Ask three carbon accounting frameworks to report the emissions of the same data center, in the same country, in the same year, and you can get numbers that differ by a factor of 7. Not because anyone is lying. Because "emissions" isn't one number — it's a methodology choice wearing a number's clothes.
The two accounting standards that disagree
Location-based accounting measures the actual emissions intensity of the grid a facility draws from — the physical mix of coal, gas, nuclear, and renewables serving that specific region, at that specific time.
Market-based accounting lets a company buy Renewable Energy Certificates (RECs) or offsets and report as if it drew clean power, regardless of what the local grid actually generated when the servers were running.
Both are "correct" under different reporting standards (the GHG Protocol permits both, reported side by side). But only one describes what actually happened at the facility. The other describes what was purchased on paper.
Why this isn't a rounding error
A data center in a coal-heavy grid can buy enough RECs to zero out its market-based number while its location-based number — the actual carbon intensity of the electrons it drew — stays high. Report the market-based number alone, and a facility running on a dirty grid at 2pm on a hot day looks indistinguishable from one running on hydro at midnight.
Neither number is fake. But they answer different questions:
Location-based: what did this facility actually draw from the grid?
Market-based: what did this company pay to offset?
Most public sustainability reporting picks whichever number is smaller and calls it "emissions," full stop — no methodology footnote, no disclosure of which standard produced it.
Why this matters more for AI inference than almost anything else
Data center load isn't constant — it spikes with usage, time of day, and now, increasingly, with model routing decisions. A router that shifts inference to "the greenest region" using a market-based, annually-averaged REC number can be routing to a facility that, at that literal moment, is drawing from a coal-heavy grid at peak demand. The certificate says clean. The grid says otherwise.
The fix isn't picking a side — it's disclosure
The honest answer isn't "location-based is right, market-based is wrong." It's that any number reported without stating which methodology produced it is not a number you can act on, compare, or route decisions against. If your carbon dashboard gives you one figure with no methodology field, you're being asked to trust an average you can't audit.
That's the bar we hold CarbonLayer's own numbers to: every inference-level carbon and water figure ships with the accounting method that produced it, not just a total.
Top comments (0)