Your "carbon-efficient" AI claim probably only measures Scope 1. Here's why that matters.

#sustainability #showdev #datascience #climatetech

If you're logging power draw at the device or rack level and calling it a carbon efficiency metric, you've built a Scope 1 measurement. That's a real number, but it's not the number most disclosure frameworks (SB-253 included) actually require — and it's not the number most "X% more efficient" marketing claims imply.

Here's the boundary problem, concretely:

Scope 1: on-site fuel combustion, on-site generation
Scope 2: purchased electricity (grid draw) — needs grid-mix intensity data, not just kWh
Scope 3: upstream/downstream — embodied emissions in hardware, supply chain, generation-side losses
Metering the device gets you kWh consumed. It does not get you:

Grid-mix intensity at the time of draw (a kWh at 2pm on a coal-heavy grid ≠ a kWh at 2am on a renewable-heavy one — same meter reading, wildly different Scope 2 number)
Embodied/upstream emissions — none of this is solved by better metering hardware, it requires external data sources joined to your measurement
So if your pipeline logs power_draw_watts and stops there, you have a real number with an undefined boundary. The fix isn't more precise metering — it's tagging every reading with which boundary it belongs to, and refusing to let a Scope 1 number get reported as if it answers a Scope 2/3 question.

A rough schema for what "boundary-tagged" actually looks like:

{
"reading_id": "...",
"value_kwh": 4.2,
"scope": "2",
"boundary": "generation-side",
"grid_region": "CAISO",
"timestamp": "...",
"baseline_ref": "permit-2024-XX"
}
The scope and boundary fields aren't metadata — they're the load-bearing part of the claim. A number without them is unverifiable by construction, no matter how accurate the sensor is.

The other failure mode worth naming for anyone building this kind of pipeline: annualized reporting hides seasonal variance that actually matters (water use at a facility can vary 10,000x by cooling config and season). If your system only emits yearly rollups, you're structurally incapable of catching the weeks that actually stress the grid or water system. Design for seasonal/periodic granularity from the start, not as a later aggregation feature.