Epoch AI predicts a significant shift in the AI industry by 2030, with inference compute power expected to outpace model training compute. The nonprofit organization projects that nearly half of all inference compute will transition to Application-Specific Integrated Circuits (ASICs), moving away from the general-purpose GPUs currently prevalent. This shift highlights a growing focus on deployment economics over development. Epoch AI estimates that by 2030, the total AI power capacity in the US could exceed 50 gigawatts, with global capacity surpassing 100 gigawatts. The organization notes that while training compute for frontier AI models is growing rapidly, inference compute will dominate, driven by the efficiency of ASICs like Google's Tensor Processing Units and Amazon's Inferentia chips. This trend suggests a substantial buildout in data centers and infrastructure to support the increased power demand.