bolster China’s homegrown chip-making industry.
We’re thinking: The AI community’s international scope is one of its greatest strengths. While individual countries must attend to their national security, progress in AI benefits all nations. Even in this era of rising protectionism, we hope members of the global AI community continue to support one another and encourage the free flow of ideas.