Summary

  • A watt reading is an observation of power at a moment and within a boundary. Energy adds time; emissions add the electricity source and an allocation method; lifecycle impact adds manufacture, cooling, replacement and end of life.
  • RFC 9845, a collective IRTF research document edited by Alexander Clemm and Carlos Pignataro, shows why network efficiency is not proportional to traffic. Idle equipment can retain a large base load, while switching it off can consume redundancy and recovery margin.
  • A defensible controller keeps an append-only receipt for the meter, interval, attribution rule, carbon source, baseline, decision, affected capacity, service constraints, observed outcome and rollback. A proxy label must never silently become measured permission.

At 02:00, a router reports 620 watts. At 02:05, after a traffic shift, it reports 570. The arithmetic is clean: fifty watts disappeared from that chassis. A sustainability dashboard can make the rest look equally simple—green arrow, estimated saving, successful policy.

But the first reading did not say how long either state lasted. It did not say whether traffic moved to another router, whether a dormant line card began drawing power elsewhere, which electricity source supplied each site, or whether a spare path was removed. It certainly did not include the carbon embedded in the equipment or the service cost of a slower recovery after failure.

This is the productive discomfort in RFC 9845, Challenges and Opportunities in Management for Green Networking. Published in October 2025 as an IRTF Informational document, it reflects consensus in the Network Management Research Group. It is not an Internet Standards Track specification and it does not certify a deployment. Alexander Clemm and Carlos Pignataro are its editors; Cedric Westphal, Laurent Ciavaglia, Jeff Tantsura and Marie-Paule Odini complete the six-person author group.

Pignataro's biography explains the perspective but does not expand the evidence. At the 1 September 2026 capture, the IETF profile described his previous Cisco sustainability, Distinguished Engineer and CTO roles, his adjunct work at North Carolina State University and his fellowship in the National Academy of Inventors. Blue Fern Consulting described him as Founder and Principal. Those facts locate a career across networking and sustainability. They do not make him the operator of the hypothetical router, the owner of an emissions factor or the sole author of the RFC.

Watts, watt-hours and carbon use different clocks

Power is a rate. A watt is one joule per second. Energy accumulates that rate through time. The difference is operationally basic but easy to erase in software: a live telemetry field arrives every few seconds, while a monthly claim asks how much energy a service consumed. The answer requires an interval, an integration method and a decision about missing samples.

Even a correct energy total does not identify greenhouse-gas emissions. One kilowatt-hour drawn where low-carbon generation is available at noon may have a different consequence from one drawn from a fossil-intensive marginal source at night. A carbon result therefore needs the location, time window, source of grid intensity and accounting rule. Annual averages, contractual renewable claims and marginal operating signals answer different questions.

Then comes the boundary outside the socket. RFC 9845 notes the energy and carbon embedded in manufacturing equipment, the power required for cooling, the development and deployment effort, and the consequences of recycling or replacement. The document deliberately focuses much of its analysis on operational energy, while warning that a holistic result would go further. Replacing an old router with a more efficient one may lower the electricity bill and increase embodied impact. Extending a platform through software may avoid manufacture but prolong a less efficient running state.

The word “green” compresses these clocks and boundaries. The fix is not a better adjective. It is a chain of claims: observed power; integrated energy; attributed energy for the service; estimated emissions under a named source; lifecycle comparison under explicit inclusions; operational decision under stated service constraints. Each transition needs its own receipt.

The idle base turns traffic efficiency into a step function

Network equipment rarely behaves like a perfect proportional meter. RFC 9845 cites studies in which an idle base system typically draws more than half of the energy used at full load. Consumption then rises in steps as components, ports or line cards become active. Once a system is awake, carrying another packet can have a small marginal cost.

That changes the optimisation problem. Spreading traffic evenly across ten awake devices may look balanced and resilient, yet preserve ten base loads. Concentrating it on six could let four enter a deeper power state. A per-bit metric can improve merely because the denominator grew, while total demand keeps rising. Conversely, a lightly loaded device can look inefficient even when it is retained for a failure scenario that the service contract requires.

