Short answer: almost all of it lands in Scope 3, Category 1 — Purchased Goods and Services. If you buy AI as a service instead of running your own GPUs, that's where it sits. This post walks through why, and how to start sorting your AI vendor list into something a Scope 3 inventory can actually use.
The framework, in one paragraph
The GHG Protocol Corporate Value Chain (Scope 3) Standard, published by WRI and WBCSD in 2011 and backed by 2013 Technical Guidance, splits indirect emissions into 15 categories. Eight are upstream, seven are downstream. Category 1 covers everything you buy that gets produced before it reaches you: raw materials, components, software subscriptions, professional services. For most companies, Category 1 is the largest slice of Scope 3. Some estimates put it at 50 to 80 percent of the total. If your organization pays for OpenAI, Azure OpenAI, Copilot seats, or Anthropic's API, that spend is a line item inside Category 1.
Why AI spend lands here and not somewhere else
Two categories get confused with AI spend. Here's the split.
Category 1 (Purchased Goods and Services) covers AI you access as a service. You don't own the servers. You're paying OpenAI, Microsoft, or Anthropic to run inference on their infrastructure and hand you a response. That's a purchased service, full stop.
Category 2 (Capital Goods) only applies if you buy and own the physical hardware, like a company that purchases its own GPU cluster instead of renting compute. Almost no company reading this owns its inference hardware. If you're paying a monthly or usage-based bill to a model provider, you're in Category 1.
One more distinction worth naming: training emissions belong to the model provider's own Scope 1 and 2, not to your inventory. What you're accounting for is the emissions embedded in the service you bought, not the emissions from building the model in the first place.
The three ways to calculate it, ranked by accuracy
The GHG Protocol ranks calculation methods by data quality, not by convenience:
- Supplier-specific data — the vendor gives you an actual emissions figure tied to your usage.
- Activity-based data — you calculate emissions from a physical unit, like tokens processed or kWh consumed, multiplied by a grid carbon intensity.
- Spend-based data — you multiply your dollar spend by an industry-average emissions factor. Lowest data burden, lowest accuracy.
A useful way to think about this for AI specifically comes from a 2026 methodology paper out of SOMA AI (Llopis, "Accounting for AI Inference in Corporate GHG Inventories"). It proposes a four-tier version of the same hierarchy, built specifically for AI services:
| Tier | What it is | Data required | Uncertainty |
|---|---|---|---|
| Tier 3 | Provider-certified carbon report | Vendor-audited figure | Low |
| Tier 2a | Token-based, exact | Token counts from API billing logs | ±50% |
| Tier 2b | Token-based, estimated | Seat or message counts | ±60% |
| Tier 1 | Spend-based EEIO | Annual spend only | Can overestimate 10–40x vs. physical methods |
That last line matters. The same paper found that a generic ICT-sector spend factor, applied blindly to AI spend, can overstate emissions by an order of magnitude compared to token-based estimates. Spend-based is a legitimate starting point. It's not where you want to stay.
Classifying spend across your actual AI vendors
Here's where this gets concrete. Your AI vendor list probably isn't one line item, it's several, and each one gives you different data to work with.
