SemiAnalysis founder Dylan Patel stated that AI inference will become one of the largest markets in the world, surpassing oil in scale and accounting for several percentage points of global GDP. Patel expects that by 2030, the combined computing power demand of OpenAI and Anthropic will exceed 100 gigawatts. He believes that the core constraints lie in the cost of ground energy and the ability to build electricity, and the migration of computing power to space will become inevitable. Patel pointed out that the efficiency improvement of AI mainly comes from model layer and cross layer collaborative optimization, and the DeepSeek expert model is specifically optimized for Nvidia Hopper architecture. He stated that Huang Renxun's support for emerging cloud computing providers is to avoid super large scale cloud vendors monopolizing computing power. InferenceX shows that inference costs decrease by about 60 times annually, with an improvement of about 40 times per watt of intelligence.
