when your GPU bill has a carbon footprint the sustainability team doesn't know about because they've never met the AI team, and now both teams have to explain their numbers to the same auditor.
Corporate Sustainability Reporting Directive (EU) 2022/2464. Requires large companies and listed SMEs to report on sustainability matters using the European Sustainability Reporting Standards (ESRS). For AI governance, the CSRD creates disclosure obligations for AI-related compute carbon emissions under ESRS E1 (Climate Change), including Scope 2 emissions from electricity consumption and Scope 3 emissions from cloud providers and hardware supply chains. Application: large companies from financial year 2024; listed SMEs from 2026. Reports are subject to mandatory limited assurance by statutory auditors.
Why It Matters
AI is one of the most energy-intensive technologies ever deployed at scale, and most organisations have no idea how much carbon their AI systems produce. Training a single large language model can emit more CO2 than the lifetime emissions of five cars. Daily inference across an enterprise AI portfolio can consume more electricity than the office building the team works in. These emissions are real, growing exponentially, and now must be reported.
The CSRD creates the reporting obligation. The ESRS E1 standard creates the measurement requirement. Together, they mean that organisations can no longer treat AI compute as an IT cost centre invisible to sustainability reporting. The sustainability team needs data from the AI team. The AI team needs to measure things they have never measured. The finance team needs to allocate carbon costs to business units. None of these teams have ever collaborated on this question before.
The gap is not conceptual. It is organisational. Sustainability teams track Scope 1, 2, and 3 emissions from traditional operations. AI teams track model performance, latency, and cloud spend. The carbon footprint of a training run sits in the gap between these two worlds. The CSRD closes that gap by making it the auditor’s problem — which makes it everyone’s problem.
The Stress Test
Your statutory auditor requests limited assurance on your CSRD sustainability report. They ask for the carbon emissions attributable to your AI operations. You provide your total cloud spend. The auditor asks for the breakdown by workload type — training, inference, data processing — and the corresponding energy consumption and carbon intensity. You do not have it. Your cloud provider reports aggregate emissions for your account, not per-workload emissions.
The auditor notes that your sustainability report claims a 12% reduction in Scope 2 emissions. Your AI team’s cloud compute consumption increased 200% in the same period. The numbers cannot both be correct unless the AI compute emissions were never included in the Scope 2 baseline. They were not. Your sustainability report is materially incomplete. The auditor cannot provide assurance on numbers that omit a growing emissions category.
In the Wild
In 2024, Google reported that its greenhouse gas emissions had increased 48% since 2019, attributing the rise primarily to data centre energy consumption driven by AI workloads. Microsoft reported a 29% increase in Scope 3 emissions, acknowledging that its AI infrastructure build-out was a significant driver. Both companies had previously committed to net-zero or carbon-negative targets. The disclosures revealed a structural tension: aggressive AI deployment targets and climate reduction targets cannot both be met without fundamental changes to compute efficiency or energy sourcing.
The companies that pioneered corporate climate commitments are now the companies whose AI ambitions make those commitments hardest to keep. Transparency did not solve the problem. It exposed it.
Research published in 2025 projected that GPU manufacturing emissions alone could reach 21.6 million metric tons of CO2 by 2030 — a 12-fold increase from current levels. The study found that the embodied carbon of a single high-end AI GPU — from raw material extraction through fabrication — is equivalent to approximately 2.5 tonnes of CO2, or roughly driving a conventional car for 6,000 miles. When operational emissions over a three-year lifecycle are added, the total per-GPU carbon footprint approaches the equivalent of driving 100,000 miles. Most CSRD sustainability reports do not include GPU embodied carbon because it falls under Scope 3 Category 2 (capital goods), which many organisations have not yet measured for IT hardware.
The GPU you bought to run your AI models has a carbon footprint before it processes its first token. If it is not in your Scope 3 reporting, your sustainability report is missing one of the fastest-growing emission categories in your entire value chain.
