CSRD and AI Carbon Emissions: What 50,000 EU Enterprises Are Required to Report

In 2019, Emma Strubell and colleagues at the University of Massachusetts published a paper that made climate researchers uncomfortable. Training a large transformer model, they calculated, could emit as much CO₂ as five cars over their entire operational lifetimes (Strubell et al., ACL 2019). The paper arrived before ChatGPT, before the mass enterprise AI rollout, and before anyone outside ML research was thinking seriously about AI's environmental cost.

Six years later, the question of AI carbon emissions is no longer academic. The International Energy Agency projects that global data centre electricity consumption will reach 945 TWh by 2030 — more than double the 415 TWh consumed in 2024, equivalent to Japan's entire annual electricity use (IEA, January 2025). AI workloads are driving the majority of that growth, with AI-specific servers consuming electricity at 30% annual growth rates compared to 9% for conventional servers.

For European enterprises, this trajectory intersects directly with binding law. CSRD requires disclosure of Scope 3 greenhouse gas emissions — which includes the AI tools your employees use every day. The question for compliance teams is no longer whether to report AI carbon. It is whether you can get the data to report it accurately.

What Does CSRD Actually Require for AI Emissions?

The Corporate Sustainability Reporting Directive (Directive 2022/2464/EU) replaced the Non-Financial Reporting Directive and extended mandatory sustainability reporting to approximately 50,000 companies across the EU, representing roughly 75% of all EU corporate turnover. Coverage spans all large EU companies (250+ employees and €40 million or more in turnover), all EU-listed companies, and non-EU undertakings generating €150 million or more in EU revenues.

Climate change disclosures are governed by ESRS E1, which covers Scope 1, Scope 2, and Scope 3 greenhouse gas emissions across all seven GHGs. Scope 3, Category 11 covers emissions from the use of products and services provided to the organisation — which is where AI tools sit. When your employees use a cloud-based AI platform to draft documents, analyse data, summarise meetings, or generate code, the energy consumed by the data centres running those models is a Scope 3 Category 11 emission your CSRD filing must account for.

Which Companies Must Report, and When?

For companies filing FY2025 reports in 2026, Scope 3 emissions — including AI tool usage — must appear in this year's disclosure. The data collection window for FY2025 has already closed; what is left is calculation, documentation, and filing.

How Much CO₂ Does Enterprise AI Actually Produce?

Understanding what needs to be reported requires understanding what AI actually costs in carbon terms. Sasha Luccioni and colleagues at Hugging Face have been among the most rigorous researchers working on AI emissions measurement. Their published work found that task-specific models use roughly 30 times less energy than general-purpose generative models for equivalent tasks (Luccioni, Hugging Face, 2024) — a variance that makes model selection a genuine sustainability decision, not just a performance or cost one.

A 2025 benchmarking study on arXiv (How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference) estimated GPT-4o's projected annual emissions at 138,000–163,000 tonnes of CO₂e, with Claude 3 Opus consuming approximately 4.05 Wh per request and producing around 1.80g CO₂ per query under typical US grid conditions (arXiv, May 2025). Goldman Sachs Research puts the macro picture starkly: data centre power demand is expected to increase 160–165% by 2030 relative to 2023 levels, with AI accounting for more than 25% of all data centre power usage by decade's end (Goldman Sachs, June 2024).

The variance between models is substantial enough to matter for ESRS E1 reporting. A company running 250 employees through heavy-use AI workflows on a large frontier model will produce meaningfully different emissions than the same company routing equivalent tasks through a lighter, more efficient model. An 8× efficiency gap between the most and least carbon-intensive commonly deployed models means the model routing policy your IT department chooses is also a sustainability decision with direct CSRD implications.

Why AI Vendors Won't Give You the Data You Need

CSRD requires you to report your AI carbon emissions, but the major AI platforms provide almost none of the data needed to do it accurately. An assessment published by carbon accounting firm CNaught found that 10 of the 13 AI companies they evaluated — including OpenAI, Anthropic, Google, xAI, and DeepSeek — disclosed zero environmental metrics to customers (CNaught, 2025). OpenAI publishes no per-query carbon data. Microsoft Azure makes broad carbon neutrality commitments but publishes no granular AI workload emissions data by customer or by model. Anthropic does not publish per-token carbon figures for its Claude models.

This is not accidental. From a commercial standpoint, there is no incentive for an AI vendor to make its carbon footprint visible and comparable at the query level. Doing so would allow customers to run the calculation that Luccioni's research implies: switching from a large frontier model to a lighter, more efficient model for routine tasks could reduce AI-related carbon emissions by a factor of several times. For the vendors selling the frontier models, that comparison works directly against them.

The result is a structural data gap at the heart of CSRD compliance for AI. The regulation requires disclosure. The vendors holding the underlying data decline to provide it. This leaves sustainability teams working from academic estimates, grid intensity averages, and approximated token counts — or engaging consultants to build bespoke estimation models at significant cost.

