Data center power draw, PUE, and water intake are directly measurable at the facility level. Attributing a slice of that to one single inference is not — it requires a model, and every modeling choice is a place to be wrong.
At CarbonLayer we publish per-inference carbon and water numbers with no offsets and no averaging across a facility. Every number is tagged measured or modeled, because collapsing that distinction is exactly how "water positive" and REC-backed claims end up hiding the actual grid draw.
A few of the harder problems in that pipeline:
Attribution math — splitting facility-level draw across concurrent workloads without just averaging it away
Grid carbon intensity — the same country/year can show a 7x spread depending on which methodology and source you use
Indirect water — the water cost of generating the electricity, which often exceeds onsite cooling water and varies heavily by grid mix and region
Location-based vs. market-based accounting — making sure a REC or offset never quietly substitutes for the real number
We're not asking for funded work. We're asking: if you've worked with grid intensity data, utility filings, PUE disclosures, or data center water permitting, we want you to try to break our model.
Tell us where the attribution logic is wrong, where we're citing a bad source, or where "measured" is doing work it hasn't earned.
Comment or reach out if that's the kind of problem you like taking apart.
Top comments (0)