Every prompt draws power, carbon and water. This instrument measures five dimensions of that draw, places your usage on a decision matrix, and scales the result to a planetary view. Set the dials to match how you actually work.
No black box. Here is the exact arithmetic behind the readings above, with your own numbers substituted into every term. If you disagree with a coefficient, change the input and watch the chain re-derive.
EIT = ( Tout · δ + Tin · π ) · C · (1 + R)
Esite = EIT · PUE
CO2e = Esite · Igrid H2O = Esite · ( Wsite + Wgen )
Low draw, real displacement. This is the pattern to scale and standardise across the organisation.
Heavy, but it earns its keep. Optimise the infrastructure and routing rather than cutting the use.
Individually negligible, collectively enormous. Trim through defaults and smaller models, not policing.
Expensive compute producing little. Eliminate first — this is where the fastest reductions live.
| Dimension | Level 1 — Critical | Level 2 — Wasteful | Level 3 — Acceptable | Level 4 — Efficient | Level 5 — Regenerative |
|---|---|---|---|---|---|
| G — Generation footprint 25 pts |
Above 8 g CO₂e per task. Untracked, unmeasured. | 2–8 g. Estimated once, never revisited. | 0.5–2 g. Measured quarterly at portfolio level. | 0.05–0.5 g. Per-workload telemetry in place. | Under 0.05 g, with per-request attribution to a named owner. |
| R — Right- sizing 25 pts |
One frontier model for everything, including formatting. | Manual model choice, biased to the largest available. | Documented guidance on which model suits which task. | Tiered routing with a small-model default and escalation rules. | Automatic complexity routing, cascade fallback, measured accuracy floor. |
| E — Interaction efficiency 20 pts |
Three or more re-prompts typical; whole history resent each turn. | Frequent retries; no caching; unbounded output. | Some prompt templates; occasional context pruning. | Prompt library, output caps, prefix caching enabled. | Single-pass success as a tracked metric; retrieval instead of resent context. |
| E — Energy provenance 20 pts |
Unknown region, unknown PUE, water-cooled in a stressed basin. | Fossil-heavy grid, PUE above 1.5. | Regional average grid, PUE near 1.4, annual REC matching. | Low-carbon region, PUE under 1.3, closed-loop cooling. | Hourly carbon-free matching, PUE under 1.15, near-zero potable water draw. |
| N — Net value displaced 10 pts |
Entertainment and novelty at production volume. | Convenience gains only; no measurement attempted. | Claimed productivity benefit, anecdotally evidenced. | Measured hours or defects avoided, tied to a baseline. | Verified displacement of a higher-carbon activity, audited. |
How to read it. Score each row independently, then sum. The weighting is deliberate: right-sizing and generation footprint carry half the total because model selection is the single largest lever a user controls — larger than grid choice for most organisations, and available immediately.
Reductions are modelled against your current settings and are not additive — each assumes the others stay as they are.
The greenest model is not the smallest one — it is the smallest one that clears the quality bar, because a model that fails the task pays full price twice. Describe the workflow, and every candidate is costed on money, carbon and capability together.
| Model | Tier | $/M in | $/M out | $ / 1K tasks | kg CO₂e / 1K tasks | Annual cost | Annual tCO₂e | Green | Abatement $/tCO₂e |
|---|
Why the abatement figures are so large. At realistic volumes the money saved by switching models outweighs the carbon saved by three to five orders of magnitude in dollar terms. That is not a rounding error — it is the central strategic fact on this page. No plausible carbon price will ever drive model selection, so do not build the business case on carbon. Build it on cost, and bank the emissions reduction as a consequence. Sustainability programmes that depend on a carbon price to justify themselves lose the budget argument; cost-reduction programmes that happen to cut emissions survive it.
A = 1 + max(0, tier_required − tier_model). This is what stops the cheapest model winning by default.$ = (T_in · P_in + T_out · P_out) ÷ 1,000,000 × calls × A, where calls is 1 for chat-style work and higher for agentic archetypes.Wh = (T_out · d + T_in · p) × calls × A, using the decode and prefill coefficients for the model's size class from section 03.gCO₂e = Wh ÷ 1000 × PUE × I_grid, taking PUE and grid intensity from your settings in section 01.MAC = (Cost_model − Cost_baseline) ÷ (CO₂_baseline − CO₂_model), in dollars per kilogram of CO₂e avoided. A negative MAC means the switch saves money and carbon at once — a no-regret move that needs no sustainability budget to justify. A positive MAC is the real price of that abatement, and should be compared against your internal carbon price before you claim it as a green initiative.Two warnings before this drives a procurement decision. First, list prices move constantly and the defaults here reflect published rates around August 2026 — verify every row against the provider's current rate card, and note that batch processing and prompt caching can each cut the cost column substantially without changing the carbon column much. Second, the capability tiers are a coarse judgement, not a benchmark result: the honest version of this table is one where you have run your own evaluation set and replaced the tier column with your own pass rates.
Sections 01 to 09 tell you where you stand. This tells you what to do. You have set roughly a dozen dials across this page — model, region, facility, tokens, workload. There are hundreds of valid combinations and no human picks the best one by eye. The agent does it arithmetically, and shows its working.
Reads every input you have set and looks for contradictions: a real-time promise on a workflow that cannot be real-time, a retry rate that signals a broken prompt, evaporative cooling on a coal grid. It tells you what is wrong before it recommends anything.
Enumerates every combination of model, region, facility efficiency, cooling, prompt caching, batching and output policy that your constraints permit — typically several hundred — and scores each on carbon and cost together.
