Despite OpenAI's GPT-5.6 being 54% more token-efficient, models like GPT-5.6 and Fable 5 consume considerably more tokens for complex agentic tasks, reinvesting efficiency into deeper reasoning. This increased token usage is driving total compute consumption per task higher, requiring more powerful and efficient computing infrastructure. The US power grid and datacenter capital expenditure, exceeding $800 billion this year, are identified as major bottlenecks, with China holding a significant energy advantage.
Frontier model efficiency gains are being reinvested into deeper reasoning, driving total compute and energy consumption per task higher, exacerbating infrastructure bottlenecks.