On July 1, Dylan Patel, founder of SemiAnalysis, stated in an interview with Sequoia Capital's podcast "Training Data" that AI inference will become one of the largest markets globally, potentially surpassing oil and accounting for several percentage points of global GDP. He believes that the number and value of tasks completed after each model iteration continue to expand at a pace faster than the growth of computing power, suggesting that a shortage of computing power may persist in the long term. Patel predicts that by 2030, the combined computing power demand of just OpenAI and Anthropic will exceed 100 gigawatts; in the next 3 to 5 years, the impact of space data centers will remain negligible, but by 2040, over half of the new computing power may be deployed in space. He indicated that the core constraint lies in the cost of ground energy and the ability to generate electricity. Once the economics of space deployment surpass those of ground-based systems, the migration of computing power to space will become inevitable. Regarding the collaborative design of hardware and software, Patel noted that the increase in AI efficiency over the past three years has not primarily come from hardware, but rather from model-level and cross-layer collaborative optimization. He cited DeepSeek as an example, stating that its expert model shape is specifically optimized for NVIDIA's Hopper architecture, resulting in excellent performance on Hopper but poor performance on TPU; Anthropic's model is more suited for TPU, while OpenAI's model leans towards the GPU route. He believes that the so-called CUDA moat is not just about CUDA itself, but rather the open-source model ecosystem that generally optimizes around GPU collaboration. Patel also mentioned that NVIDIA CEO Jensen Huang's strong support for emerging cloud computing companies is aimed at preventing large-scale cloud providers from monopolizing the computing power landscape and promoting a multipolar market. Additionally, the real-time inference benchmarking system InferenceX built by the SemiAnalysis team shows that under equivalent quality, inference costs have decreased by approximately 60 times annually, and intelligence per watt has improved by about 40 times.
SemiAnalysis 創辦人：AI 推論市場可能超越石油，成為全球最大的市場之一