The footprint of every answer measured, cited and reproducible

Executive summary

Every query to the Fluxia assistant consumes energy (the electricity of the GPUs running the model) and, indirectly, water (datacenter cooling plus the cooling of the power plants feeding the grid).

We publish the full calculation, the calibrated coefficients with their sources, and the uncertainty band. Nothing hides behind “approximate”: if a figure surprises you, you can rebuild it step by step from these formulas and check it against the cited sources. The published figure is operational: it measures inference, not model training or hardware manufacturing (see the lifecycle contrast below).

The formula (methodology v3, 2026-08-08)

Wh_server    = Wh_input  × (input_tokens  / 1,000)
             + Wh_output × (output_tokens / 1,000)
kWh_facility = Wh_server / 1,000 × provider_PUE
liters_total = kWh_facility × (WUE_onsite + WUE_source)

Three ideas distinguish it from the previous version:

1. Input and output at separate rates. The document context (the prefill) is processed in parallel and costs ~5 % per token of generating the answer token by token (the decode). A RAG query carries tens of thousands of context tokens: pricing them at the decode rate — what the previous version did — inflated the figure ~20×.
2. The datacenter profile belongs to the model’s PROVIDER. Mistral inference runs in Mistral’s datacenters in France (nuclear-dominant grid, ~40-60 gCO₂eq/kWh); embeddings run on Google’s fleet. The region of Fluxia’s web server is irrelevant: no model runs there.
3. Per-model rates derive from public measurements. EcoLogits’ fit over the ML.ENERGY benchmark (H100 GPUs, batched) ×3 fleet factor: in Google’s production measurement, the active accelerator is only 58 % of total energy (the rest is host, memory and idle capacity).

Calibrated coefficients (version 2026-08-08)

Per-model rate (Wh per 1,000 tokens, server level, before PUE)

Infrastructure profile per provider

Uncertainty band ±50 %

No provider publishes the energy per token of its inference, and the best public sources disagree with each other:

• Google (2025) measured its real fleet: 0.24 Wh and 0.26 mL for the median prompt — operational, including host, idle capacity and PUE.
• ML.ENERGY / EcoLogits measure the GPU in benchmark conditions with optimal batching — the floor of the range.
• Mistral (2025) published its lifecycle analysis, but without the energy per response or the training/inference split, and mistral-medium’s parameters are not public (estimated ~55B by price position).

The ±50 % band reflects that real spread between sources, not a statistical variance. The previous version published ±30 %: that was spurious precision.

Lifecycle contrast (what the figure does NOT include)

Our figure is operational. Mistral itself, in its lifecycle analysis with ADEME and Carbone 4 (July 2025), estimates 45 mL of water and 1.14 gCO₂e per average 400-token response once model training and hardware manufacturing are amortized. That contrast is also cited inside the assistant (ask it about its own consumption): we prefer the gap between both system boundaries to be visible rather than seemingly non-existent.

Plausibility check

A typical query to the assistant (~35,000 context tokens on the small model + ~5,000 context and ~1,500 output tokens on the synthesis tier):

small : (0.008 × 35 + 0.16 × 0.5)  ≈ 0.36 Wh
medium: (0.011 × 5  + 0.22 × 1.5)  ≈ 0.39 Wh
total : 0.75 Wh × PUE 1.2          ≈ 0.9 Wh  ≈ 0.0009 kWh
water : 0.0009 × (0.2 + 1.6)       ≈ 0.0016 L ≈ 1.6 mL

~1-2 milliliters of water and ~1 Wh per query (operational). Google’s median prompt — much shorter than a Fluxia RAG query — measured 0.24 Wh and 0.26 mL: same order of magnitude once the size difference is accounted for. Mistral’s lifecycle anchor (45 mL/400 tokens) sits about an order of magnitude above, as expected for a boundary that includes training.

What we are NOT counting

We prefer declaring the gaps to inflating the number’s precision:

• Vector-store cost: negligible next to model compute.
• Network cost (CDN, egress): negligible.
• Model training and hardware manufacturing: outside the operational boundary — Mistral’s lifecycle anchor (above) is the reference for that boundary.
• Mistral’s real datacenter PUE/WUE: not published; we use typical modern-EU-datacenter values, declared as estimates.

Reranking and live web search ARE counted (with their own conservative rates) — the previous version of this page said otherwise and was out of date.

Bibliography

1. Mistral AI (2025). Our contribution to a global environmental standard for AI. Lifecycle analysis of Mistral Large 2 with ADEME and Carbone 4. mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai: 45 mL + 1.14 gCO₂e per 400-token response (lifecycle).
2. Elsworth, C., Patterson, D. et al. (2025). Measuring the environmental impact of delivering AI at Google Scale. arXiv:2508.15734. arxiv.org/abs/2508.15734: 0.24 Wh / 0.26 mL for the median prompt, fleet-measured; the active accelerator is 58 % of total energy.
3. EcoLogits (GenAI Impact) + ML.ENERGY Leaderboard. ecologits.ai/latest/methodology/llm_inference: GPU energy per output token as a function of active parameters.
4. Google (2024). 2024 Environmental Report. sustainability.google/reports/google-2024-environmental-report: fleet PUE and on-site WUE (embeddings profile).
5. Patterson, D. et al. (2021). Carbon Considerations when Choosing a Machine Learning Platform. arXiv:2104.10350. arxiv.org/abs/2104.10350: WUE_source per grid mix.
6. Li, P. et al. (2023). Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv:2304.03271 (published in Communications of the ACM, 2025). arxiv.org/abs/2304.03271: split between on-site water and water embedded in electricity.