AI systems are significant energy consumers — from training large foundation models to running inference at scale. As CSRD becomes mandatory for large companies and AI usage proliferates, the environmental impact of AI becomes a material sustainability topic. This article explains what companies must disclose about their AI-related environmental footprint under CSRD and ESRS E1.
Why AI's Environmental Impact Is a CSRD Issue
ESRS E1 requires reporting on greenhouse gas emissions, energy consumption, and climate transition plans. For companies that:
- Train AI models at scale (foundation model providers, research labs, large enterprises with custom models)
- Deploy AI inference at scale (AI SaaS companies, large users of AI APIs)
- Operate data centres that host AI workloads
...the energy used for AI is a material component of their overall Scope 1, 2, and 3 emissions.
Key drivers of AI's environmental footprint:
Training: Training large language models is energy-intensive. GPT-4 training has been estimated at 50+ GWh of energy. Custom foundation models trained by enterprises involve similar orders of magnitude. Training runs are typically infrequent but have a large one-time carbon cost.
Inference: Ongoing inference (serving AI responses to users) consumes energy continuously. For high-usage AI products, inference can exceed training energy consumption over the product's lifetime.
Cooling: AI chips (GPUs, TPUs, custom silicon) generate significant heat, requiring additional cooling energy in data centres.
What CSRD Requires for AI Energy Reporting
ESRS E1 — Scope 2 and Scope 3 Emissions
Scope 2 (purchased electricity): The electricity used to train and run AI systems is a Scope 2 emissions source. Companies with significant AI workloads must include this energy in their Scope 2 disclosure.
If using cloud AI services (AWS SageMaker, Azure AI, Google Cloud AI), the energy is consumed by the cloud provider's data centres. This is your Scope 3 Category 1 (purchased goods and services) or Scope 3 Category 6 (business travel and cloud services — classification varies by methodology).
Scope 3: Companies using third-party AI services (API calls to OpenAI, Anthropic, Google, etc.) have an upstream Scope 3 footprint. While Scope 3 calculation for API usage is still developing methodologically, forward-looking CSRD reporters should begin building estimates.
Energy Disclosure
ESRS E1 requires total energy consumption — this must include energy used by AI workloads in own operations. For companies running on-premises AI infrastructure, energy metering at the GPU cluster level is needed for accurate disclosure.
Transition Plan
Companies with significant AI-related energy consumption should address this in their climate transition plan:
- Current energy mix for AI workloads
- Targets for renewable energy coverage
- Efficiency improvement targets (performance per watt, model compression, more efficient architectures)
- Timeline for achieving net-zero for AI operations
Emerging Disclosure Expectations Beyond ESRS E1
While ESRS E1 is the core framework, several other disclosure areas are emerging:
Water consumption (ESRS E3): AI chips require effective cooling, and some cooling technologies (evaporative cooling towers, liquid cooling) use significant water. Companies in water-stressed regions with large AI footprints may find ESRS E3 material.
Hardware supply chain (ESRS E5, S2): AI relies on specialised chips (GPUs, TPUs) with significant embedded carbon in their manufacture — extraction of rare earth materials, semiconductor fabrication. This is a Scope 3 Category 1 footprint and a supply chain sustainability consideration.
Algorithmic impacts (ESRS G1, S4): Beyond environmental footprint, CSRD's governance and social standards are increasingly being read by analysts to require disclosure of AI governance — bias, fairness, and societal impact. This is an emerging area; ESRS does not yet require explicit AI governance disclosure, but narrative reporting under ESRS 2 governance may touch on it.
What AI Companies Should Do Now
Companies That Build AI Products
-
Measure training energy: For any model training run, record the GPU hours, cloud provider, energy consumption, and associated emissions. Tools: CodeCarbon, ML CO2 Impact calculator, cloud provider emissions dashboards.
-
Measure inference energy: Instrument your AI serving infrastructure to measure energy consumption per inference or per unit of compute.
-
Attribute cloud AI energy to Scope 3: For AI workloads on cloud providers, use cloud provider carbon reporting tools (AWS Customer Carbon Footprint Tool, Google Carbon Footprint, Azure Emissions Impact Dashboard) to estimate associated Scope 3 emissions.
-
Disclose in ESRS E1: Include AI energy consumption in total energy and emissions disclosures. Segment AI energy separately in supplementary disclosures where material.
-
Transition plan: Commit to renewable energy for AI workloads — a credible transition plan for a significant AI company should specifically address AI energy consumption.
Companies That Use AI Tools
-
Account for AI vendor Scope 3: API usage of AI services (ChatGPT, Claude, Gemini) is a Scope 3 Category 1 purchase. While precise accounting methodologies are still developing, include an estimate in your Scope 3 disclosure.
-
Ask vendors for carbon data: As CSRD value chain requirements embed, request carbon footprint data from your AI API providers. Some providers (Google, Microsoft) publish data centre energy and emissions data; others are less transparent.
-
Consider efficiency in procurement: Energy efficiency per task is a legitimate sustainability criterion in AI tool procurement — models that produce equivalent outputs with lower compute have a lower emissions footprint.
The Regulatory Outlook
The AI Act's provisions on high-risk AI and GPAI include some energy disclosure requirements — GPAI model providers must report energy consumption as part of technical documentation. As AI regulation and sustainability reporting both mature, integrated disclosure on AI's environmental impact will become a more explicit requirement.