To Heck With Export Controls! Nvidia Reportedly Plotting Cut-down B300 For Chinese Market

Nvidia is reportedly prepping a new Blackwell-based GPU for the Chinese market that'll outperform its controversial H20 accelerators.

The chip, allegedly codenamed the B30A, would be a cut-down version of the GPU giant's B300 accelerators announced at GTC in March, according to a Reuters report, citing two unnamed sources familiar with the matter.

Rather than the dual-GPU module of the B300, the B30A is said to utilize a single-die design and deliver roughly half the raw floating point performance of its full fat accelerator.

If the report is accurate, it suggests that the chip will deliver roughly 7 petaFLOPs of dense FP4 performance, 144GB of HBM3e, and 4TB/s of memory bandwidth. Normalized to dense FP8, that'd make the B30A about 7.6x faster and capable of delivering 50 percent more memory capacity than the H20. In fact, such a chip would rival Nvidia's own H100 and H200 accelerators.

The performance disparity is partly because the H20 was designed specifically to comply with US export restrictions on AI accelerators imposed by the Biden administration. The B30A would exceed all of these performance caps, which means that Nvidia would need approval from the US Commerce Department and would have to cut Uncle Sam a check for 15 percent of the revenue before it could ship the chip to Chinese customers.

Last week, 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.

However, there's no guarantee that the Trump administration will green light shipments of such a powerful chip to China. Lawmakers, including the Chair of the House Select Committee on China, Representative John Moolenaar (R-MI), are already unhappy about the Trump administration's decision to rescind a sales ban on H20 accelerators.

While Nvidia has said it's working on a new Blackwell-based part for the Chinese market, details remain thin.

"We evaluate a variety of products for our roadmap, so that we can be prepared to compete to the extent that governments allow. Everything we offer is with the full approval of the applicable authorities and designed solely for beneficial commercial use," an Nvidia spokesperson told El Reg.

Previous reports have suggested that Nvidia may offer a cut-down version of its RTX Pro 6000 server chip detailed at GTC this spring, which according to Reuters, could begin sampling to customers in China as soon as next month.

• Pot calls kettle black as China dubs US 'surveillance empire' over chip tracking
• Dodgy Huawei chips nearly sunk DeepSeek's next-gen R2 model
• Tencent doesn't care if it can buy American GPUs again – it already has all the chips it needs
• Beijing doesn't want Nvidia's H20s anywhere near sensitive government workloads

While the Reuters report claims that the B30A is bound for China, the Middle Kingdom may not be the only market for it. Unlike the Hopper generation, Nvidia's most powerful Blackwell parts — those packing high-bandwidth memory — aren't offered in a PCIe form factor.

By effectively cutting the B300 in half, the chip could be brought under the 600 watt power limit of PCIe devices, making the so-called B30A a viable replacement for the aging H200 NVL. That part remains Nvidia's most powerful PCIe-based GPU available today.

The B30A could reflect a hedging of bets should US trade policy shift. Even if Nvidia were not allowed to sell the new GPU in China, other customers might want it because of its power consumption and form factor. 

Demand for Nvidia accelerators in China may also be waning as Beijing pressures cloud providers and model devs not to employ Nvidia kit, particularly for any sensitive government workloads. ®