🧠 Food for thought

Implications, context, and why it matters.

Blackwell chips stay restricted in China, so Trump’s comments have little practical impact

• Nvidia’s Blackwell series (B100/B200/B300, its latest AI accelerators) stays under US export controls with no China-ready versions approved for sale 1
• The US has granted limited, case-by-case licenses to ship the lower-spec H20 to China 1. Chinese regulators then barred buyers from the H20 and the RTX Pro 6000D, while a hypothetical B30A at roughly half a B300 could deliver about 12 to 17x the H20’s compute if licensed 1
• Trump’s talk of “leaving it to Beijing” holds little practical sway, since any shift needs explicit US action to approve new variants or grant special licenses 2
• The US keeps its AI compute edge through volume, with about 3.67 million B300-equivalent chips expected in 2025 versus China’s 40,000 to 146,000 B300-equivalents from Huawei 1. “B300-equivalent” means performance normalized to a single Nvidia B300 1

Openings for startups building AI workload migration tools for Chinese developers

• Many Chinese firms still favor Nvidia hardware despite curbs, which frustrates teams pushed to domestic options like Huawei’s Ascend (Huawei’s line of AI accelerators) 3
• Huawei’s Ascend 910C hits about 60% of Nvidia H100’s inference speed and runs into software bugs and crashes 45. The H100 is Nvidia’s flagship data center graphics processing unit (GPU) from the prior generation 45
• A wide gap between Nvidia’s Compute Unified Device Architecture (CUDA) ecosystem and Huawei’s Compute Architecture for Neural Networks (CANN)/MindSpore creates pull for tools that smooth the switch 5
• Vendors that simplify PyTorch (an open-source AI framework) to Ascend migrations or tune models for weaker hardware can win budget from Chinese AI companies adapting to chip limits 65