Most "green AI" claims stop at Scope 1 and 2 — on-site emissions, purchased electricity. That's the easy part: metered, grid-mix data exists, the math is boring and solid.
Scope 3 is where it falls apart, because the boundary is a choice, not a measurement. Where do you draw the line around an inference request? Chip fabrication upstream? The datacenter's water draw? The marginal grid load at the exact hour the query ran? Every one of those is defensible — and every one gives you a different number, often by an order of magnitude.
The industry's current move is to pick one boundary, quote the resulting number, and present it as the footprint — no disclosure of what's excluded. That's not measurement. That's a chosen boundary wearing a measurement's clothes.
The fix isn't more precision. Nobody needs per-request, per-hop tracking — that's a fantasy and a distraction. The fix is disclosure: state the boundary, state what's in and out, update it on a sane cadence. A number without a stated boundary isn't defensible, it's just a number someone liked.
"Chase the choice, not the number" is a fine principle — but the choice still needs a floor underneath it. Boundary-tagged metering is that floor: not a claim to perfect accounting, but a refusal to let an undisclosed boundary pass as one.
That's the standard we're building CarbonLayer's inference-level accounting against — bounded, disclosed, checkable. Not because it's more precise than the alternatives, but because it's the only version of "defensible" that survives someone asking "defensible according to what boundary?"
Top comments (0)