What is carbon accounting?

Carbon accounting measures the greenhouse gas emissions an organization causes, across its own operations (Scope 1), purchased energy (Scope 2) and its value chain (Scope 3). It's also called GHG accounting.

What is the spend-based method?

It estimates emissions by multiplying the money spent on a good or service by an industry-average emission factor per unit of currency. It's GHG Protocol Formula 1.5 for Scope 3 Category 1.

Which EPA emission factor should I use for AI and cloud services?

For model APIs and cloud compute, NAICS 518210 (data processing, hosting and related services) is the closest match. The v1.3 factor is 0.093 kg CO2e per 2022 USD, with margins. For AI software seats, 511210 (software publishers) at 0.080 is a reasonable alternative. Document whichever you choose.

Why did the EPA factor for data hosting drop?

v1.3 uses 2022 GHG data and 2022 dollars, where v1.2 used 2019 data and 2021 dollars. It also switched from AR4 to AR5 GWPs. Part of the drop reflects lower emissions intensity, and part reflects the change in price level and method. It isn't all real decarbonization.

Does the spend-based method overstate AI emissions?

The evidence says yes. The MIT CTL study found spend-based estimates 5 to 9 times higher than AWS's usage-based figures. A separate working paper found 10 to 40 times for AI inference. Invoices include margin and R&D costs that don't consume electricity.

Is the spend-based method accepted under CSRD?

Yes. ESRS follows the GHG Protocol, and the GHG Protocol accepts spend-based data. You disclose the share of Scope 3 that comes from primary data, and assurance providers will expect you to improve it over time.

Carbon Accounting for AI Services: The Spend-Based Method (2026)

Carbon accounting measures the greenhouse gas emissions an organization causes. For AI services you buy, such as model APIs, cloud GPUs and AI copilots, those emissions fall under Scope 3, Category 1 (purchased goods and services). The spend-based method estimates them by multiplying what you spent by an industry-average emission factor per dollar. It's defined in the GHG Protocol's Technical Guidance for Calculating Scope 3 Emissions, Chapter 1, Formula 1.5:

Emissions (kg CO2e) = Σ (spend on the good or service ($) × emission factor (kg CO2e per $))

The headline number: across three releases, the EPA's spend-based factor for data processing and hosting, the code most AI spend maps to, fell 46%, from 0.172 to 0.148 to 0.093 kg CO2e per dollar. That puts $1 million of AI spend at 172, 148 or 93 tonnes of CO2e depending only on which version you use.

Methodology

Source: US EPA Supply Chain GHG Emission Factors. v1.3 (2022 GHG data, 2022 USD, AR5 GWPs), v1.2 (2019 GHG data, 2021 USD, AR4 GWPs), and v1.1.1 (commodity-level USEEIO factors, 2018 data)

Factor type: supply chain emissions with margins, at purchaser price, as EPA recommends for spend data

Codes compared: nine NAICS-6 codes that AI, cloud and IT spend usually maps to

How I pulled it: directly from EPA's published CSV files, September 2026. The v1.1.1 value comes from Climatiq's record of that dataset.

Caveat: each version uses a different dollar year and GHG year, so some of the change comes from price levels and not only from cleaner hosting. v1.1.1 uses a commodity classification, not NAICS-6.

Four findings you should know before you file a number

1. The factor version changes your result by up to 46%

EPA version	Data processing and hosting factor	$1M of AI spend
v1.1.1	0.172 kg CO2e/$	172 t CO2e
v1.2	0.148 kg CO2e/$	148 t CO2e
v1.3	0.093 kg CO2e/$	93 t CO2e

If your team set up the calculation a few years ago and never updated the factor, your AI emissions could be overstated by almost half. The reverse applies too. If you switch to v1.3 without explaining it, the drop will look like progress you didn't make.

2. The NAICS code you pick moves the result by about 16%

No NAICS code covers "AI." Spend gets mapped to the closest one, and the closest ones don't agree.

AI purchase	Likely NAICS code	v1.3 factor
Model API calls, hosted inference	518210 Data processing, hosting	0.093
Cloud GPU compute	518210 Data processing, hosting	0.093
AI implementation or integration work	541512 Computer systems design	0.089
Custom model development	541511 Custom programming	0.084
AI copilot or SaaS seats	511210 Software publishers	0.080
Other AI-related IT services	541519 Other computer services	0.080

Put the same invoice under 511210 or 518210 and the result differs by 16%. Both choices are defensible. The mapping is a judgment call, so write it down.

