Bypassing export regulations, Nvidia allegedly devising a reduced B300 model specifically for the Chinese market
Nvidia is set to introduce a new GPU, the B30A, specifically designed for the Chinese market. This move comes in response to U.S. export restrictions and increasing pressure on cloud providers and model developers not to employ Nvidia kit for sensitive government workloads in China.
The B30A is based on the Blackwell Ultra microarchitecture, featuring a single compute chiplet and offering 144 GB of HBM3E memory. It is estimated to have around 700W or less power consumption, significantly lower than the B300's 1400W.
Key specifications of the B30A include:
- Memory: 144 GB HBM3E
- Memory Bandwidth: Around 4 TB/s
- Compute: Single chiplet with approximately half the computational capacity of the dual-chiplet B300
- Architecture: Blackwell Ultra microarchitecture with 5th Gen Tensor Cores and NVFP4 support at boosted rates
- Power: Estimated TDP around 700W or less
The B30A is expected to be used in AI inference workloads and data center applications tailored for the Chinese market. It aims to serve high-performance computing and AI acceleration needs in Chinese data centers and cloud environments where Nvidia's traditional top-tier chips cannot be sold due to U.S. export controls.
If the report is accurate, the B30A would deliver approximately 7 petaFLOPs of dense FP4 performance, making it a viable replacement for the aging H200 NVL due to its power consumption and form factor. It may also be used in scaled systems that integrate NVLink for multi-GPU configurations but likely with some limitations on cluster-scale connectivity to maintain compliance.
The B30A's introduction reflects Nvidia's strategic attempt to maintain a foothold in China’s growing AI market under increasing U.S. export restrictions. Its single-die design and performance tier represent a compromise balancing U.S. restrictions with the need to compete technologically. Nvidia’s continued revenue from the Chinese AI market relies on such compliant designs, as previous products like H20 generated tens of billions in revenue under Trump-era revenue-sharing agreements but face increased uncertainty under current geopolitical tensions.
Despite the innovation in B30A, Nvidia’s market share in China is expected to decline from 66% to about 54% by 2025 due to intensifying competition from domestic Chinese alternatives and broader trade frictions. However, the B30A could potentially be sold to markets other than China, as the most powerful Blackwell parts with high-bandwidth memory aren't offered in a PCIe form factor.
The B30A's potential release comes amid ongoing pressure from lawmakers, including the Chair of the House Select Committee on China, about the potential sale of H20 accelerators to China. If the B30A exceeds the performance caps set by the US export restrictions on AI accelerators, it would require approval from the US Commerce Department and a 15 percent revenue share for the US government. President Trump suggested he was open to the idea of allowing sales of Blackwell-based chips to China if the performance were "enhanced in a negative way" by 30-50 percent.
In summary, the Nvidia B30A is a Blackwell Ultra GPU tailored for China’s AI market, offering high memory capacity and strong AI performance while obeying U.S. export restrictions. It plays a crucial role in Nvidia’s tactical response to geopolitical challenges posed by tightening US-China trade controls on advanced AI semiconductors.
[1] TechCrunch: Nvidia is reportedly preparing a new GPU, codenamed B30A, for the Chinese market [2] AnandTech: Nvidia's B30A: A New GPU for the Chinese Market [3] Tom's Hardware: Nvidia's B30A GPU: A New Chip for the Chinese Market [4] Wired: Nvidia's New GPU Is a Hedge Against US-China Tensions [5] The Register: Nvidia's B30A GPU: A Compliant Solution for China’s AI Market
