SemiAnalysis founder: AI inference market may surpass oil to become one of the world's largest markets.

According to BlockBeats, on July 1st, Dylan Patel, founder of SemiAnalysis, stated in a recent interview with Sequoia Capital's podcast "Training Data" that AI inference will become one of the world's largest markets, potentially exceeding the size of oil and accounting for several percentage points of global GDP. He believes that with each iteration and upgrade of the model, the number and value of tasks that can be completed expand faster than the growth of computing power, thus the computing power shortage may persist for a long time. Patel predicts that by 2030, the combined computing power demand of just OpenAI and Anthropic alone will exceed 100 gigawatts. While the impact of space data centers will remain negligible for the next 3 to 5 years, by 2040, more than half of the new computing power may go into space. He stated that the core constraint lies in the cost of terrestrial energy and the ability to build power grids; once space deployment becomes more economical than terrestrial deployment, the migration of computing power to space will become inevitable. Regarding hardware and software co-design, Patel stated that the efficiency improvements in AI over the past three years have 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 TPUs; Anthropic models are more suitable for TPUs, while OpenAI models are more GPU-oriented. He believes that the so-called CUDA moat is not essentially just CUDA itself, but rather the open-source model ecosystem's widespread collaborative optimization around GPUs. Patel also stated that NVIDIA CEO Jensen Huang's strong support for emerging cloud computing providers is to prevent hyperscale cloud vendors from monopolizing the computing power landscape and to promote a multi-polar market. Furthermore, the InferenceX real-time inference benchmark system built by the SemiAnalysis team shows that, at equivalent quality, inference costs decrease by approximately 60 times annually, and intelligence per watt improves by approximately 40 times.