When stock in Nvidia crashed an unprecented 18% allegedly because DeepSeek allegedly having used less capable chips, many were confused. Some pointed at the Jevons paradox which describes that efficiency gains cause an increase in use to a degree where the gains evaporate. But as X user Alexander Doria was quick to point out:

DeepSeek has trained on Nvidia H800 but is running inference on the new home Chinese chips made by Huawei, the 910C.

He attached a screenshot of Huawei offering “DeepSeek-R1-Distill”, presumably fine-tuned, variants of Qwen and Llama. I didn’t make the connection back then, but users @teortaxesTex and @olalatech now provide additional background. In what Yuchen Jin, CEO & Co-Founder at Hyperbolic Labs, portraits as unsubstantiated rumors from Zhihu/RedNote, Huawei is claimed to position their offerings as fit for “reasoning inference”. The article mentions “compressed inference” (压缩推理) as a key enabling technology for Huawei chips, which ties back to Alexander’s R1-Distill screenshot. So the market crash could be due to a confusion of actual R1 with R1-Distill, with the wrongful conclusion drawn that Huawei could also deliver on frontier model demands, while in fact it can only deliver on the very low-end.

However, CUDA may not be as much of of a moat to Nvidia than I previously thought: from a report by Unite.ai:

[Huawei Ascend 910C] software compatibility, including support for Huawei’s MindSpore AI framework and other platforms like TensorFlow and PyTorch, makes it easier for developers to integrate into existing ecosystems without significant reconfiguration.

[Update 2025-02-02] IT News outlet Slashdot offers a different take, sharing a story at Marketwatch.com about “The blogger who helped spark Nvidia’s $600 billion stock collapse and a panic in Silicon Valley”. This references a blog post by Jeffrey Emanuel in which the author shares basically nothing new, and the 12.000 words boil down to disingenuity likes this:

when DeepSeek can match GPT-4 level performance while charging 95% less for API calls, it suggests either NVIDIA’s customers are burning cash unnecessarily or margins must come down dramatically.