Nvidia Offers Alternative Chip for China to Clear U.S. Export Hurdles

The A800 replaces the A100, a chip widely used in servers and artificial-intelligence applications by China’s tech giants including Alibaba Group Holding Ltd.,

Tencent Holdings Ltd.

TCEHY 0.36%

and

Baidu Inc.

BIDU -0.21%

According to a memo Nvidia sent to its channel distributors last Thursday, the A800 has the same computational performance but a narrower interconnect bandwidth, the capacity of a chip to send and receive data from other chips, crucial for training large-scale AI models or building supercomputers.

“The A800 meets the U.S. Government’s clear test for reduced export control and cannot be programmed to exceed it,” the company said. Nvidia’s plans to offer the new chip was earlier reported by Reuters.

Nvidia’s new chip is a response to the Biden administration’s sweeping new rules restricting exports of American chip technology to China, which U.S. officials say is aimed at slowing the country’s military advance. The rules unveiled by the Commerce Department last month blocked an array of semiconductor technology from being shipped to China without a license, including advanced chip-manufacturing equipment and cutting-edge chips used for AI and supercomputing.

Those rules expanded on earlier restrictions placed on Nvidia’s China exports of the A100. In August, the company disclosed that export restrictions on its A100 chips would cost it $400 million in lost sales. About a quarter of Nvidia’s $26.9 billion in revenue in its most recent fiscal year came from China and Hong Kong, the company said.

The company is preparing for high levels of demand for the new chip, according to people familiar with the matter, with the company set to start giving quotations in China as soon as Wednesday, the people said. Shipments are expected to start in the coming weeks and potential buyers include U.S. companies such as

Dell Technologies Inc.

for products they sell in China, these people said. Dell didn’t immediately respond to a request for comment.

Where the A100 can send 600 gigabytes of data per second, the A800 can send 400 gigabytes. The change is akin to reducing the lanes of a highway from six to four, and would have the largest impact on the performance of supercomputers, which string thousands of graphics-processing chips together. Simpler tasks that require only one or a few chips, such as AI inference—running an AI model after it has been trained—will be minimally affected.

The U.S. restrictions on advanced chip exports announced last month limited the interconnect bandwidth of chips that could be exported to China without a license to less than 600 gigabytes per second, alongside other performance thresholds.

Write to Dan Strumpf at [email protected], Karen Hao at [email protected] and Raffaele Huang at [email protected]