The hidden energy cost of AI: why efficiency is becoming a selection criterion for SMBs

Global data center electricity consumption could double by 2026, according to the International Energy Agency's Electricity 2024 report, on a trajectory that could reach 530 TWh for data centers alone by 2030. That figure remains, for now, a macroeconomic and environmental topic, far from the daily concerns of a ten-to-fifty-person SMB. Yet it has a concrete translation, with a lag of months to years, into the price of the AI tools that same SMB uses every day.

Electricity can account for up to 54% of a data center's operating costs. Over a ten-year operating period, electricity costs can neutralize, or even exceed, the effect of a discount on hardware itself. AI vendors hosting their models in these facilities eventually pass this economic reality on in their pricing. An SMB building heavy usage around an oversized model for the intended task pays a price today that does not yet fully reflect this trend, and may feel it more in the coming years.

France sits in a fairly favorable position on this specific front. Non-domestic electricity prices there fell 14.1% year over year, versus 5.4% on average across the European Union, making it an attractive destination for data center projects and potentially limiting, within the country, the scale of the pricing pass-through seen elsewhere.

This reality invites a simple habit, rarely applied by SMBs, sizing the AI model to the task rather than systematically defaulting to the most powerful model available. For a simple task, rewording an email, summarizing a short document, a fast model like GPT-5.5 Instant or Claude Haiku 4.5 produces an equivalent result to a heavier model like GPT-5.5 or Claude Opus 4.7, at markedly lower cost and latency. Reserving the most powerful models for tasks that genuinely need them, complex analysis, multi-step reasoning, is not just a gesture toward sustainability, it is an immediate budget optimization.

This habit matters more for high-volume uses than occasional ones. An employee using AI a few times a day for varied tasks has a marginal impact, whatever model is chosen. A conversational agent handling several thousand customer requests a day, on the other hand, sees the model choice flow directly and cumulatively into the monthly bill. It is precisely on these large-scale automated uses that the energy criterion deserves a place in tool selection, alongside price, security, and regulatory compliance.

No binding obligation currently applies to SMB users of AI regarding energy transparency for their uses, European discussions on the topic remain, in 2026, at a preliminary stage. But getting ahead of this criterion, before it becomes a regulatory requirement or a vendor's sales pitch, means approaching tool choices with a head start rather than under the pressure of a new obligation.

This logic ties into, without replacing, a broader sustainability concern that several French SMBs already build into their CSR policy. Documenting that an AI tool choice took energy footprint into account, even simply, becomes credible messaging with clients or partners increasingly attentive to this kind of criterion, without requiring a complex reporting exercise.

This is what the Framing step of the IMPACT method should incorporate today, model sizing to the task as a selection criterion, on par with price or security. This is not a topic that will dictate your AI choices in 2026, but it is a criterion that is starting to matter, and will matter more as usage scales up. A five-business-day diagnostic factors this criterion into your tool choices.