| Vendor / Product | Billing model | Best available tier | What to pull |
|---|---|---|---|
| OpenAI API (direct) | Usage-based, token metered | Tier 2a | Token counts and model used, from OpenAI's usage dashboard or billing export |
| Azure OpenAI Service | Usage-based, billed through Azure | Tier 2a/3 | Microsoft's Emissions Impact Dashboard, which already calculates Scope 3 Categories 1, 2, 4, 5, 9, and 12 for Azure customers using a third-party validated methodology |
| Microsoft 365 Copilot | Seat-based subscription | Tier 2b | Licensed seat count and tier, since Copilot doesn't expose token-level billing to customers |
| Anthropic API (Claude, direct) | Usage-based, token metered | Tier 2a for the token data, Tier 1 for cross-checking | Token counts from the Anthropic console. Note Anthropic has not published an audited Scope 1/2/3 report as of this writing, so there's no vendor-side figure to validate against |
| Claude via AWS Bedrock or Google Vertex | Usage-based, billed through the cloud platform | Tier 2a | Bedrock or Vertex usage export, same as direct API |
| AI features bundled into unrelated SaaS (a CRM's AI assistant, for example) | Opaque, no separable line item | Tier 1 | Vendor category and total contract spend, since usage isn't broken out |
Two things jump out once you build this table for your own vendor list. First, the providers with usage-based billing (OpenAI, Anthropic, Bedrock, Vertex) actually hand you activity data for free, it's sitting in your existing invoices. Second, seat-based tools like Copilot are the hardest to get past Tier 1 or Tier 2b, because the provider doesn't expose per-query usage to you at all.
What data to collect, by source
Before you calculate anything, build the inventory. You need:
- A complete AI vendor list. Not just the AI tools IT approved. Pull actual spend from accounts payable, because AI subscriptions get expensed on corporate cards and departmental budgets outside procurement all the time.
- Billing model per vendor. Token-metered, seat-based, or bundled into a larger contract. This determines which tier you can reach.
- Token or usage exports where available. OpenAI, Anthropic, and Bedrock/Vertex all expose this in billing consoles or admin dashboards.
- Seat counts and license tiers for subscription products. Copilot, ChatGPT Enterprise, and similar tools.
- Regional deployment info if you have it. Grid carbon intensity varies enormously by region, and some providers let you see or choose data center region.
- A documented methodology note per vendor. Under CSRD and similar frameworks, you can't quietly switch methods year to year. Write down which tier you used for each vendor and why.
One data-quality trap worth flagging: shadow AI. Free-tier ChatGPT, Claude, or Gemini accounts that employees sign up for on their own never hit a procurement statement. A spend-based Category 1 estimate built only from AP data will miss this entirely and report a number smaller than actual usage. A short internal survey or SSO audit closes most of that gap.
The context that makes this material
This isn't a rounding error forever. The IEA's 2026 update to its Energy and AI report projects data center electricity demand roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, with AI-optimized data centers growing even faster. AI servers already accounted for 24 percent of server electricity demand in 2024. Whatever your AI Category 1 number looks like today, the trend line points up, not down. That's the case for building a real methodology now instead of a placeholder spend-based figure you'll have to redo in two years.
One methodology note before you calculate
If you graduate a vendor from spend-based to token-based data, pull that vendor's spend out of your spend-based calculation entirely. Otherwise you've counted the same purchase twice, once as activity data and once as spend data. This is a known trap in Scope 3 accounting generally, not unique to AI, but it's easy to miss when you're building out a new vendor category for the first time.
That's the classification layer. The next piece in this series, How to Calculate Your AI Emissions, walks through the actual math: which emission factors to use for each tier, how to convert token counts into kWh, and a worked example using a real vendor mix. If you're building your inventory now, start with the table above. Knowing which tier each vendor can reach is most of the work.
Sources
- GHG Protocol, Corporate Value Chain (Scope 3) Accounting and Reporting Standard (WRI/WBCSD, 2011) and Scope 3 Calculation Guidance (2013)
- Llopis, G., "Accounting for AI Inference in Corporate GHG Inventories: A Four-Tier Methodology for Scope 3 Category 1 Reporting," SOMA AI preprint, June 2026 (arXiv:2606.10660)
- EPA, Supply Chain Greenhouse Gas Emission Factors v1.3 (NAICS 518210, 541511, 541512, 541519)
- Microsoft Learn, "Azure Emissions Calculation Methodology"
- IEA, Energy and AI (2026 update), executive summary
- Sam Altman, "The Gentle Singularity" (OpenAI blog)
- Normative, "How to Calculate Scope 3 Emissions" (double-counting guidance)