Graphique en barre

AI’s ROI is measured in dollars. By 2027, it will also have to be measured in energy.

In 2026, one debate was settled: AI applied to building and running information systems has a proven return on investment, paid back in under a year. The conversation then shifted to the cost of the token, now a budget line in its own right. But that equation is still incomplete. Every token consumed carries an energy, water, and carbon cost that almost no organization tracks, and this variable becomes structural as early as 2027.

Key takeaways

  • The cost of the token is now actively managed, but only in dollars. As AI moves from pilot to large-scale production, its energy bill follows the same curve. It is the invisible half of the “cost per outcome.”
  • The Jevons paradox defeats the dashboards. Energy per task is collapsing at an unprecedented rate; token volume, however, is exploding even faster. The result: total consumption doubles despite the efficiency gains.
  • Regulation locks the topic in. The EU AI Act, CSRD, and ISO 42001 bring AI’s footprint into mandatory documentation and governance. The key deadline falls on August 2, 2027.
  • The decision. Move from an AI FinOps counted in dollars to a dual-metric approach: cost per outcome in dollars and in grams of CO₂. Measure before optimizing, route before governing.

The debate has moved on, but it is missing a column

The 2026 consensus is clear: it is by equipping engineering, the “build” and “run” of IT, that AI creates its most measurable value. The token has thus become a structural OpEx line, on par with the enterprise’s major cost centers. At Mistral, whose core business is AI, it already represents about 10% of payroll. Gartner expects that by 2028, the cost of coding tokens could exceed the average developer’s salary.

These analyses reason in dollars. None counts the energy. Yet a token is not an accounting abstraction: it is a physical computation, run in a data center, that consumes electricity and water and emits carbon. As long as AI stayed at the pilot stage, this dimension was negligible. At the scale of industrial production, it becomes a budget line, not a footnote.

The paradox that makes the topic counterintuitive

The most common intuition, “AI is getting more efficient, so its footprint will shrink,” is true for the token but false for the total.

On the efficiency side, the progress is spectacular. In one year, Google cut the energy of a median Gemini text prompt by 33x, and its carbon footprint by 44x, largely through software optimization. The International Energy Agency observes the same dynamic: energy per task is falling at a pace without precedent in the history of energy.

On the volume side, the trajectory is the opposite. Usage is exploding: according to Goldman Sachs, global token consumption could multiply 24-fold by 2030. And each task now consumes more: a reasoning model can use more than 70 times the energy of a lightweight model to answer, because it generates far more intermediate tokens. Autonomous agents, which chain calls together, amplify the phenomenon further.

The net result is a textbook case of the Jevons paradox: the more efficient and cheap the token, the more of it we consume. The IEA thus projects a near-doubling of data center electricity consumption, from 485 TWh in 2025 to nearly 950 TWh in 2030, roughly 3% of global electricity demand, with the share dedicated to AI tripling over the period. A “cost per outcome” that falls in dollars can therefore come with an energy and carbon footprint that climbs. This is precisely the column today’s dashboards do not calculate.

Not all tokens are equal

This is also what makes the topic actionable rather than fatalistic. The life-cycle assessment published by Mistral with ADEME and Carbone 4 establishes a clear correlation: for the same number of tokens, a model ten times larger generates impacts an order of magnitude higher. In other words, the first lever of frugality is not technical: it is an allocation choice, the right model for the right task.

Sending a ticket classification, a data extraction, or a summary to a frontier model wastes budget and energy at the same time. Reserving reasoning power for high-value cases, routing volume to lightweight or specialized models, optimizing the context: these choices reduce the financial bill and the footprint in a single move. Both metrics point in the same direction, which makes AI frugality a performance issue, not a constraint.

Regulation turns good practice into obligation

What was yesterday a matter of voluntary responsibility is becoming a compliance imperative.

The EU AI Act now requires providers of general-purpose AI models to document energy consumption, known or estimated, for their models. Models already on the market before August 2025 benefit from a grace period that expires on August 2, 2027: that date sets the horizon. In parallel, the CSRD directive brings the footprint of digital technology and AI into large companies’ non-financial reporting, and the ISO 42001 standard integrates sustainability into the AI management system.

The French context adds an urgency of its own. According to ADEME, digital technology already accounts for 4.4% of the national carbon footprint, and data centers make up nearly half of it. The decisive point: these figures do not yet reflect the rise of generative AI. The trajectory to be documented in 2027 will therefore start from an already sharply rising baseline.

What this changes for a CIO in 2027

Four decisions, in this order.

1. Measure before optimizing. Attach to the existing AI FinOps a dual-metric dashboard: dollars and grams of CO₂ per outcome. The trigger is not the pilot, but the move into production at scale, because that is where the variable becomes significant. You cannot manage what you do not measure.

2. Route by task and by model. Make model choice an explicit, governed decision, not a default reflex toward the most powerful option. Model cascading and specialized models deliver the same service level at a fraction of the cost and footprint.

3. Govern at the right level. Bring the energy variable into AI budget decisions on the same footing as the dollar, with cost and carbon checkpoints ahead of decisions. The three milestones that shape the agenda: entering the CSRD scope, the August 2, 2027 EU AI Act deadline, and an ISO 42001 initiative.

4. Optimize and build on it. Lean prompts, caching, and right-sized context windows reduce tokens, cost, and energy in a single move. And if the price of the frontier token keeps rising under the compute crunch, the shift toward open-weight or self-hosted models will gain relevance, both financial and environmental.

In one sentence

AI applied to IT has proven its ROI; the next budget battle will be fought over cost per outcome, and that cost is only complete when it is also expressed in energy and carbon. By 2027, this will no longer be a governance option, but a line on the books.

Run the diagnostic before 2027 makes it mandatory. Does your organization already know how much each outcome its AI produces costs, in dollars and in kilowatt-hours? Fruggr brings the financial cost and the environmental footprint of your AI usage together in a single dashboard, so you can manage the true cost per outcome. 

https://pre-prod.fruggr.io/en/