BlockBeats news: On July 1, Dylan Patel, founder of SemiAnalysis, said in an 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 the oil industry and accounting for several percentage points of global GDP. He believes that after each model iteration, the number and value of tasks it can accomplish continue to grow faster than compute capacity, meaning compute shortages may persist long-term.
Patel estimates that by 2030, the combined compute demand from just OpenAI and Anthropic will exceed 100 gigawatts; while the impact of space-based data centers will remain negligible over the next three to five years, more than half of new compute capacity could be deployed in space by 2040. He notes that the primary constraint lies in the cost of terrestrial energy and power infrastructure—once space-based deployment becomes more economical than ground-based options, the migration of compute capacity to space will become inevitable.
Regarding hardware-software co-design, Patel stated that the AI efficiency gains over the past three years have not primarily come from hardware, but rather from model-level and cross-layer co-optimization. He cited DeepSeek as an example, noting that its expert model architecture is specifically optimized for NVIDIA’s Hopper architecture, resulting in excellent performance on Hopper but poor performance on TPUs; Anthropic’s models are better suited for TPUs, while OpenAI’s models lean more toward the GPU path. He argues that what is often called the “CUDA moat” is not merely CUDA itself, but rather the fact that the open-source model ecosystem has generally been co-optimized around GPUs.
Patel also stated that NVIDIA CEO Jensen Huang is actively supporting emerging cloud providers to prevent hyperscale cloud vendors from monopolizing the computing power landscape and to promote a multi-polar market. Additionally, SemiAnalysis’s real-time inference benchmarking system, InferenceX, shows that inference costs have decreased by approximately 60 times annually, while intelligence per watt has improved by about 40 times, at equivalent quality levels.