In April 2025, the US government imposed additional restrictions on exports of Nvidia’s H20 chips to China.
Nvidia is planning to launch a new AI chipset for the Chinese market, designed to comply with US export rules. According to a report from Reuters, the GPU will form part of Nvidia’s Blackwell family of chips, with mass production expected as early as June.
A second chip, also based on Blackwell architecture and designed for the Chinese market, is likely to begin production in September, Reuters further reported.
Citing sources “familiar with the matter,” Reuters reported that the first chip set for production will be based on Nvidia’s RTX Pro 6000D offering that uses GDDR7 memory instead of HBM. It will be priced between $6,500 and $8,000 – lower than the cost of the company’s now-restricted H20 chip – reflecting its diminished specifications and manufacturing requirements.
No further details about the second chip are currently known.
First announced in 2023, Nvidia’s H20 GPUs are a less sophisticated version of its H100 processors and were designed specifically for the Chinese market in compliance with US export controls.
However, in April 2025, the US government imposed additional restrictions on exports of Nvidia’s H20 chips to China, a move that Nvidia said would cost it approximately $5.5 billion in charges associated with H20 products for inventory, purchase commitments, and related reserves. The company has since said it also had to walk away from $15bn in sales as a result of the new restrictions.
Speaking to reporters during the Computex event in Taiwan earlier this month, Nvidia CEO Jensen Huang told reporters that if the US government continues to curb exports of semiconductors to China, it will bolster the sale of Huawei’s chips.
Details of Huawei’s incoming AI chips were first reported last month. The company’s Ascend 920 AI chips – which were expected to challenge Nvidia’s H20 GPUs – are produced using 6nm process node technology and expected to offer more than 900 teraflops (BF16 performance) per card, whilst boasting 4Tbps of memory bandwidth.
