Nvidia CEO Jensen Huang visits Beijing following US chip export restrictions – News

Nvidia Corp. Chief Executive Officer Jensen Huang visited Beijing today to meet with Chinese officials and the chipmaker’s local customers.

According to the Financial Times, DeepSeek founder Liang Wenfeng was among the clients who participated in the discussions. DeepSeek rose to prominence earlier this year after it released a reasoning model called R1. The algorithm, which is available for free under an open-source license, has demonstrated similar performance as the most advanced paid reasoning models.

During the visit, Huang and Nvidia clients reportedly discussed new chip designs. The U.S. government has implemented export controls that ban the sale of the most advanced artificial intelligence accelerators to China. The new chip designs that Huang discussed reportedly wouldn’t be subject to those restrictions.

The reported designs represent the third time Nvidia has adjusted its processor portfolio to comply with U.S. export restrictions on China.

In 2022, the Biden administration rolled out rules that banned the sale of Nvidia’s then-flagship H100 graphics card to Chinese clients. In response, the chipmaker designed a scale-down version of the chip called the H800 that wasn’t affected by the restrictions. According to Reuters, the main design change was that Nvidia cut “the chip-to-chip data transfer rate to about half.”

Graphics cards are usually deployed in clusters. To coordinate their calculations, the graphics cards in a cluster must regularly exchange data with one another. Reducing chip-to-chip data transfer speeds, the change that Nvidia reportedly implemented in the H800, limits processors’ ability to coordinate their work and thereby slows cluster performance.

In 2023, the Biden administration extended the export restrictions to the previously compliant H800. Nvidia responded by designing a new, even slower chip called the H20. It reportedly provides less than 4% of the performance offered by the Blackwell B200, Nvidia’s flagship graphics card.

Shortly before Huang’s visit to Beijing today, the company disclosed that the U.S. government had restricted its ability to sell H20 chips to Chinese customers. Nvidia stated on Tuesday that it expects to take a $5.5 billion charge as a result of the regulatory change.

The following day, a House committee sent the company a letter seeking information about its dealings with DeepSeek. Lawmakers are investigating whether Nvidia supplied the AI developer with chips that are subject to export restrictions. In February, Reuters reported that the U.S. Commerce Department is also looking into the matter.

DeepSeek’s flagship R1 reasoning model is an upgraded version of a neural network called DeepSeek-V1 that debuted in January. According to the company, DeepSeek-V1 was trained using H800 graphics cards, the scaled-down chips that Nvidia designed in 2022 to comply with U.S. export restrictions. The development project used a cluster equipped with about 2,000 of the chips.

Photo: Nvidia