The European Financial Reporting Advisory Group (EFRAG) published implementation guidance on ESRS E1 (Climate Change) specifically addressing technology sector challenges. The guidance clarified that AI compute emissions are within scope of ESRS E1 reporting, that cloud provider emissions must be allocated to customers on a usage basis for Scope 3 reporting, and that organisations must disclose material changes in emission profiles — including those driven by new AI deployments. The guidance noted that many technology companies had treated cloud compute as a “Scope 3 data gap” that would be filled later. EFRAG stated that the materiality of AI-driven compute growth means this gap can no longer be deferred.
EFRAG closed the deferral option. AI compute emissions are material, they are in scope, and “we will report next year” is no longer an acceptable answer.
How to Govern It
CSRD compliance for AI requires connecting two worlds that have never talked to each other: AI operations and sustainability reporting.
Within the AI Control Index, CSRD governance operates across the FinOps shield and the Strategy layer:
- FinOps (S5) — Compute consumption tracking per AI workload, per model, per lifecycle stage (training vs. inference). FinOps translates cloud spend into resource consumption, which translates into energy consumption, which translates into carbon. Without FinOps, there is no measurement. Without measurement, there is no reporting.
- Strategy (L1) — AI system inventory enriched with compute intensity classifications. The inventory links each AI system to its resource consumption profile, enabling portfolio-level carbon footprint estimation.
- GRC (S1) — Evidence Factory captures compute consumption data, carbon intensity calculations, and sustainability reporting artifacts. The same evidence discipline that governs AI Act compliance governs CSRD compliance.
- Observability (S4) — Runtime monitoring of compute resource consumption feeds the FinOps and sustainability reporting pipelines. Post-market monitoring includes environmental impact as a governance dimension.
- Risk Appetite — Board-level articulation of acceptable AI carbon intensity per unit of business value, enabling trade-off decisions between model capability and environmental impact.
When It’s Relevant
All large companies and listed SMEs within CSRD scope that deploy, train, or operate AI systems. The CSRD’s phased application means large companies are already reporting from financial year 2024, with listed SMEs following from 2026.
CSRD-AI urgency is highest when:
- Your organisation is rapidly scaling AI compute without corresponding sustainability measurement
- Your sustainability team and AI team have no shared metrics or reporting processes
- Your cloud compute emissions are not allocated to business units or AI workloads
- Your sustainability report does not include AI-related compute in its emissions baseline
- Your organisation has made public climate commitments while aggressively expanding AI capabilities
Related Terms
References
- [1] European Parliament and Council of the European Union (2022) Directive (EU) 2022/2464 amending Regulation (EU) No 537/2014, Directive 2004/109/EC, Directive 2006/43/EC and Directive 2013/34/EU, as regards corporate sustainability reporting. Official Journal of the European Union, L 322.
- [2] European Financial Reporting Advisory Group (2023) European Sustainability Reporting Standards: ESRS E1 Climate Change. EFRAG, Brussels.
- [3] International Energy Agency (2024) Electricity 2024: Analysis and Forecast to 2026. IEA, Paris.
- [4] Google (2024) 2024 Environmental Report. Alphabet Inc., Mountain View, CA.
- [5] Microsoft (2024) 2024 Environmental Sustainability Report. Microsoft Corporation, Redmond, WA.
- [6] Luccioni, A.S., Viguier, S. and Ligozat, A.-L. (2023) ‘Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model’, Journal of Machine Learning Research, 24(253), pp. 1–15.
- [7] Gupta, U., Kim, Y.G., Lee, S., Tse, J., Lee, H.-H.S., Wei, G.-Y., Brooks, D. and Wu, C.-J. (2022) ‘Chasing Carbon: The Elusive Environmental Footprint of Computing’, IEEE Micro, 42(4), pp. 37–47. doi: 10.1109/MM.2022.3163226.
AI Control Index v6.0 · Glossary · June 2026 · i-DEPOT 158508 (BOIP) · CC BY-NC-ND 4.0
By Jeroen Janssen, Apparens