Does the CSRD really require reporting the environmental impact of AI workloads?

Yes, for companies within its scope, which Omnibus I (Directive (EU) 2026/470) narrowed to companies above 1,000 employees and EUR 450 million in net turnover, from financial year 2027: the CSRD requires verifiable Scope 1, 2 and 3 emissions reporting, which includes IT systems and AI workloads hosted in the cloud or on-premise.

Does a smaller model really consume less energy than a large one?

Yes: for the same task, an SLM (Small Language Model) requires less compute both in training and inference than a large model, with a proportionally lower energy impact.

What is Green IT and digital sustainability?

The practices for reducing the energy footprint of IT infrastructure and AI workloads, now also a CSRD reporting obligation.

Green IT and digital sustainability describe the set of practices a company uses to reduce the environmental impact of its technology infrastructure: data center energy efficiency, choosing regions and providers powered by renewable energy, and optimizing workloads to reach the same result while consuming less compute. This is not a new topic, but the arrival of generative AI at scale has made it urgent: training and running large models carries a measurable energy cost, and that cost is growing faster than any other IT segment. Until recently, Green IT was mostly a matter of corporate image; today it is also a regulatory constraint, with reporting obligations that make the energy footprint of digital systems a figure as scrutinized as any other line in the balance sheet. It touches the choice of cloud region and provider just as much as architectural decisions about which model to use and how much compute it will take to train or run it in production.

Why it is a data governance topic, not just facility management

According to Gartner, global data center electricity demand will grow 26% in 2026, reaching about 132 gigawatts from 104 in 2025, with AI-optimized data centers accounting for 31% of total consumption. The International Energy Agency confirms the same trajectory: data center electricity use surged sharply in 2025, with power availability becoming the real bottleneck for AI compute capacity, more so than chip supply. In parallel, the CSRD (Corporate Sustainability Reporting Directive), as amended by Omnibus I (Directive (EU) 2026/470), requires companies above 1,000 employees and EUR 450 million in net turnover, from financial year 2027, to report verifiable Scope 1, 2 and 3 emissions, including IT systems and AI workloads hosted in the cloud or on-premise: no longer a voluntary initiative for a press release, but a figure that goes into financial statements and gets audited.

The link to technical choices

Architecture matters as much as good intentions. A smaller model (SLM) that solves the same task as a huge one consumes a fraction of the energy, both in training and inference. Choosing to fine-tune an existing model instead of training one from scratch has a direct, measurable impact on consumption. Cloud repatriation decisions, that is moving workloads back from public cloud to on-premise, are not just about cost or data sovereignty either: they also change the energy footprint, in one direction or the other, depending on where and how the energy used is generated.

Why it matters for a business

For a company subject to the CSRD, Green IT stops being a branding topic and becomes a reporting requirement with real consequences for audits and credibility with investors and customers. For anyone designing AI systems, it means the choice between a huge model and a smaller one, or between training from scratch and fine-tuning, is not only a performance and cost decision: it is also one that ends up in the company's sustainability report.