Biden’s Commerce Department is issuing new semiconductor export rules that affect Nvidia. This rule categorizes countries subject to GPU export restrictions, and has an impact on Nvidia’s market. Critics argue that the rules could stifle AI innovation. Supporters say it will keep the United States in first place.
The Biden administration’s Commerce Department on Monday announced 168 pages of new regulations for the U.S. semiconductor industry that could significantly change this year for Nvidia.
The new rules target exports of graphics processing units, a class of high-performance chips made by Nvidia, and challenger AMD. The world’s data centers are filled with GPUs, and Nvidia has so far captured an estimated 90% of that market share.
Although highly complex chips such as GPUs are primarily manufactured in Taiwan, most of the companies that design them are based in the United States, so their products fall within the purview of the Department of Commerce.
A White House statement said: “To strengthen America’s national security and economic power, it is essential that this critical technology is kept overseas and that the world’s AI runs on American rails.” In the wrong hands, advanced computing could “develop weapons of mass destruction, support powerful offensive cyber operations, support mass surveillance, and other human rights abuses.”
In response to previous export restrictions, Nvidia created a less powerful chip model specifically for the Chinese market in order to continue operating there after the Biden administration changed the rules in 2022.
The new regulations go further by grouping countries into three categories, each with different export controls.
The first is a group of 18 allies to which we can freely ship GPUs. These are Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, and the United Kingdom.
The second group is listed as “countries of concern” where exports of cutting-edge GPUs will be completely banned. These are China, Hong Kong and Macau, Russia, Iran, North Korea, Venezuela, Nicaragua, and Syria.
All other countries are subject to the 100,000 GPU limit. The rule sets out a verification process for large orders, and companies looking to establish large clusters in these countries will need US government approval to do so.
The administration said the regulations include provisions to maintain small orders of chips to research institutes and universities.
Nvidia, along with the Semiconductor Industry Association, opposes the regulation.
Ned Finkle, Nvidia’s vice president of government affairs, said: “Though masked as ‘anti-China’ measures, these rules do nothing to strengthen U.S. security.” said in a statement on its website.
Impact on Nvidia
Any restrictions on GPU sales anywhere would inevitably hurt Nvidia’s sales.
Related articles
“The Biden administration is now attempting to restrict access to mainstream computing applications with unprecedented and erroneous ‘AI proliferation’ rules, which threatens to derail innovation and economic growth around the world,” Finkle said. I wrote it.
But will this regulation reduce sales or change sales?
Chris Miller, author of “Chip Wars” and a leading semiconductor industry expert, told Business Insider that demand for NVIDIA products is so high that total GPU sales will be materially affected. He said it was unclear.
“I think these rules will generally have the effect of shifting data center construction to U.S. companies,” Miller said.
If demand does indeed decline, “the picture will change with reduced GPU demand from countries and companies that don’t want to rely on U.S. cloud providers,” Miller said.
The draft rules have been circulating ahead of Monday’s announcement, and reactions from technology industry leaders have been fierce.
Oracle EVP Ken Glueck first blogged about them in mid-December and again in early January.
Both Finkle and Glück highlighted national caps as the most important element introduced.
“An extreme ‘country cap’ policy would affect mainstream computing in countries around the world and would not promote national security, but rather push the world toward alternative technologies,” Finkle said. said in an emailed statement Friday.
Of particular note, Glueck noted that Singapore, Mexico, Malaysia, the UAE, Israel, Saudi Arabia, and India are not included in the list of unrestricted countries.
Miller said that if some countries in the Middle East are excluded, the direction of global AI infrastructure construction could change significantly.
“The main impact of these regulations is that cutting-edge AI systems are much more likely to be trained in the United States rather than the Middle East,” Miller said.
“Without these regulations, wealthy governments in the Middle East would have had some success in persuading U.S. companies to train high-end AI systems in the Middle East by offering subsidized data centers. Because that won’t be possible, American companies will be training their high-end AI systems in the Middle East, “American systems,” Miller said.
According to the Federal Register, Glück wrote in the draft regulation, which will be formally published on Wednesday, that national quotas are a terrible concept.
Oracle’s Glueck wrote in December that “GPU control makes no sense when you can simply add more GPUs to solve the problem, even at lower performance, to achieve parity.” “The problem with this proposal is that it assumes there are no other non-US suppliers available for GPU technology,” he continued.
Republican support
The fate of Biden’s unprecedented export control rules is unclear given their timing.
Nvidia’s Finkle’s Monday statement referenced the Trump administration, saying that during his first term, the Trump administration “laid the foundation for America’s current strength and success in AI.”
The new rule has a 120-day comment period before it becomes enforceable. President Biden will have left office when it takes effect.
Although this rule originated with the outgoing Democratic Party administration, it has also gained some support from the incoming president.
Republican Rep. John Moolenaar and Raja Krishnamoorthi, chairman and ranking member of the House Select Committee on the Communist Party of China, support the framework.
“Countries that host GPUs or Huawei’s cloud computing infrastructure should be restricted from accessing model weights for closed-weight dual-use AI models,” the lawmakers said in a written statement.
The editorial was written by Matt Pottinger, who served on the National Security Council during the Trump administration’s first term and currently chairs the China Program at the Foundation for Defense of Democracies along with Anthropic CEO Dario Amodei. It was published in the Wall Street Journal on January 6th. They suggest that while existing export restrictions have been successful, further restrictions are needed as they still leave room for China to set up data centers in friendly third countries.
Pottinger and Amodei write: “Skeptics of these restrictions argue that countries and companies covered by the rules will simply switch to Chinese AI chips. It overlooks superiority and gives countries an incentive to follow U.S. rules.”
“Countries that want to reap huge economic benefits will have an incentive to follow the US model rather than use China’s inferior chips,” they continued.
Miller acknowledged that the fact that China is still buying Nvidia’s “defective” GPUs is enough evidence that domestically designed chips are still not competitive.
“As long as China imports U.S.-made GPUs, it cannot export them. In that case, these regulations would be effective because there would be no alternative source of high-end GPUs,” Miller said.
But Huawei is catching up, said Forrester senior analyst Alvin Nguyen. In his view, additional U.S. export controls could accelerate that effort.
“They caught up a generation later than Nvidia,” Nguyen said.
Another concern is that restricting the distribution of advanced chips could fragment the economic opportunity for AI to spread equally around the world.
“If you’re not working with the best infrastructure and the best models, you may not be able to leverage the data you have, creating haves and have-nots,” Nguyen said.
