How to Report AI Compute in Scope 3 Under SEC Climate Rules

The SEC finalized its climate-related disclosure rules in March 2024, and large accelerated filers are now preparing their first filings. For companies that use AI - whether through direct API calls, cloud-hosted models, or embedded AI features - there is a new question on the table: how do you report the carbon footprint of your AI compute?

What the SEC Actually Requires

The SEC rules adopted in Release No. 33-11275 focus on climate-related risk and emissions disclosure. While Scope 1 and Scope 2 are mandatory for large filers, Scope 3 reporting is currently voluntary but explicitly encouraged. More importantly, if Scope 3 emissions are material to your business, omitting them creates legal risk.

For companies with significant AI usage - SaaS platforms, fintech, healthtech, legal tech - AI compute is increasingly the largest single contributor to Scope 3 Category 1 (Purchased Goods and Services) emissions. Ignoring it is no longer defensible.

The Three Components of AI Compute Emissions

To report AI emissions accurately, you need to account for three things:

• Energy consumption (kWh): How much electricity each model call uses. This varies dramatically by model - GPT-4o uses roughly 3.5 Wh per 1,000 tokens, while GPT-3.5 Turbo uses 0.5 Wh. Output tokens consume approximately 3x more energy than input tokens due to autoregressive generation.
• Carbon intensity (gCO2e/kWh): The emissions factor of the electricity grid where the data center operates. Virginia (310 gCO2/kWh) produces 15x more carbon per kWh than Sweden (20 gCO2/kWh). This is the biggest variable most companies overlook.
• Water usage (L/kWh): Data centers consume water for cooling. The Water Usage Effectiveness (WUE) ranges from 0.3 L/kWh in cool climates to 3.0+ L/kWh in arid regions. Water is increasingly material for ESG reporting, especially in water-stressed regions like California and the Middle East.

Step-by-Step: Calculating Your AI Scope 3

Here is a practical approach:

1. Inventory Your AI Usage

Pull your token usage from your AI provider dashboards (OpenAI, Anthropic, Google Cloud, Azure). Most providers give monthly token counts broken down by model. If you cannot get exact numbers, estimate from your billing data - cost per 1M tokens is published for every major model.

2. Map to Data Center Regions

Identify which regions your AI calls are routed to. OpenAI and Azure primarily serve US traffic from Virginia data centers. Google Cloud and Anthropic use a mix of US regions. If you are unsure, use your cloud provider's region setting or default to US East (the most common).

3. Run the Calculation

Use the AI Impact Calculator API to convert your usage data into emissions. The v2 API accepts model, token count, and region, and returns energy (kWh), carbon (gCO2e), and water (liters) with full data source citations.

curl -X POST https://aiimpactcalculator.com/api/v2/calculate \
  -H "Authorization: Bearer YOUR_KEY" \
  -d '{"model":"gpt-4o","tokens":5000000,"region":"us-east-virginia"}'

// Response includes:
// energy_kwh: 17.5
// carbon_gco2e: 5,425g (5.4 kg)
// water.liters: 31.5L
// water.stress_level: "high"

4. Aggregate and Report

Sum across all models and regions for your reporting period. Include the methodology note citing EIA State Electricity Profiles and WRI Aqueduct Water Risk Atlas as your data sources. This gives your filing audit-grade traceability.

Why Region Matters More Than Model Choice

Most companies focus on which model they use, but geographic location often has a larger impact on total emissions. The same 1 million GPT-4o tokens produce:

• 1,085g CO2 when served from Virginia (coal + gas grid)
• 70g CO2 when served from Sweden (hydro + wind grid)
• 1,750g CO2 when served from UAE (gas grid + extreme cooling)

That is a 25x difference for the exact same workload. If your cloud provider offers region selection, choosing a low-carbon region is the single most impactful optimization you can make.