Particle.news

Download on the App Store

Nvidia Readies Blackwell B30A for China as DeepSeek Shifts Off Huawei for Training

A new licensing framework and domestic chip quotas are reshaping how Chinese AI developers choose and deploy accelerators.

Overview

  • Reuters reports Nvidia is developing a China-focused Blackwell part, tentatively called B30A, using a single-die design with roughly half the B300’s compute and more performance than H20, with samples targeted as early as next month pending U.S. approval.
  • Sources also point to an RTX6000D card tuned for inference with 1,398 GB/s bandwidth, with small-batch deliveries to Chinese clients expected in September under existing export thresholds.
  • DeepSeek delayed its R2 model after repeated training failures on Huawei Ascend chips despite on-site support, then moved training to Nvidia hardware and kept Ascend for inference.
  • Chinese authorities have reportedly required publicly funded data centers to use more than 50% domestic chips, creating integration hurdles as teams translate CUDA-based stacks to Huawei’s CANN software.
  • Nvidia and AMD agreed to remit about 15% of China chip revenue to the U.S. in exchange for export licenses, and President Trump has signaled openness to scaled-down Blackwell sales while Chinese regulators scrutinize H20 purchases for security risks.