A peer-reviewed paper out of Cambridge (published in One Earth) quantified something the sustainability world has quietly accepted for years: corporate Scope 3 supply chain emissions estimates can be off by a factor of 10x. Their comparison — it's the equivalent of a company reporting $10M in revenue when the real number is $100M. No auditor signs off on that in financial statements. In carbon accounting, it's business as usual.
Why the error is so large
Most Scope 3 data — which covers 75%+ of a typical company's footprint — isn't measured. It's estimated using industry-average emission factors applied to spend or activity data. That approach is credible enough to satisfy disclosure requirements, and scalable enough to apply across thousands of suppliers. What it isn't is actionable: an average can't tell you which supplier, which process, or which decision to change.
The same failure mode, one layer down
We're building CarbonLayer around a parallel problem in AI infrastructure. Facility-level carbon and water claims for data centers are typically built the same way: averaged grid carbon-intensity factors, applied over annual timeframes, often reconciled with RECs or offsets that reflect annual net-matching rather than hour-by-hour grid draw. It's precise enough to publish. It's useless for the decision that actually matters — which region to route an inference workload to, or which hour to run backup generation, given the actual grid mix at that place and time.
The fix isn't more rigor on the average — it's resolution
The Cambridge paper argues emissions accounting needs to be credible, scalable, and actionable simultaneously. We think the only way to hit all three is to measure at the resolution where a decision actually gets made: per-supplier for Scope 3, per-inference and location-based for AI compute. Anything coarser can satisfy a report. Nothing coarser can drive a reduction.
Top comments (0)