Overview

• Reuters reports Nvidia is developing a Blackwell-based B30A accelerator for China with a single-die design targeting roughly half the compute of the B300 and higher performance than H20, with sample testing planned as early as September pending U.S. approval.
• Nvidia is also preparing an RTX6000D card tuned for inference with memory bandwidth capped at 1,398 GB/s to sit just under new U.S. thresholds, with small-batch deliveries to Chinese clients expected in September, according to the reporting.
• President Donald Trump has signaled openness to allowing scaled-down next‑generation GPUs for China under a deal requiring Nvidia and AMD to remit 15% of revenue from some advanced chip sales in the country, after H20 sales were halted in April and partially resumed in July.
• DeepSeek’s R2 launch remains delayed after failed training runs on Huawei’s Ascend 910C despite on‑site support from Huawei engineers, leading the company to switch model training to Nvidia H20 hardware and reserve Ascend for inference compatibility.
• Chinese authorities have told state-linked data centers to source more than half of their chips from domestic suppliers, creating integration challenges between Nvidia’s CUDA-based stacks and Huawei’s CANN ecosystem and spurring domestic rivals such as Cambricon to raise new funds.