So three statements that resemble one another must remain separate. “This box uses fewer watts at the present load” is an equipment observation. “This traffic shift reduced total network energy over the hour” is a system result. “This design reduced emissions without degrading the service” adds grid and reliability evidence. None follows from the previous line without inspecting the rest of the estate.

The counterfactual matters as much as the reading. What would the same demand have consumed without the intervention? Was the load displaced in place, geography or time? Did a later traffic spike force a rapid restart with additional energy and packet loss? Did a capacity planner add new hardware because the now-tighter margin appeared unsafe? A result without the counterfactual can celebrate movement rather than reduction.

Sleeping capacity carries an availability price

Powering down unused resources is one of the most promising opportunities in the RFC, and one of the clearest examples of why the metric cannot authorize the action. Network demand is bursty. Failures are not scheduled. A dormant device takes time to discover, initialize, synchronize state and rejoin forwarding. Routes may reconverge. Queues can build while the controller decides that yesterday's demand model no longer applies.

Spare capacity is not automatically waste. It may be the physical form of an availability commitment. Redundant paths, hot standby systems and headroom for a sudden load are resources held against uncertainty. Removing them can reduce measured power in the ordinary state and increase the probability or duration of an exceptional outage.

A carbon-aware controller therefore needs more than an objective function. It needs guardrails for minimum capacity, diverse failure domains, latency, packet loss, wake-up time and recovery. It needs a scenario for a link cut during the deepest saving state. It needs permission scoped to specific resources, and a rollback that works before the service-level breach becomes irreversible.

Heng Lu's running-code principle is useful here. A deployed mechanism can prove that a particular router entered a low-power state and returned under test conditions. It cannot by that act claim authority over every operator's resilience policy. The minimum specification should expose measurements and safe controls, leaving the later operational choice local and attributable. The code is evidence of mechanism; the decision remains with the party bearing the outage.

A label is not a meter, and a meter is not attribution

RFC 9845 is candid about instrumentation. Device-, line-card- and port-level power data are not uniformly available, and standardized management models have lagged the need. Until direct measurement exists, equipment may advertise a category or proxy—perhaps a rating, a processor type or a nominal profile.

A proxy can be useful for inventory and planning. It becomes dangerous when a processing chain removes its type. efficiency_class: gold is not measured_power_watts: 570. A nominal maximum is not a live reading. A model estimate is not a utility meter. The record must say which it is, which version produced it and when it should expire.

Virtualization adds another join. A network function may run in a container on a shared CPU, inside a server sharing fans and memory, in a facility sharing cooling and power conversion. The host meter can be correct while the service allocation is arbitrary. CPU time, reserved cores, bytes forwarded and peak capacity distribute shared costs differently. A per-flow carbon number without its allocation rule is a conclusion with the method removed.

Carbon-aware routing makes these weaknesses active. Once a controller treats an emissions estimate as path cost, stale grid data or a manipulated device label can move real traffic. The RFC accordingly points to compliance, misleading data and attack surfaces that can turn resources off or waste energy. Freshness, provenance and failure behaviour are not reporting details; they are inputs to control-plane safety.

The useful object is an operational receipt

A compact green badge is attractive because it travels easily. A receipt is more demanding because it preserves disagreement. For every intervention, retain the meter or proxy identity; device, component, host, path or flow boundary; unit and sampling interval; raw and derived values; missing-data treatment; attribution rule; traffic denominator; grid-intensity source and timestamp; lifecycle exclusions; baseline and counterfactual; controller and policy version; service constraints; affected resources; observed result; rollback trigger and recovery record.

That list is not a demand for one universal repository. It is a demand that the joins remain reconstructable. Equipment teams may own measurements. Facilities teams may own electricity and cooling. Sustainability teams may own accounting methods. Network operations owns resilience constraints and the action. The shared identifier must connect their records without letting a later score overwrite the original evidence.

The best sentence after an intervention is therefore modest: under this interval, boundary, grid source and service constraint, the observed system used less energy than the stated counterfactual, with no detected breach during the observation window. That is narrower than “greener network.” It is also testable, correctable and useful.

Sources