Carbon / power observability (estimate)

What this is

The sustainability (ESG) view of network traffic: an estimate of how much electricity (kWh) and carbon (gCO2e) the network spends moving your bytes, broken down by traffic class, service, and team. It is the same idea as Kepler — attributing energy to workloads — applied to the network plane, and it is computed entirely from telemetry probectl already has.

It lives in the control plane (internal/carbon) and reuses the FinOps engine's plumbing: the same flow stream and the same zone/owner attribution, so the dollar view and the carbon view always agree on who owns which traffic. It is fully local: no external calls, and the grid carbon intensity is operator-set configuration — per the no-phone-home rule in the Non-negotiables, nothing is fetched and nothing leaves your network.

The estimate, and its honesty

Two multiplications:

kWh   = (bytes / GiB) × kWh-per-GB(traffic class)
gCO2e = kWh × grid intensity (gCO2e/kWh)

• Energy coefficients — fixed-network transmission energy, from the published literature (the Aslan et al. 2017 band and its successors). probectl scales the coefficient by path locality, because traffic crossing more network hops costs more energy to move. The defaults span roughly 0.004 to 0.06 kWh/GB:

  Traffic class	kWh per GB
  same_zone	0.004
  inter_az	0.01
  inter_region	0.03
  internet_egress	0.06
  unknown (zones unmapped)	0.01 (conservative mid-band)

• Grid intensity — PROBECTL_CARBON_GRID_GCO2E, the carbon your electricity carries per kWh. It defaults to roughly the world average (436 gCO2e/kWh). Set your real figure — your utility or national grid operator publishes it, and it varies enormously by region (a hydro-heavy grid is a fraction of a coal-heavy one).

• Attribution — the exact same zone and owner maps the FinOps engine uses (PROBECTL_COST_ZONES / PROBECTL_COST_SERVICES), so dollars and grams line up on the same services and teams.

The honesty contract is structural, not a footnote. Every response carries a methodology block: measured: false, the coefficient source, the grid intensity used, and a note stating plainly that these are coefficient-based estimates of transmission energy, not measured device power. They are good for relative attribution ("service A moves ~3× the carbon of service B"), trends over time, and ESG reporting that cites its methodology — they are not audited absolute footprints, and the engine never pretends otherwise.

Serving

GET /v1/carbon (permission metrics.read) — the tenant's totals, the breakdowns by class / service / team, a 7-day hourly trend, and the methodology block. In the UI, the Carbon / energy (estimate) card folds into the Cost page — the sustainability sibling of dollar showback. carbon_running: false distinguishes an engine that is not wired in from one that simply has no data yet.

Configuration

Deferred by design:

• Device-measured power. The methodology block's measured flag is the structural seam for one day distinguishing real device power (e.g. PSU watts over SNMP) from coefficient estimates — today it is always false, no supported device reports watts, and the engine never fabricates them.
• Per-region grid intensities. For now a deployment uses one grid figure; there is no per-region split.