SemiAnalysis Founder: AI Inference Market Size Could Surpass Oil, Becoming One of the World's Largest Markets

BlockBeats News, July 1st. Recently, Dylan Patel, the founder of SemiAnalysis, stated in an interview on the Sequoia Capital podcast "Training Data" that AI inference will become one of the world's largest markets, potentially surpassing oil and accounting for several percentage points of global GDP. He believes that after each model iteration upgrade, the speed of task completion and value expansion continues to outpace the growth of computing power, indicating that a computing power shortage may persist in the long term.

Patel predicted that by 2030, the combined computing power requirements of only two companies, OpenAI and Anthropic, will exceed 100 gigawatts. In the next 3 to 5 years, the impact of space data centers will still be negligible, but by 2040, more than half of the new computing power may be deployed in space. He stated that the core constraint lies in ground energy costs and power plant capacity, and once space deployment becomes more economical than ground-based solutions, the migration of computing power to space will become inevitable.

In terms of hardware and software co-design, Patel mentioned that over the past three years, the efficiency improvement in AI did not primarily come from hardware but from model-level and cross-layer collaborative optimization. Using DeepSeek as an example, he mentioned that its expert model shape was specifically optimized for the NVIDIA Hopper architecture, performing well on Hopper but poorly on TPU. Anthropic's models are more suitable for TPU, while OpenAI's models lean towards the GPU route. He believed that the so-called CUDA moat is essentially not just about CUDA itself but about the widespread open-source model ecosystem that commonly focuses on GPU collaborative optimization.

Patel also stated that NVIDIA CEO Jensen Huang strongly supports emerging cloud computing businesses to prevent a monopolistic computing power landscape by large-scale cloud companies and to promote a more diversified market. Furthermore, the real-time inference benchmark system, InferenceX, developed by the SemiAnalysis team, shows that under equivalent quality, the cost of inference decreases by about 60 times per year, with an approximately 40 times improvement in intelligence per watt.