“The adoption of inference-time reasoning, where an AI model may ping itself 100 or 1,000 times to solve a multi-part question, is expected to dramatically increase compute demand.”