Nvidia's costs could double due to new US export rules for AI chips

Contents

The US government's new export restrictions on AI chips have apparently taken Nvidia and its Chinese customers by surprise. In addition, the financial impact is likely to be significantly higher than the costs of 5.5 billion US dollars initially quoted by Nvidia due to the loss of deliveries of H20 chips specially tailored for export to China. Intel and its AI chips for servers are also affected.

On Tuesday this week, new US rules for the export of AI chips to China came into force, costing Nvidia billions, as the company initially reported. Nvidia had officially put this at 5.5 billion dollars, but the Group is likely to lose significantly higher amounts of revenue in the future due to lost orders. Losses of more than 10 billion dollars are expected. After all, Nvidia generated sales of 12 to 15 billion dollars in China last year alone.

Nvidia was still optimistic after Trump meeting

This is according to the Financial Times, which also states that Nvidia and its Chinese customers feel blind sided by the new export controls. The US government had announced a possible tightening of AI chip exports at the beginning of April. However, Nvidia had assumed that the H20 chip, which fell under the previous export rules due to technical restrictions, would also be exempt from the new export restrictions.

In addition, leading Nvidia managers had met with Donald Trump at his estate in Florida this month and were subsequently convinced that they would be spared further tightening of export controls. Trump was also probably impressed by Nvidia's plans to invest 500 billion dollars within the USA.

Licensing unclear so far

Nvidia had therefore explained to its major Chinese customers such as Alibaba, ByteDance, and Tencent that their H20 orders would not be affected by potentially stricter US export controls. However, the US government is now demanding licenses for the export of such AI chips. It is not yet clear where and how customers can apply for such licenses and on what basis they will be granted.

Nvidia usually requires more than six months delivery time for such AI chips, so most of this year's orders have not yet been fulfilled and are likely to fall under the new US export rules. However, demand has picked up since the beginning of this year, as Chinese companies are ordering more H20 chips from Nvidia due to the new Chinese AI model DeepSeek.

Intel provides details on chips subject to licensing

In addition to Nvidia, Intel is also impacted by the stricter export controls, as the processor manufacturer also has an AI accelerator for data centers in its range, Gaudi 3. However, while Nvidia has not yet commented on the requirements of the new export restrictions, Intel has already informed its Chinese customers accordingly.

Accordingly, the US government requires that chips with a memory bandwidth of 1400 GByte/s, an I/O bandwidth of 1100 GByte/s and a total bandwidth of 1700 GByte/s or more must be licensed before export. Both Nvidia's H20 and Intel's Gaudi series clearly exceed these requirements, meaning that exports to affected countries such as China must be explicitly approved with immediate effect.

(fds)