Nvidia is developing a China-specific artificial-intelligence accelerator, tentatively called the B30A, according to people familiar with the plans. Built on the company’s latest Blackwell architecture, the single-die chip is designed to deliver roughly half the raw computing power of the flagship dual-die B300 while outperforming the Hopper-based H20 that Washington currently allows it to ship to China. The part will incorporate high-bandwidth memory and Nvidia’s NVLink interconnect, with sample deliveries to Chinese customers targeted as early as September, the people said. The project underscores Nvidia’s efforts to preserve its dominant position in the world’s second-largest AI market while navigating tightening U.S. export controls. Although Washington in July re-authorised sales of the H20—on the condition that 15% of China revenues go to the U.S. Treasury—Beijing has since raised security-backdoor concerns about the chip, prompting state agencies to caution local firms against purchases. Sources say Nvidia has asked suppliers including Foxconn, Samsung Electronics and Amkor Technology to halt some H20 production work while it works through existing inventory and evaluates demand. Chief Executive Officer Jensen Huang confirmed during a brief stop in Taipei on 22 Aug. that the company is in talks with the U.S. Commerce Department about offering China a successor to the H20, but said any decision rests with Washington. Huang reiterated that the H20 contains no security backdoors and argued that allowing U.S. chips to remain in China’s AI stack is in America’s interest, calling shipments "not a national security concern." Huang, who met Taiwan Semiconductor Manufacturing Co. executives during the visit, also disclosed that six chips based on Nvidia’s next-generation Rubin architecture have already been taped out at TSMC. He praised the foundry’s growth prospects, saying investors who buy TSMC shares are "very smart," and indicated the company will detail further progress when it reports earnings next week.