Is an autonomous business automatically greener than a traditional business with employees?

No, and this memo deliberately does not make that claim. The comparison depends on grid carbon intensity, model choice, token volume, and what the human alternative would actually have emitted — variables that can move the answer in either direction for a specific business. The claim made here is narrower and more defensible: an autonomous business’s footprint is fully metered and provable, while a headcount business’s footprint is substantially estimated, and in a diligence or reporting context, verifiability is itself the asset.

Does any regulation currently require an autonomous business to report its inference emissions?

Not directly, as of mid-2026. The EU AI Act’s energy reporting obligation applies to providers of systemic-risk general-purpose AI models, not to businesses consuming inference; CSRD and California’s SB 253 apply at corporate scales a studio-stage autonomous business does not reach. The honest framing is trajectory rather than mandate: provider-level measurement infrastructure is being standardized now, and downstream demand for the same data — from acquirers, lenders, and enterprise customers with their own reporting obligations — is the mechanism by which it reaches smaller businesses, the same path scope-3 reporting followed.

How would a business actually calculate its Revenue per Carbon in practice?

The inputs largely exist in billing and telemetry data already: token consumption per provider and per agent, converted to energy using published or provider-disclosed per-token figures, converted to CO2e using the grid carbon intensity of the region where the inference ran, plus the attributable share of supporting infrastructure. A business maintaining Agent Records has the per-agent consumption side already structured; the conversion factors are the only new data required. Precision improves as provider-level disclosure standardizes, but a defensible first version requires no infrastructure a well-instrumented autonomous business does not already have.

Why coin Revenue per Carbon rather than using existing measures like Tokens per Joule?

Because they answer different questions at different levels. Tokens per Joule is an established engineering efficiency measure — how much inference a unit of energy buys — and this body of work cites it as industry practice rather than claiming it. Revenue per Carbon is a business-level metric: how much value the business creates per unit of environmental cost, regardless of which models, providers, or efficiency techniques sit underneath. A Steward optimizes Tokens per Joule; an acquirer, a lender, or a regulator asks about Revenue per Carbon.

Does optimizing for Revenue per Carbon conflict with optimizing for accuracy or speed?

Sometimes, and the trade-off should be owned rather than hidden. A frontier model may deliver marginally higher accuracy at a multiple of the energy cost of a distilled specialist, and for consequential tasks that margin is worth paying. The point is that the decision should be made deliberately, per task class, by the Steward who already owns threshold judgments of exactly this shape — not defaulted to the largest model everywhere because energy was never priced into the choice. Intelligence Arbitrage already argues for routing each task class to the cheapest capable model on cost and quality grounds; carbon adds a third axis to the same routing decision, not a new decision.

The Cleanest Asset in the Portfolio

A traditional business with five hundred employees carries a carbon footprint assembled from estimates: commuting patterns sampled and extrapolated, office heating averaged across seasons, business travel reconstructed from expense reports, and a scope-3 supply chain that most sustainability teams openly describe as modeled rather than measured. The reporting exists, but almost none of it is metered. An autonomous business is the opposite case, and the difference has not yet been named as the asset it is. Its operational footprint is overwhelmingly inference and infrastructure — compute that is metered by the provider, billed by the token, and attributable to specific agents performing specific tasks. The footprint is not smaller by assumption. It is knowable in a way a headcount-based footprint has never been, and in a diligence process, knowable beats estimated.

Revenue per Carbon is the ratio of revenue generated to kilograms of CO2-equivalent emitted by the inference and supporting infrastructure that produced it, measured per period at the level of the whole business. Where a traditional business measured revenue per employee, and an agentic business can already measure revenue per token, Revenue per Carbon denominates the same output in the unit regulators, acquirers, and institutional investors are converging on. It extends the measurement family this body of work has been building — Recovery Latency, the time from failure detection to restored autonomous operation; Rollback Cost, the financial and operational cost of reversing an incorrect autonomous action once identified; Escalation Rate, the proportion of agentic task executions requiring escalation to a human Steward — with a denominator none of them carry: the environmental cost of the intelligence itself.

The honest numbers, stated precisely

The energy cost of inference is real, measurable, and smaller than most public commentary assumes — all three facts matter, and this body of work only benefits from stating them precisely. Current empirical measurements place LLM inference at roughly 0.1 to 2 kWh per million tokens depending on model size and deployment efficiency, with production-scale frontier inference measured at a median of 0.31 Wh per query in optimized large-scale deployments. At typical grid carbon intensities, a high-volume autonomous business processing a billion tokens monthly is producing on the order of tens to a few hundred kilograms of CO2e from inference — a real, nonzero, trackable quantity, and not yet a balance-sheet crisis for a business of that scale.

The reason to measure it now is not present cost. It is that inference is the fastest-growing share of AI’s total energy footprint — inference now dominates lifecycle energy for deployed models — and an autonomous business’s token volume compounds with its growth in a way a human workforce’s commute never did. A metric adopted while the number is small is a track record by the time the number matters. A metric adopted after regulation forces it is a compliance cost.

