AI agents can consume 1,000 times more tokens than a single chatbot query, forcing a rethink of chip ratios, server architecture, and power budgets