Scope 3 emissions and AI: where usage goes, how to count it

For most companies Scope 3 is the largest part of the inventory and the part built on the roughest estimates. This guide covers what the scope is, which of its fifteen categories AI usage falls into, which rules now require it, and why AI is the unusual case where the number can be measured.

The three scopes

The GHG Protocol, the standard almost every reporting rule points to, splits a company’s emissions by where they occur. Scope 1 is direct: the fuel in company vehicles and boilers, the processes in its plants. Scope 2 is the electricity, heat and steam it purchases, emitted at the power plant but counted by the buyer. Scope 3 is everything else the company causes without owning: the goods and services it buys, the travel and commuting of its staff, the transport of its products, their use and their disposal, its investments. The protocol sorts Scope 3 into fifteen categories.

Where AI usage lands

When a company calls a model through OpenAI, Anthropic, Google, AWS or a gateway, the inference runs on the provider’s hardware in the provider’s data centre. The electricity is the provider’s Scope 2 and the hardware is the provider’s Scope 3. For the company doing the calling, the whole thing is a purchased service, which puts it in Category 1. That is the treatment the corporate framework by Bistline and colleagues sets out, and it is the one CrbonFree’s methodology follows: the operational layer corresponds to the provider’s Scope 2 boundary, and the lifecycle layer adds the embodied hardware and training carbon that sit in the provider’s Scope 3, carried as their own line so nothing is blended in silently.

There are two edge cases. A company running models on its own servers has Scope 2 electricity and Category 2 capital goods, not Category 1. A company renting dedicated capacity may find the guidance for Category 8, upstream leased assets, a closer fit. For the ordinary case of paying per token, Category 1 is where the line goes.

Why Scope 3 is hard, and why AI is not

The protocol’s guidance offers four ways to calculate a Category 1 figure. Supplier-specific data uses the supplier’s own emissions for the product or service. Hybrid methods combine supplier data with estimates. Average-data methods multiply a physical quantity by an average factor. Spend-based methods multiply money spent by an emissions-per-dollar factor for the industry. Data quality runs in that order, and most Scope 3 reporting sits at the bottom of it, because for most purchases the only number a company has is the invoice.

AI usage is different in one specific way: the activity data exists. Every provider reports tokens per model through its usage API, and tokens are a physical measure of the work done, so the line can be built with the average-data method on measured quantities and published per-token factors, with the supplier’s own PUE and grid intensity where the supplier publishes them. A spend-based figure for AI fails the first question an auditor asks, because a dollar buys very different numbers of tokens on different models and a model’s energy per token varies by an order of magnitude across tiers.

Is it mandatory?

In the European Union, companies in scope of the Corporate Sustainability Reporting Directive report Scope 3 under the ESRS E1 climate standard. In California, SB 253 requires companies over one billion dollars in revenue doing business in the state to report Scopes 1 and 2 from 2026 and Scope 3 from 2027, in conformance with the GHG Protocol. In the United States federally, the SEC excluded Scope 3 from its 2024 climate rule and in 2026 proposed rescinding the rule. Outside the rules, investor questionnaires, customer supplier surveys and science-based targets all ask for Scope 3, and AI usage is now large enough on most invoices that a company without a figure for it will be asked why.

How tokens become the line

1. Meter the tokens, per model and per provider, from each provider’s usage API, the CLI for coding agents and the browser extension for chat; spend does not count as activity data.
2. Match each model to a tier and apply published per-token factors for input, output and cached tokens. CrbonFree’s methodology v1.2 publishes eight tiers with their factors.
3. Convert energy to CO₂e with the facility’s power usage effectiveness and the grid’s intensity, then add the facility layer and the lifecycle layer, keeping the three available separately.
4. State the uncertainty. The methodology carries a plus or minus 28.3% band from IPCC Tier 1 error propagation, and an auditor will want it.
5. Pin the factor version. A report filed today must not move when the factors improve, so the version is recorded on each period and each receipt.
6. Keep measurement and climate action apart. If verified credits are retired against the AI line, the line is reported unchanged and the retirement is reported alongside it with its serial numbers.

CrbonFree does all six from a provider key, the CLI or the extension, and the sample account shows a year of real traffic worked through this way with no login required. The pricing page has what the audit packs and receipts cost.