Returns one deployable configuration, the projected annual cost and carbon against your current baseline, and an eight-line trace of how it got there — including the runner-up, so you can see whether the decision was close.
A worked example of the value. On a support-triage workflow at 120,000 tasks a month, most teams default to a mid-tier frontier model on whatever region their cloud account defaults to. The agent typically finds a smaller model that clears the same quality bar, on a cleaner grid, with a tighter output cap — and returns something in the order of a few thousand dollars and a fraction of a tonne of CO₂e saved a year, for the same work. Run it and read the numbers for your own settings rather than mine.
Set the four constraints below to describe what you are actually allowed to change. If you are unsure, the defaults are the common enterprise case.
Where the law or your contracts say inference may physically run. Picking a region is the single biggest carbon lever — if you cannot move, say so and the agent will stop offering it.
How fast the answer must come back. Batch tolerance unlocks a fifty per cent cost cut; real time rules out multi-step agents entirely.
How much quality you will trade for savings. Strict protects the answer; cost-first accepts a weaker model and pays for the extra attempts it needs.
← all carbon · all cost →
Which you care about more when the two disagree. They usually agree — slide it to both ends and see whether your answer even changes.
Your per-query figures, scaled to a planetary user base. Adjust the assumption and watch the sensitivity — the honest finding of this exercise is how much the answer depends on inputs nobody publishes.
Read the gap, not just the number. For ordinary chat usage this projection lands well below actual data centre consumption — and that gap is the finding. Serving prompts is only a slice of AI's footprint; training runs, idle capacity held warm for latency, networking, storage and the embodied carbon of chips and buildings make up the rest. Push the model class to agentic and the gap closes fast, which is the clearest signal in the whole instrument: workload design, not chat volume, is what will decide AI's environmental trajectory.
Partly, and in a specific way. The environment is not being ignored — it is being measured on the wrong axis. Efficiency per task is celebrated because it is improving spectacularly. Total load is downplayed because efficiency is improving spectacularly. Both are reported honestly and the combination is misleading, because the quantity that decides the outcome is neither one alone: it is whether workload growth, net of efficiency, outruns the growth of clean supply.
The asymmetry is the finding. Capital, compute and demand are all compounding at double or triple digits. The one line that has not moved is disclosure — not a single frontier lab publishes full per-model energy telemetry, which is why every framework on this page, including this one, is estimating rather than auditing.
A carrying-capacity model with three variables. Workload volume grows. Efficiency cuts the energy each unit of work needs. Clean supply grows to meet what is left. Equilibrium holds only while the first two, combined, stay below the third.
Net demand growth = (1 + g) · (1 − e) − 1
Equilibrium holds while (1 + g) · (1 − e) ≤ (1 + c)
g = workload growth · e = efficiency gain · c = clean supply growth
Each condition below moves one variable in the equation above. The order is deliberate — it runs from what you control this week to what takes a decade and a parliament.
Efficiency is losing a race it did not choose. At the default settings the industry needs a far larger annual efficiency gain than it is achieving, and the gap is made of workload volume, not model design. Cap agent step counts, route to small models by default, kill the retry tax. This is the only variable that responds in days.
Quantisation, distillation, sparsity, better serving stacks and speculative decoding have delivered genuine order-of-magnitude gains and must continue. But an efficiency gain announced without the corresponding volume figure is a half-truth, and the half that is missing is the one that decides the answer.
Buying existing clean electricity displaces another consumer onto fossil generation and changes nothing at system level. Only additional capacity, contracted ahead of the load it serves, moves c. Hourly matching makes the claim auditable; annual certificates do not.
A crossover is a peak problem before it is an annual problem. Grid-interactive data centres that curtail or shift training and batch inference into clean hours buy real years of runway without reducing a single unit of work done.
None of these variables can be managed at the values they actually hold, because nobody publishes them. Independent estimates for the same model differ by an order of magnitude. Until per-model energy telemetry is standard disclosure, equilibrium is a claim, not a measurement — and this framework is a well-reasoned guess.
The honest summary. The environment can sustain AI adoption indefinitely at some rate. The question was never whether, but at what rate — and the current rate is set by capital allocation and product ambition, with the environmental constraint discovered afterwards through interconnection queues, water permits and local opposition rather than chosen deliberately in advance. Equilibrium is available. It is simply not what anyone is currently optimising for.
Tick what you will genuinely commit to. The figure at the bottom applies diminishing returns rather than adding percentages, because these measures overlap.
The order matters. Right-sizing and retry reduction are behavioural, free and available today. Region and cooling choices are procurement decisions with a lead time of weeks. Grid and policy measures take years. Do the first tier while you argue about the third.
Identical compute. Identical answer. Eighteen times the emissions between the cleanest and dirtiest column. This is why siting is a bigger environmental decision than usage policy — and why it is made by people who never see a prompt.
Efficiency is not reduction. Power per AI task is falling at a rate unprecedented in energy history, while total consumption climbs. Both statements are true, and each camp quotes only its own line. The gap between them is the rebound effect, and agentic workloads are its engine.
Where this framework is weak, stated plainly. Per-query energy figures are estimates, not measurements — no major lab publishes full telemetry, and independent estimates for the same model vary by an order of magnitude. Embodied carbon in chips, buildings and networks is excluded from the per-query figures and folded into a single crude amplifier. Water intensity varies by site and season far more than any single coefficient can capture. Treat every number here as a decision aid with wide error bars, not an audit result.