100000 Blackwell trained GPT-6 Astra, where is the difference in computing power of Chinese large models? 】 Dynamic Beating AI News: NVIDIA CEO Huang Renxun revealed this morning that GPT-6 Astra training used over 100000 Grace Blackwell GPUs, forming a high-speed interconnected cluster through NVLink72. He also claimed that the next batch of 400000 GPUs will be launched, but did not specify the specific model and ownership. Using Astra as a benchmark, there is still a significant gap in the publicly available computing power of China's top model companies. ByteDance is currently the closest, having accessed approximately 36000 B200 cards through Malaysia this year. This batch of chips and the GB200 used by Astra belong to the Blackwell generation, but the computing power is deployed overseas. When ByteDance publicly introduced its domestic clusters in China this year, it mainly used the Chinese special edition H20 and H800 of the Hopper architecture. Kimi was recently exposed to have obtained approximately 20000 Hopper GPUs through Alibaba. Bloomberg sources claim it is specifically H200, but Alibaba denies the claim of H200 model. H200 belongs to the previous generation Hopper; B200 has been upgraded to Blackwell. GB200 further combines Grace CPU and Blackwell GPU, and NVL72 can put 72 GPUs into the same high-speed interconnect domain. DeepSeek has not released the complete training hardware for V4. According to a transcript of the investor exchange meeting with Liang Wenfeng, the company had approximately 20000 H-equivalent computing power cards in May, most of which had just arrived, and the next purchases will be mostly from Nvidia. Huawei has provided DeepSeek with a production capacity of approximately 16000 sheets for 950. Liang Wenfeng said that this batch of domestically produced cards is equivalent to 4000 Nvidia B-series, which is only enough for training the current generation model. And the Blackwell used by Astra is no longer Nvidia's latest generation. Vera Rubin has entered full mass production. NVIDIA estimates that the number of GPUs required for Rubin to train a large MoE model can be reduced to one fourth of Blackwell's. In public information, no Chinese company has disclosed a case of using 100000 advanced GPUs of the same generation for a single model training. So the publicly visible gap between Chinese and American AI companies is no longer just the number of cards, but also the generation of cards they can obtain and how many cards can be collected for a model at a time.