As the United States evaluates stricter trade regulations to hinder the transfer of advanced chip technology to China, Nvidia, a prominent American chip manufacturer, is reportedly developing a variant of its new artificial intelligence chips that will adhere to these regulations.
A source revealed to Reuters that Nvidia is creating a version of its Blackwell AI chips for the Chinese market and plans to collaborate with a local distribution partner named Inspur to introduce and market the chip, which is tentatively referred to as the “B20,” in China.
The B20 is anticipated to begin shipping in the second quarter of 2025. Nvidia has chosen not to comment on the matter.
The company currently offers three chips engineered to comply with U.S. export controls, including the H20. Nvidia recently lowered prices for the H20 due to sluggish sales in order to compete against chips from Chinese rival Huawei. However, sources indicate that sales of the H20 are beginning to rebound. Nvidia is anticipated to sell over one million H20 chips in China this year, equating to approximately $12 billion, despite existing U.S. trade restrictions, as reported by the Financial Times, referencing data from SemiAnalysis. This sales projection is nearly twice that of Huawei’s expectations for its Ascend 910B chip.
On a different note, analysts from Jefferies have stated that the H20 chips may face additional risks due to potential U.S. trade measures. In their assessment ahead of the U.S. annual review of semiconductor export controls in October, they suggested that it is “highly likely” the H20 will be banned for sale to China. Such a prohibition could manifest in various forms, including a specific product ban, a reduction in the allowable computing power, or limitations on memory capacity.
Moreover, the U.S. may broaden its export controls to include chips sold to other countries in the region, such as Malaysia, Indonesia, and Thailand, or even extend these controls to Chinese companies operating abroad, although analysts believe the latter would be more challenging to implement.