DeepSeek, a Chinese startup known for its innovative approach to artificial intelligence, has dramatically increased AI computing requirements by 100 to 1,000 times compared to Nvidia's standards. The company recently revealed that it uses 2,048 Nvidia chips to train its V3 model, employing a hardware-software co-design strategy as it prepares for its upcoming V4 and R2 models. This development follows U.S. President Donald Trump's announcement of a $500 billion public-private investment in America's AI future. Meanwhile, Nvidia has advanced its own technology with the release of the B200, a desktop device equipped with four RTX 4090 GPUs capable of supporting up to 70 billion parameter unquantized models. Independent benchmarking shows that DeepSeek's R1 model achieves over 1,000 output tokens per second, which is more than ten times faster than some competitors, marking it as the fastest R1 endpoint tested so far. Additionally, Shanghai Goku Technologies, a Chinese quantitative fund, has published a paper on an AI training breakthrough, signaling increased activity and innovation in China's AI sector.