On Tuesday, Nvidia said that the U.S. decision that exports of its H20 artificial intelligence chip to China would be restricted will cost it $5.5 billion in charges.
U.S. export limits have concentrated on Nvidia’s AI processors, with officials attempting to prevent the sale of the most advanced chips to China as the US competes with China in the AI area. Following the implementation of such restrictions, Nvidia began creating tiles that were as near to US borders as feasible.
After hours, Nvidia shares were down almost 6%.
Nvidia wants to be involved in the rapidly growing Chinese AI market, and the H20 is now the company’s most potent product to sell there. However, according to a Reuters story last month, Chinese firms including Tencent, Alibaba, and TikTok (owner of ByteDance) had been placing large orders for H20 chips for their DeepSeek AI venture because of the country’s high need for affordable AI models.
When AI models are performing what is known as inference—using models to answer questions—the H20 chip performs similarly, if not better, than Nvidia’s chips that are sold outside of China. However, it is not as fast as those chips for training AI models. The largest segment of the AI chip market is rapidly evolving into inference. Jensen Huang, the CEO of Nvidia, stated last month that the company is in a strong position to control that change.
But Nvidia said Tuesday the U.S. government is blocking the Chinese from buying H20 on the fear that the chips could be used in a supercomputer. The H20 has more cooling power than many Nvidia chips and a high speed, but still can mate with chips and other memory as fast as any other Nvidia chip.
Since 2022, the United States has limited the sale of processors that can be used in the supercomputers it delivers to China, and the memory and networking features of the H20 may actually be helpful in the construction of China’s supercomputers. According to the Institute for Progress, a nonpartisan think organization based in Washington, D.C., the systems are already in Chinese hands and need to be prohibited from doing so.
And at least one of these buyers has already used H20s in at least one place where big model training takes place, probably in breach of the rules that do not allow supercomputers above a certain threshold to control chips. A group of researchers added that the same restrictions are likely to apply to their supercomputer used to train their V3 model.
Nvidia said on Tuesday that the U.S. government notified it that its H20 chip needs a license in order to be exported to China, and that on April 14 it would be told that rules for those licenses would be in place indefinitely.
However, it’s not certain how many (if any) of those licenses would be granted by the U.S. government.
The filing is the extent of what Nvidia would say beyond that. There was no comment immediately available from the U.S. Department of Commerce, which controls U.S. export controls.
By charging $5.5 billion according to H20 products for inventory, purchase commitments and related reserves, the company Nvidia has added it.
It follows rumors that Nvidia and its partners, including TSMC, are contemplating building servers in the US and potentially bringing in as much as $500 billion of revenue.
