Unlike standard AI chat, autonomous agents like OpenClaw loop through reasoning, error handling, retries, consuming 100x more tokens. To run these agents, the cloud needs more H200/B200 GPUs, while local setups drive demand for high-performance cards like RTX 5090. NVDA remains the sole “arms supplier” in this war.