The most explosive news yesterday was that former OpenAI researcher @willdepue discovered that the latest six open-source models from DeepSeek, Kimi, MiniMax, and Ant Financial all used NVIDIA Blackwell GPUs for training. This GPU is explicitly banned from sale to China by the US, yet Chinese models are now using these banned GPUs for training, indicating that the "chokehold" isn't entirely effective. The report estimates that five training sessions totaled approximately 12,000 GPUs. At $70,000 per GPU, the GPUs alone amount to $840 million. How he discovered this: 1. Before model training, billions of cells are filled with random numbers, the order of which is tied to the number of computing cores on the GPU. 2. After training, some cells remain untouched, retaining the original random numbers. 3. By working backward from these remnants, the H100 generation read 132, while Blackwell reads 148 or 152. Earlier models from the same companies still read the H100 generation. The GPU replacement is concentrated between April and October 2026, which coincides with the time when Chinese open-source models begin to catch up with top-tier US models. According to the report's estimated training using NVIDIA Blackwell GPUs: - Kimi K3 approximately 4,500 frames - DeepSeek-V4-Pro approximately 3,000 frames - Approximately 1,500 frames each for the other three models These figures are estimates and do not represent actual data, nor do they indicate that all training uses NVIDIA GPUs. However, for large model companies, time is money; halving the training time doubles the number of GPUs. Kimi K3's lead over other Chinese models in terms of training time and parameter count may be due to the use of more memory cards. Finally, it's important to note that fingerprints only prove which memory card was used for training; they don't indicate the card's location or ownership. It's possible the card was rented overseas. Previous reports indicated that Kimi used overseas computing power for training, which wasn't illegal at the time, as the US doesn't explicitly prohibit Chinese companies from renting overseas computing power for model training. The report itself states that this cannot be used to determine smuggling. DeepSeek's latest V4.1-Flash didn't show fingerprints, and the Qwen series wasn't specifically mentioned, reflecting its true capabilities. The ban controlled the sale of memory cards, but not their rental. The US may close this loophole in the future.