NVIDIA is reportedly working on a new AI chip for China, built on its latest Blackwell architecture, with performance expected to surpass the currently approved H20 model.

The chip, temporarily named B30A, is said to use a single-die design, meaning all major circuit components are integrated onto one silicon piece rather than split across multiple chips. While its raw computing power may reach only about half of NVIDIA's flagship B300 dual-die accelerator, it will still feature high-bandwidth memory (HBM) and NVLink technology for fast data transfer—similar to the H20.

Although U.S. President Donald Trump recently signaled some flexibility on chip exports to China, regulatory approval remains uncertain. Washington is cautious about giving China wider access to advanced U.S. AI hardware.

Insiders say NVIDIA aims to deliver B30A samples to Chinese clients as early as September for testing. The specifications are still being finalized.

In addition, NVIDIA is preparing another Blackwell-based chip, designed mainly for AI inference tasks. Called RTX 6000D, it targets the Chinese market with a lower cost than the H20, thanks to simpler specifications and a GDDR-based memory system. Its bandwidth of 1.398TB/s is just under the U.S. government's export limit of 1.4TB/s. Limited shipments of the RTX 6000D could also start in September.

NVIDIA emphasized in a statement that all its products remain fully approved by regulators and are intended strictly for commercial use.

The company only regained permission to sell the H20 chip in July, after being forced to halt sales in April 2025 despite the part being tailored for China since 2023 export rules. With Washington now requiring NVIDIA and AMD to give up 15% of their China-related chip revenues, NVIDIA is under pressure to balance compliance with keeping its foothold in the Chinese AI market—where maintaining ecosystem compatibility is crucial to prevent developers from switching to domestic alternatives.