Nvidia is developing a new AI chip for China that outperforms its current H20 model. Tentatively named B30A, this chip is based on Nvidia’s cutting-edge Blackwell architecture and is meant to deliver improved performance while still conforming to U.S. export restrictions
Unlike the less powerful H20, the B30A uses a single-die Blackwell design. It incorporates high-bandwidth memory and NVLink technology—similar features that helped make the H20 viable—but is expected to offer a noticeable boost in raw computing power
The chip may deliver about half the performance of Nvidia’s flagship dual-die B300 accelerator, aligning with comments from U.S. officials that the new variant would be “30 to 50% off” in power compared to unrestricted models
Nvidia is planning to send B30A sample units for testing in China as early as next month, though final regulatory approval remains pending
The launch follows a recent U.S. policy shift, where President Trump signaled support for allowing scaled-down AI chip exports to China under strict conditions. As part of the agreement, Nvidia and AMD must remit 15% of their China chip revenue to the U.S. government in exchange for resumed sales Reuters
Key Highlights
- Chip Name: B30A
- Architecture: Advanced Blackwell (single-die)
- Performance: Outperforms H20; approximately 50% of B300
- Features: High-bandwidth memory, NVLink
- Timeline: Samples expected next month
- Export Context: Part of U.S. agreement; 15% revenue share required
- Strategic Aim: Maintain Nvidia’s foothold in China (13% of global revenue)
