What exists
- Prometheus scraping every validator on port 9615: CPU, memory, disk, network, block and finality metrics.
- A finality exporter with p50/p95/p99 and a Grafana dashboard for decentralization.
- A convention for dated measurement artefacts (
artifacts/performance/,artifacts/fees/,artifacts/decentralization/).
What is missing, in order
1. Metered power. The hosting provider's per-server draw, or a rack PDU exporter. Without a watt figure, everything downstream is estimation. The two validators are in a shared-hosting provider's data centre; whether metered draw is available per server is a question for the provider.
2. PUE. The data centre's power-usage-effectiveness ratio, from the provider. Not estimated from industry averages.
3. Emissions factor. The grid's carbon intensity for the data centre's location and hour, from the grid operator's public feed. Spain publishes this in near real time.
4. A boundary and a denominator. Which servers: validators and RPC (two), Ailoos coordinator and inference (one, doing real compute), app backends (a cloud provider's regions), the phones that train (their owners' electricity). And what to divide by: blocks (constant, six seconds), transactions (tiny and bursty on a testnet), finalized state (meaningful but abstract). Publishing the per-transaction number without the transaction count is how proof-of-stake chains produce absurd figures in both directions.
5. Publication. A Prometheus metric (empoorio_power_watts{host}, empoorio_emissions_gco2_per_hour) and a dated artefact, with method.
Why the AI servers must be in
Ailoos does real work: inference on a quantized general model, EmpoorioLM training coordination, dataset serving. Its one dedicated server likely draws more than both validators. A feed that counts only consensus would flatter the chain and mislead about the ecosystem. The boundary is the ecosystem's.
Why phones are out, and noted
A phone training a LoRA adapter for thirty steps uses its owner's electricity and would have been charging anyway. It is not the ecosystem's footprint to report; the design choice to train there is an energy choice and belongs in the method note.
Priority
Below weights, audits, a faucet and an indexer. Above publishing any energy figure at all — which is why none appears on this site until the feed exists.
Based on CONSENSUS_STATUS.md (metrics), VALIDATOR_GUIDE.md (monitoring) and the ecosystem's rule on unmeasured figures.