Why AI Compute Regulation is Heading Straight Into the Scope 3 Trap…and How to Fix it.
During my time at JPMorgan, I watched financial institutions grapple with the push toward Scope 3 carbon disclosure — first through voluntary international frameworks, and eventually under mandatory regimes like Europe’s CSRD.
Scope 3 forced organizations to quantify indirect, value-chain emissions using porous boundaries and proxy estimates. An entire consultancy industry emerged, yet the outputs remained imprecise, un-auditable, and endlessly complex.
Now, as policy conversations turn toward regulating AI data centers based on "inference capacity" or "effective model output," we are setting up the exact same trap.
Why measuring inference capacity is a fool’s errand:
Software optimization moves too fast: Techniques like quantization (INT4 vs. FP16) and Mixture-of-Experts (MoE) dramatically change effective model capability on the exact same hardware.
The boundary problem: A single query routed across multi-tier APIs, fine-tuned edge models, and cloud clusters creates the same porous, un-auditable boundaries that broke Scope 3.
The Solution: Keep regulation precise, physical, and pared down.
Instead of trying to audit fluid software execution, policy should stick to direct, metered "Scope 1 & 2" physical baselines:
Utility Interconnect Approvals (MW): Hard electrical boundaries metered directly at the substation.
Thermal Design Power (TDP): Clear energy-draw baselines for facility cooling and power architecture.
Physical Silicon Tallies: Verifiable hardware counts at the facility wall.
Rule #1 of practical policy: If a metric can’t be measured deterministically at the physical boundary wall, don't build a mandate around it.
Regulators should focus oversight on concrete, physical inputs, but keep fluid, software-defined inference metrics off the ledger.
#AIRegulation #TechPolicy #DataCenters #CorporateGovernance #AIInfrastructure
