Why no single carbon value can convert AI compute’s emissions into dollars, and the architecture that makes cost, energy, and carbon composable instead.
AI unit economics has a founding question on the table: should compute efficiency be governed by intelligence per watt or by total cost of ownership per intelligence? This paper argues the question hides a broader one. Financial cost, energy demand, and greenhouse gas emissions are distinct dimensions of the same production, and converting emissions into dollars requires a carbon value whose meaning changes across compliance systems, voluntary claims, policy appraisal, internal planning, and abatement decisions. No single value can serve as a universal conversion factor. The paper develops the alternative: a measurement vector normalized to a defined outcome, an instrument register kept beside the physical accounts, reporting results derived under named rules, and a decision layer where the emissions basis, the carbon value, and the decision purpose must be mutually compatible. A worked deployment case shows the accounting basis redrawing the efficient frontier, a Pareto-efficient option no carbon price ever selects, and capacity pruning the choice set before valuation matters. The constructive close is a Minimum AI Compute Impact Reporting Profile, an Outcome Contract for the denominator, and one completed record.
The placement trade-offs the worked case formalizes, cost, energy, and carbon by GPU type and region, can be explored interactively in the GPU Placement Explorer, an open tool by the same author (source).
This is a living draft. Versions are numbered and dated on the title line. Comments and criticism are welcome: contact@jamesemurphy.com.
James E. Murphy, “No Exchange Rate: Carbon in the Economics of AI Compute,” working draft v1.0, August 2026. Available at https://jamesemurphy.com/ai-economics/carbon-ai-compute/ .