■A Small Preface (Research Log)
The three-day weekend in mid-July was packed — elections, building a research system, and school events for the kids. Quite fulfilling.
Iorin-san (@Ioriiii8) told me that “it’s handy to save papers in Markdown,” so I had Claude Code build a local paper analysis tool with GitHub integration. Being able to whip up tools tailored to your personal needs on the spot is genuinely delightful.
■The Generative Energy Arena
The Generative Energy Arena (GEA) (https://arxiv.org/abs/2507.13302) studied how explicitly showing energy consumption in LLM user evaluations changes behavior. After analyzing 694 voting logs from MOOC participants, it was reported that an average of 46% of respondents switched their votes to more energy-efficient models after being shown energy information.
■Verifying the Break-Even Time in the DeepRx Paper
Regarding 6G, ITU-R is set to call for candidate technologies in 2027, with a commercialization target around 2030 — and 6G is aiming to incorporate Native AI as a standard feature.
While I happened to be summarizing related papers, I found an error in Energy Efficiency in AI for 5G and Beyond: A DeepRx Case Study (https://arxiv.org/abs/2507.10409).
The paper states that “if inference continues for 10 minutes, the energy spent on training is surpassed.”
The assumptions presented in the paper (values taken from the text and figures) are as follows:
| Parameter | Symbol | Value |
|---|---|---|
| Energy per training step | Estep | 20 J |
| Number of training steps | Nstep | 2.5×105 |
| Energy per inference | Einf | 0.002 J |
| Inference rate | r | 400 inferences s−1 |
Etrain = Estep × Nstep = 5.0×106 J
tBE = Etrain / (Einf × r) = 6.25×106 s ≈ 72.3 days
In other words, the correct figure is not “10 minutes” but “approximately 72 days” — the value in the paper is underestimated by roughly 10,400 times.
■Energy Savings Evaluation Scope in the 6G AI-Native RAN Paper
Towards AI-Native RAN: An Operator’s Perspective of 6G Day 1 Standardization (https://arxiv.org/abs/2507.08403), published by China Mobile, reports a large-scale trial across 31 cities and 5,000 base stations:
“AI achieved a 34.16% energy reduction.”
The network power consumption model presented in the paper is:
E = NBS × Top × (PPA + PTRX + PDIF + PBB + Pstatic)
The breakdown covers only telecom hardware — the power for AI inference servers and cooling facilities is not included as a parameter (as interpreted from the explanation in Section V-E). Therefore, the “34% reduction” represents energy savings in the communications portion only, and the overall system energy balance accounting for the power that AI processing itself adds is not presented.
■Discussion and Concerns
6G standardization is scheduled for ITU’s call for proposals in 2027. If partial optimization data covering only the communications portion gets adopted first, there is a risk that total energy consumption actually increases. We need a framework that mandates disclosure of total energy balance and enables third-party verification. I’m concerned that specification development may be proceeding with flawed energy assessments.