How to Calculate AI Emissions Without Vendor Data

Even without vendor cooperation, a defensible ESRS E1 disclosure for AI tool usage can be constructed from four inputs. The methodology is reasonable and auditor-accepted where it is documented, consistently applied, and draws on published research sources.

1. Token counts by model. The primary driver of AI energy consumption at the inference stage is the number of tokens processed — the volume of text going in and coming out of the model. Token counts must be tracked by model because energy-per-token varies substantially between GPT-4o, GPT-4o-mini, Claude Opus, Claude Haiku, and other models in common enterprise use. API billing records are the most accessible data source where the AI platform itself does not provide usage exports.
2. Energy per token by model. Published research provides estimates — Luccioni's Hugging Face work and the arXiv 2025 benchmarking studies are the most frequently cited academic sources. These figures are approximations, as vendors do not publish precise hardware configurations or power draw per inference request. They are nonetheless defensible for CSRD reporting purposes where precision is bounded by available data.
3. Grid carbon intensity. Energy consumed by a data centre in a high-renewable-energy market produces a different CO₂e figure than the same energy from a high-fossil-fuel grid. US Azure data centres operate at approximately 367g CO₂ per kWh; European data centres typically range lower depending on grid mix. The IEA publishes country-level grid intensity figures that provide a reasonable basis for Scope 3 estimates when vendor-specific data is unavailable.
4. Attribution to reporting entity. CSRD requires emissions to be attributed to your organisation. This means knowing how many tokens your employees sent to which models over the reporting period — data that, if your AI platform does not track it, must be reconstructed from API billing records, usage dashboards, or estimates based on seat count and average usage patterns.

Running this calculation manually across multiple AI tools, multiple teams, and a full reporting year is feasible but slow. For a mid-sized enterprise using four or five AI platforms, a manual CSRD calculation typically takes three to four weeks of sustainability team time and involves significant uncertainty at each step. The practical step-by-step guide to CSRD AI emissions reporting walks through this calculation in full, including how to handle multiple platforms, document the methodology, and produce the ESRS E1 format output that sustainability auditors expect.

What Are the Penalties for Incomplete CSRD Reporting?

CSRD itself does not specify fines — enforcement is left to EU member states. The divergence in national penalty regimes is significant. Germany has set penalties up to €10 million or 5% of annual turnover for material non-compliance (CoreFiling, August 2024). France has established fines up to €75,000, with the possibility of prison sentences for corporate directors who obstruct audits. Novata's analysis of CSRD enforcement notes that criminal liability provisions — including prison sentences up to five years for obstructing sustainability auditors — exist in several jurisdictions.

Beyond direct penalty exposure, CSRD non-compliance creates a secondary risk that is arguably larger than any individual fine: ESG rating consequences. MSCI, Sustainalytics, and ISS ESG all incorporate CSRD disclosure quality into their ratings methodologies. Institutional investors operating under the EU Taxonomy or SFDR require portfolio companies to produce verified CSRD data. A disclosure that omits or materially underestimates a significant and growing emission source — AI tool usage — creates rating exposure that compounds over reporting cycles.

The FY2025 reporting cycle, with disclosures due in 2026, is likely to be the first year where CSRD auditors specifically scrutinise AI emissions methodology. Companies that have collected no usage data and are constructing retrospective estimates from memory will face greater audit challenge than those with documented tracking processes in place.

Does Model Selection Affect Your CSRD Exposure?

One implication of tracking AI carbon at the model level is that it makes a previously invisible decision visible: which model you route employees' work to. This is not a question that most enterprise AI policies currently address, but it becomes material when ESRS E1 reporting is in scope. For routine tasks — summarising documents, drafting emails, answering internal knowledge questions — routing through a lighter model rather than a frontier model can reduce per-interaction emissions by 3–8× without material quality loss for those task types.

The model selection decision is simultaneously a cost decision (smaller models are cheaper per token), a performance decision (appropriate model for task), and a sustainability decision with CSRD implications. Enterprises building AI governance frameworks now have an additional optimisation axis. The discussion of multi-model enterprise AI and BYOK architecture covers how routing policies across multiple models can serve cost, performance, and governance goals — and CSRD reporting adds a fourth dimension to that optimisation.

What the Next CSRD Reporting Cycle Looks Like

For companies whose FY2025 reports are due in 2026, the window to collect usage data is already closed — the year has elapsed. The options are to estimate retrospectively from available billing records, or to disclose in the methodology section that AI-specific tracking has now been implemented for future periods. Both are defensible, but the former requires careful documentation and the latter is an explicit disclosure of a known gap.

From FY2026 onward, the standard for what a reasonable CSRD disclosure on AI emissions looks like will be set by early filers. Companies with automated tracking will produce precise, auditable data. Companies relying on estimates will need increasingly rigorous documentation to satisfy auditors comparing their methodologies against peers with better data. The IEA's projection that AI energy consumption will double by 2030 means this disclosure item will grow in materiality alongside the technology's adoption.