The company aims to navigate U.S. restrictions with a China-adapted model that maintains competitive performance at a lower cost
NVIDIA is preparing to launch a special edition of its new Blackwell architecture, specifically designed for the Chinese market. According to industry sources, this new model could be called NVIDIA 6000D or B40, and mass production is set to begin in July of this year, with an estimated price of around $7,000 per unit.
A model with GDDR7 and no HBM: key adjustments to meet regulations
One of the main differences of this GPU compared to standard Blackwell models (like the B100 or B200) is the replacement of HBM memory with GDDR7 memory, which would allow for high performance—estimated bandwidth of up to 1.7 TB/s—while adhering to regulatory limits imposed by the U.S. on the export of high-performance hardware to China.
The new GPU will also maintain CUDA compatibility, NVIDIA’s parallel computing platform, ensuring its integration into well-established AI environments in the Asian country.
Additionally, the NVLink interconnect bus is expected to operate at an estimated speed of 550 GB/s per direction, indicating a high-performance architecture designed to scale in multi-GPU systems in data centers and AI training servers.
Up to one million units by 2025
According to leaker @Jukanlosreve on social media platform X, NVIDIA’s forecast is to achieve a distribution of up to one million units of this new GPU by the end of the year. While this is an estimate, it suggests that the company anticipates massive adoption in a booming AI market in China.
This aligns with a recent announcement by NVIDIA’s CEO, Jensen Huang, during Computex 2025, where he indicated that the H20 model—based on Hopper—was no longer viable for the Chinese market due to its inability to be modified to comply with U.S. restrictions. Thus, this new version based on Blackwell would be the definitive alternative.
A strategic bet for a strained market
Since the onset of U.S. trade sanctions on Chinese tech companies, NVIDIA has had to redesign part of its offering to avoid being sidelined in what is currently one of the largest global markets for AI chips. NVIDIA’s market share in China has noticeably declined—from 90% to nearly 50% in just one year, according to Huang himself—and this new model represents a direct response to that situation.
Although more technical details about the compute units, tensor cores, or total number of transistors have not been revealed, it seems that the 6000D/B40 will be optimized for large language model (LLM) tasks, large-scale inference, and training tasks that do not require the extreme performance of the B200 but need more power than current mid-range models.
The 6000D or B40 model is shaping up to be a turning point in NVIDIA’s strategy to remain in the competitive Chinese market without violating international regulations. With a more budget-friendly price, CUDA support, and the power of the Blackwell architecture, this GPU could become one of the company’s most significant releases in 2025.
We will have to wait for the start of production in July and the official announcements to confirm specifications, final name, and availability dates. But one thing is clear: the tech battle between the U.S. and China is still heating up, and NVIDIA is not willing to give ground.
References: Twitter X, TweakTown, and El Chapuzas informático.