Limited supply of Nvidia H20 AI GPUs due to export restrictions, prompting Nvidia to develop a new B30 model specifically for the Chinese market to comply with these regulations.
Nvidia Stays Away from H20 Production, Focuses on B30 for China Market
In a strategic move, Nvidia has decided not to resume production of its H20 HGX GPUs, despite a recent policy shift allowing exports to China. The tech giant plans to fulfill only existing orders with the current inventory and has no immediate plans to restart H20 production.
This decision comes after the impact of previous U.S. government export restrictions that forced Nvidia to halt H20 silicon production earlier this year, leading to significant inventory write-downs and unused production capacity at TSMC. Although the U.S. administration recently relaxed export rules, permitting the sale of H20 GPUs to China without a license, Nvidia has not signaled a resumption of H20 production, and TSMC has not confirmed new H20 orders.
Instead, Nvidia is reportedly preparing a new AI accelerator model, the B30 GPU, based on its latest Blackwell architecture, specifically for the Chinese market. This strategic pivot suggests Nvidia's intention to develop a more advanced GPU to comply with current export restrictions and technological requirements, potentially replacing the H20 in China.
The B30 model, if confirmed, could enable Nvidia to maintain a presence in China’s AI hardware market under the new export regime while moving forward with newer technology rather than reviving the older H20 product. The B30 GPU is said to be smaller than the H20 GPU while still meeting U.S. export control requirements.
It's important to note that the information about the B30 GPU should be taken with a grain of salt until more details are available. The price of the B30 GPU is reported to be about 30% to 40% lower than the H20 GPU, which could help Nvidia maintain or even increase its gross margins for China-bound AI GPUs.
Nvidia needs to offer competitive performance compared to AI accelerators from companies like Biren or Huawei while meeting U.S. export restrictions for high-performance GPUs. However, the B30 GPU is rumoured to be 10% to 20% slower than the H20 GPU.
The successor to the H20 GPU may be based on the original Blackwell architecture or the Blackwell Ultra architecture with enhancements for FP4 performance. The cycle time for N4-based chips at TSMC is around three months, but even if production of H20 GPUs resumed today, end users would only get them in late October or early November.
TSMC, the manufacturer of H20 GPUs, did not confirm new orders for H20 GPUs last week. The fabs are operating at full utilization and have reassigned the H20 capacity to other customers. Nvidia has not confirmed anything about these reported Blackwell-based GPUs for the Chinese market.
In conclusion, Nvidia's decision to focus on fulfilling existing orders of H20 GPUs and developing the B30 GPU with Blackwell architecture for China market compliance and future sales reflects the company's balancing act between regulatory uncertainty, inventory management, and technological evolution in a complex geopolitical environment.
The tech giant, Nvidia, is shifting its focus from the H20 HGX GPUs and is instead reportedly developing a new AI accelerator model, the B30 GPU, based on its latest Blackwell architecture, specifically for the Chinese market. This strategic pivot in the business industry indicates a move towards technology that complies with current export restrictions and technological requirements.
The B30 model, if confirmed, could enable Nvidia to maintain a presence in China’s AI hardware market under the new export regime while also meeting the challenges in finance, as the price is reported to be about 30% to 40% lower than the H20 GPU, potentially maintaining or increasing gross margins for China-bound AI GPUs.
