This page may contain stale information. Last updated: 2026-07-05
Definition
Agent energy consumption measures the electricity and GPU compute required for AI agents that autonomously plan, call external tools, and iterate through multi-step reasoning loops — as distinct from single-turn chatbot inference.
KAIST Quantification (HPCA 2026)
First real-world quantitative analysis under service conditions (kaist, Prof. Minsoo Yoo):
| Metric | Agent vs Chatbot |
|---|---|
| Energy per query | Up to 136.5× higher |
| Response time | Up to 153.7× longer |
| GPU idle time | Up to 54.5% of runtime (waiting on tools) |
| 70B agent average | 348.41 Wh per question |
Why Agents Cost More
- Repeated LLM invocations (Reflexion, LATS-style loops)
- Serialized tool-calling: expensive GPUs idle during external API/web waits
- Test-time scaling: accuracy traded for compute
198.9 GW projection assumes 13.7 billion daily agent requests — hypothetical scale-up scenario for infrastructure stress-testing, not a forecast.
Implications
- Agent architecture choices (tool loops, model size, batching) directly affect data center capex
- “Smarter AI” competition shifting toward “more efficient AI” per KAIST framing
- Open-source benchmark code released for reproducibility