Overview
- KAIST published its measurements on Monday, July 6, 2026, reporting that agentic AI can use far more energy per completed task than single-turn chatbots, with a top-measured case of 136.5× and 348.41 watt-hours per query on a 70-billion-parameter model.
- The team traced the gap to agent behavior: agents loop through planning, call external tools, wait for results, and invoke the language model many times, producing much higher latency and GPU idle time—KAIST measured response-time increases up to 153.7× and idle GPU fractions as high as 54.5%.
- Subsequent reporting emphasized that the 136.5× figure is the top of KAIST’s measured range and varies by agent framework, model size and task mix, and that the study’s large-scale 13.7 billion-requests → ~198.9 GW number is an illustrative what-if rather than a forecast.
- KAIST’s authors recommend co-design of models, chips, data centers and power systems to cut costs and reduce grid risk, and they released the agent implementations and benchmarks as open source to let researchers and planners test real workloads.
- The practical implication is immediate for operators and buyers: per-completed-task costing matters more than per-query pricing, so enterprises and cloud vendors should reassess compute budgets, power procurement and service pricing to avoid hidden operating and stranded-capex risks.