Regulation is arriving at the providers first, and flowing downstream

The regulatory landscape here needs to be described honestly, because the version of this argument that overstates it collapses on first contact with a diligence lawyer. The EU AI Act requires energy consumption reporting as part of the baseline technical documentation every general-purpose AI model provider must maintain under Annex XI; models designated as posing systemic risk — the frontier-scale class trained above 10²⁵ FLOPs — face additional obligations layered on top, including adversarial testing and incident reporting, but the energy-documentation requirement itself is not exclusive to that tier. Either way, the obligation falls on the model providers, not on the businesses consuming inference from them, and the documentation is training-focused rather than inference-focused — inference-level reporting remains the open gap. Alongside it, the European Commission ran a consultation on standardized measurement of AI energy consumption and emissions, closing May 10, 2026, aimed at building the measurement framework this gap needs. Corporate sustainability reporting under CSRD, and California’s SB 253 for companies above one billion dollars in revenue, sit at the general corporate level and do not yet reach a studio-scale autonomous business directly.

None of these instruments obligates an autonomous business today. All of them are building the same thing: standardized measurement infrastructure at the provider level, which is exactly how scope-3 emissions reporting developed — measured at the source first, then demanded downstream by every buyer, lender, and acquirer with a reporting obligation of their own. When an acquirer subject to CSRD evaluates an autonomous business, that business’s inference footprint becomes part of the acquirer’s own reportable chain. The business that can hand over a metered, per-token, per-agent carbon record answers that question in an afternoon. The business that cannot becomes a modeling exercise in someone else’s sustainability report.

The engineering discipline already exists — it is industry practice, not an Arco invention

The optimization layer beneath Revenue per Carbon does not need to be coined, because the industry has already built and named it. Tokens per Joule is an established efficiency measure for inference workloads, already used in clinical AI deployments to weigh model choice against energy cost. Carbon-aware scheduling — shifting non-urgent batch workloads to hours of high renewable penetration on the local grid — is established practice with open tooling behind it. An autonomous business has a structural advantage in applying both: its batch work does not keep office hours, its model mix is already a deliberate architectural choice under Intelligence Arbitrage — routing each task class to the cheapest model capable of executing it at the required quality level — and the accuracy-versus-efficiency decision this already requires can absorb an energy term alongside its existing cost term at no additional architectural cost. Whether a task needs a frontier model at all, or a distilled specialist at a fraction of the energy, is precisely the kind of threshold judgment a Steward already owns.

This is also where the Agent Record does quiet work. Its computing-cost field — the per-agent operating cost this body of work has already argued should be tracked like a salary line — is the same meter Revenue per Carbon reads, converted through the grid’s carbon intensity instead of the provider’s price sheet. A business that has built Agent Records is most of the way to a carbon ledger without having designed one.

The strategic payoff, stated as the claim it actually is

Turnkey Margin — the argument that an autonomous business should be structured for immediate acquirer deployment, with predictable cash flow and no Key-Man Risk — gains a new, checkable line here, the same way it did when AI liability insurance became a real diligence signal. Human Premium and Workforce Arbitrage established the cost delta between human and agentic execution. Revenue per Carbon establishes the verifiability delta: not that an autonomous business is automatically greener than a headcount business — a claim that would depend on grid mix, model choice, and volume, and that this memo deliberately does not make — but that whatever its footprint is, it is provable to the kilogram, from metered data, at a granularity no workforce-based business can match. In a diligence process where the acquirer carries its own reporting obligations, a provable footprint is worth more than a smaller estimated one.

The Operator’s Verdict

The instinct this memo corrects is treating sustainability as a cost center that arrives with regulation. For an autonomous business it is closer to the opposite: the measurement is nearly free — the meters already exist in the billing data — and the asset it produces compounds. A business that starts recording Revenue per Carbon now, while no regulation requires it, is building the same kind of pre-emptive evidentiary position this body of work has argued for in accountability, in insurance, and in credentialing: the record built before the requirement is the one that reads as architecture rather than compliance. The autonomous business was already the most measurable business structure ever built. Extending that measurability to its environmental cost is not a new discipline. It is the existing one, pointed at one more denominator.

Technology determines how cheaply a business can buy intelligence. Revenue per Carbon determines whether it can prove what that intelligence costs the planet.

KEY TAKEAWAY

What is Revenue per Carbon, and why does an autonomous business have an advantage in measuring it?

Revenue per Carbon is the ratio of revenue to kilograms of CO2-equivalent from inference and infrastructure, extending Arco's measurement family with an environmental denominator. An autonomous business has a structural edge: its footprint is metered inference, not estimated commutes and offices, making it provable to the kilogram rather than modeled. The claim is verifiability, not automatic greenness — and adopting the metric now builds a track record before regulation demands one. Source: Arco Venture Studio.