Nvidia is preparing to launch a new artificial intelligence chip for the Chinese market at a significantly lower price than its previous H20 model, which is now restricted by US sanctions. According to sources close to the matter, mass production will begin in June, with an estimated price tag of between $6,500 and $8,000, compared to $10,000 to $12,000 for the H20.

A tailored, lower-cost Blackwell architecture

This new GPU is part of Nvidia's latest generation of AI processors, based on the Blackwell architecture. It is said to be derived from the RTX Pro 6000D, a model designed for servers, and to use more conventional GDDR7 memory instead of the more powerful but more strictly controlled HBM (High Bandwidth Memory).

Another notable simplification is that the chip will not use TSMC's advanced packaging, known as CoWoS (Chip-on-Wafer-on-Substrate), a key technology in high-end chips. These choices are intended to comply with US restrictions while maintaining a competitive product.

A strategic but constrained Chinese market

China remains a major market for Nvidia, accounting for 13% of its sales in the last fiscal year. But since the first US restrictions on GPU exports in 2022, its market share in the country has fallen from 95% to around 50%, according to CEO Jensen Huang in Taipei last week. He also warned that further restrictions would push more Chinese customers towards Huawei, its main local rival.

This new GPU, despite its reduced performance, would allow Nvidia to remain present in the Chinese market while complying with the regulatory limit on memory bandwidth, now capped at around 1.7-1.8 terabytes per second. By comparison, the H20 reached 4 TB/s.

A second model is in the works for September

At the same time, Nvidia is reportedly working on another version of the Blackwell chip for China, with production set to begin as early as September. No technical details have been released yet. The company had initially planned a downgraded version of the H20, but had to abandon the project due to the limitations of the older Hopper architecture, which was deemed unsuitable for the new US regulations.

The new chip, which could be called 6000D or B40 according to a recent note from GF Securities, would be a regulatory-compliant alternative without compromising performance. However, we will have to wait for the technical details to make a comprehensive comparison between Nvidia's new GPU and Huawei's Ascend chip.

A heavy cost for Nvidia

The ban on the H20 in April forced Nvidia to take a $5.5 billion charge for inventory write-downs. Speaking on the Stratechery podcast, Jensen Huang also revealed that the company had to forego $15 billion in sales.

Despite these setbacks, Nvidia retains a crucial competitive advantage: its CUDA software platform, which is essential for AI developers and difficult to replace, even for Chinese players such as Huawei, whose chips such as the Ascend 910B are advancing rapidly.

According to analyst Nori Chiou, Nvidia could maintain a one- to two-year lead thanks to its integrated software ecosystem, even if the hardware performance gap narrows.