Responsible AI · measurement · transparency
Mirmi adapts compute to each task, self-hosts the open-source models used by its proprietary engine and measures monitored cloud resources in France at Scaleway.
Dated measurements, documented perimeters and accessible sources.
From compute to measurement
Reading framework
Route
Use the level of compute suited to the task.
Measure
Track the cloud resources covered by the report.
Publish
Connect each measurement to its date and source.
June 2026 measurements with documented sources and perimeters.
Design and operations
Deterministic processing and lighter models handle operations that do not require a heavy model. The most intensive resources are reserved for analyses that justify them.
Each task invokes components matched to its complexity.
Mirmi's proprietary engine orchestrates self-hosted open-source models on SaaS infrastructure located in France at Scaleway, according to the selected deployment.
Location, services and data flows are documented with the architecture.
Mirmi publishes a monthly snapshot of its cloud perimeter and presents comparative benchmarks separately.
Date, unit, source and perimeter accompany every published measurement.
Document flows
20,000–100,000
Range from Arnold & Porter's Green Initiative survey, as cited in the report; the upper bound is equivalent to roughly 50 pages per hour.
~8,300
Conservatree reference point based on Tom Soder's calculation, as cited in the report.
Thousands
Estimate published in the Mirmi presentation for digitised document workflows.
The engine favours in-flow validation and reuse of a single structured source to avoid unnecessary processing, transfers and storage copies.
Cited sources : Mirmi AI: Green Legal Automation, pages 2–3; Arnold & Porter LLP, Green Initiative Survey & Environmental Audits; Conservatree, estimate based on Tom Soder's calculation.
Monthly snapshot
| Operational category | Monthly water | Monthly carbon |
|---|---|---|
| Compute & instancesStandard and L4 nodes | 0.494 L | 6.270 kg CO₂e |
| Managed databasesPostgreSQL / MySQL DEV | 0.240 L | 5.611 kg CO₂e |
| StorageVolume Low Latency & Object Storage | 0.262 L | 8.866 kg CO₂e |
| Total covered perimeterSum of the three categories above | 0.996 L | 20.747 kg CO₂e |
The presentation describes 0.996 L as less than one litre of water per month and compares ~21 kg CO₂e with a 100–110 km car journey.
Infrastructure reference points
The benchmark uses a PUE of 1.25 and WUE of 0.014 L/kWh alongside the published comparisons below.
1.25
Energy-efficiency indicator used for AI workloads.
0.014 L/kWh
Water-use indicator used in the DC5 benchmark.
~100×
Comparison of the 0.014 WUE with a 1.8 L/kWh AI data-center reference.
30–40%
Reference presented against traditional infrastructure in the report.
Benchmark source : Scaleway Data Center Sustainability & DC5 Water Efficiency Metrics, values reproduced in the Mirmi AI: Green Legal Automation presentation, page 4.
Per-query efficiency
0.084 ml
Mirmi water per query
Optimized pipeline value in the benchmark.
0.7–1.5 ml
Short GPT-4o query
External reference used by the report.
8–12×
Less water
Published ratio for the per-query benchmark.
3.884 kg CO₂e/month
Mirmi benchmark footprint
Monthly value for the specific perimeter presented on page 5.
4–8 g CO₂e
GPT-4o query
External reference used by the benchmark.
3–6×
Lower emissions
Published carbon ratio for this separate benchmark.
The monthly snapshot and per-query benchmark cover two complementary perimeters. Each figure retains its own unit, source and period.
Benchmark sources : Scaleway Mirmi AI Environmental Footprint Report — June 2026; How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference, N. Jegham et al., 2025.
Reference documents
Mirmi presentation reproducing a monthly snapshot attributed to the cited Scaleway environmental report and a separate benchmark, kept separate on this page.
Download the PDF presentationOverview of the measurement approach, carbon and water categories and stated alignment with the ADEME reference framework.
View the methodology2025 study cited to contextualize the energy, water and carbon footprint of language-model inference.
Read the studyEvery figure is published with its date, unit and perimeter.
Method questions
The June 2026 snapshot (20.747 kg CO₂e) and the per-query benchmark (3.884 kg CO₂e/month) cover two distinct perimeters. They are presented separately with their units and sources.
The monthly total covers the categories published in Mirmi’s presentation and attributed to the cited Scaleway report. User equipment and the company’s full lifecycle remain outside this snapshot.
The 20,000–100,000 sheets and ~8,300 sheets per tree come from the cited external references. Thousands of pages potentially not printed each month is the estimate published in the Mirmi presentation.
Each new publication retains its date, source and perimeter to track changes in usage, infrastructure, covered products and methodology.
Responsible architecture
We document the architecture, data flows, resources involved and available indicators for your deployment with your teams.
Mirmi