CSRD and AI: How European Companies Should Report AI Emissions Under ESRS E1

The EU Corporate Sustainability Reporting Directive (CSRD) is the most comprehensive climate disclosure framework in the world. Starting with reports published in 2025 (for the 2024 fiscal year), and expanding to all large companies by 2026, CSRD requires detailed Scope 1, 2, and 3 emissions disclosure under the European Sustainability Reporting Standards (ESRS) - specifically ESRS E1 on Climate Change.

For European companies that rely on cloud AI services, this creates a new reporting obligation: accounting for the carbon and water footprint of your AI compute. Unlike the SEC rules, CSRD's Scope 3 reporting is mandatory, not voluntary.

What CSRD Requires for AI Compute

Under ESRS E1, companies must disclose:

• Scope 3 Category 1 (Purchased Goods and Services): AI API calls from OpenAI, Anthropic, Google, or Azure fall here. The emissions from running your prompts on their infrastructure are your Scope 3.
• Scope 3 Category 11 (Use of Sold Products): If your product embeds AI features, the downstream compute emissions from your customers using those features may need reporting.
• Water consumption: ESRS E3 (Water and Marine Resources) requires disclosure of water usage across the value chain. Data center water consumption is increasingly scrutinized.

The Double Materiality Challenge

CSRD introduces "double materiality" - you must report on how climate change affects your business (financial materiality) and how your business affects the climate (impact materiality). AI compute is relevant to both:

• Impact materiality: Your AI usage directly causes emissions. If you run 10 million tokens per day through GPT-4o on Virginia servers, that is roughly 5.4 kg of CO2 per day, or about 2 tonnes per year - from a single application.
• Financial materiality: Carbon pricing is coming. The EU ETS already prices carbon at around 60-80 EUR/tonne. As AI compute emissions become reportable, they become priceable. Companies with unoptimized AI usage face direct cost exposure.

Practical Reporting Guide for AI Emissions

Step 1: Gather Your AI Usage Data

Export token usage from your providers. Most enterprise agreements with OpenAI, Azure, or Google Cloud include usage dashboards. You need: model name, total tokens (ideally split into input and output), and the cloud region.

Step 2: Select the Right Emission Factors

Generic global averages are not CSRD-compliant. You need region-specific emission factors. European data centers vary enormously in carbon intensity:

• Sweden: 20 gCO2/kWh (hydro + wind) - nearly zero-carbon
• France: 55 gCO2/kWh (nuclear-dominated)
• Ireland: 300 gCO2/kWh (gas-dominated, major data center hub)
• Germany: 350 gCO2/kWh (coal transition)
• Netherlands: 330 gCO2/kWh (gas + offshore wind mix)

The AI Impact Calculator API v2 provides all 14 region-specific emission factors with WRI Aqueduct water stress classification and WUE data, sourced from EIA and Ember Climate.

Step 3: Calculate and Document

Use the API to convert usage into emissions with full traceability:

POST https://aiimpactcalculator.com/api/v2/calculate
{
  "model": "claude-3.5-sonnet",
  "input_tokens": 2000000,
  "output_tokens": 800000,
  "region": "eu-ireland"
}

// Response:
// energy_kwh: 8.48
// carbon_gco2e: 2,544g (2.5 kg)
// water.liters: 7.63L
// water.stress_level: "low"
// data_sources: ["Ember Climate (2025)", "WRI Aqueduct"]

The response includes data source citations and methodology references - exactly what CSRD auditors expect to see in your supporting documentation.

Water: The Overlooked CSRD Dimension

ESRS E3 requires companies to disclose water consumption across their value chain. Most companies have never considered the water footprint of their cloud services. But data centers are increasingly significant water consumers:

• Microsoft reported a 34% increase in water consumption due to AI scaling
• Google's data centers consumed 5.6 billion gallons of water in 2023
• A single conversation with GPT-4 can consume approximately 500ml of water in evaporative cooling

Our API uniquely provides region-specific WUE (Water Usage Effectiveness) data, so you can report water impact with the same precision as carbon.

Optimization: Reduce Before You Report

CSRD is not just about disclosure - it requires companies to set emission reduction targets. For AI compute, the most effective strategies are:

• Region selection: Route AI calls to low-carbon regions (Sweden, France) where possible. This can reduce emissions by 15x without any other changes.
• Model right-sizing: Use smaller models (Claude Haiku, GPT-3.5) for routine tasks. An 8-10x reduction in energy per call.
• Output token optimization: Output tokens use 3x more energy than input. Shorter, more focused prompts reduce the output token count and the footprint.
• Caching: Cache repeated AI responses. A cached response has zero marginal emissions.