3. Emission factor databases can lag the source

On Climatiq, one of the most widely used factor databases, the v1.2 hosting factor (0.148) is correctly marked as replaced. The older v1.1.1 factor (0.172) is still labeled "Current." That's almost twice the v1.3 value.

If your carbon accounting software pulls factors from a third-party database, check the version string on each factor you use. Don't assume "current" means latest.

4. Spend-based overshoots measured cloud emissions by 5 to 9 times

A May 2026 capstone from MIT's Center for Transportation & Logistics compared spend-based estimates with AWS's own usage-based emissions reporting for MIT's AWS account. Using spend-based:

Emissions came out about 5x higher than AWS's location-based figures
Emissions came out about 9x higher than AWS's market-based figures
Monthly spikes tracked billing cycles, not actual compute use

That backs up a separate working paper by Guillermo Llopis, cited by Stanton Chase, which found spend-based AI estimates running 10 to 40 times higher than electricity-based calculations. The Llopis paper hasn't been peer-reviewed, and the MIT study comes from real invoices. Both point the same direction.

The reason is structural. An AI or cloud invoice includes the vendor's margin, R&D and model development costs. None of that burns electricity when your query runs.

Full factor comparison: EPA v1.2 vs v1.3

NAICS code	Description	v1.2	v1.3	Change
334111	Electronic computer manufacturing	0.112	0.058	−48.2%
518210	Data processing, hosting and related services	0.148	0.093	−37.2%
541330	Engineering services	0.125	0.103	−17.6%
511210	Software publishers	0.097	0.080	−17.5%
541511	Custom computer programming	0.097	0.084	−13.4%
517311	Wired telecommunications carriers	0.082	0.075	−8.5%
541519	Other computer related services	0.087	0.080	−8.0%
541611	Management consulting	0.084	0.078	−7.1%
541512	Computer systems design services	0.077	0.089	+15.6%

kg CO2e per USD, purchaser price, with margins. v1.2 is in 2021 USD and v1.3 is in 2022 USD.

Most factors went down. Computer systems design went up by 16%. If you model a trend over time, check each code on its own. You can't assume they all moved the same way.

Worked example: the spend-based method, step by step

This is an illustrative company. The inflation adjustment is an assumption for demonstration. Use the published price index for your own calculation.

A company spent $500,000 on AI services in 2026:

Purchase	2026 spend
LLM API calls	$200,000
Cloud GPU compute	$180,000
AI copilot seats	$120,000

Step 1. Map each line to a NAICS code. API calls and GPU compute go to 518210. Copilot seats go to 511210.

Step 2. Convert spend to the factor's dollar year. v1.3 factors are in 2022 USD. The GHG Protocol and EPA both recommend adjusting for inflation. This example assumes a 12% cumulative price change from 2022 to 2026, so divide 2026 spend by 1.12.

Step 3. Multiply by the factor.

Purchase	2022 USD	Factor (v1.3)	Emissions
LLM API calls	$178,571	0.093	16.6 t CO2e
Cloud GPU compute	$160,714	0.093	14.9 t CO2e
AI copilot seats	$107,143	0.080	8.6 t CO2e
Total			40.1 t CO2e

Step 4. Check what the old factors would have given. The same spend under v1.2 factors comes to 58.0 t CO2e. That uses 2021 USD and an assumed 17% cumulative price change from 2021 to 2026. It's about 45% higher, for the same purchases.

For EUR reporters: convert euro spend to USD at the average exchange rate for the reporting year, then deflate. For a European factor set, look at EXIOBASE. Whichever set you use, record the version and region.

When the spend-based method is still the right call

Spend-based carbon accounting isn't wrong. It's a starting point, and the GHG Protocol treats it that way. The Category 1 hierarchy runs:

Supplier-specific method
Hybrid method
Average-data method
Spend-based method

Spend-based makes sense for:

Screening. Finding out whether AI spend is material at all.
The long tail. Small SaaS tools and vendors that publish no emissions data.
Categories that aren't material, where a more precise figure wouldn't change a decision.

It falls short when AI spend is large, when you want to show that emissions are falling, or when an auditor asks why your number moved.

How to make a spend-based AI figure defensible

Pin the factor version. Write "EPA Supply Chain GHG Emission Factors v1.3, 2022 USD, with margins" in your methodology note.
Document the NAICS mapping for each AI purchase type, and keep it consistent from year to year.
Adjust for inflation to the factor's dollar year.
Flag the data quality as low in your inventory.
Move your largest vendors to supplier data. AWS, Microsoft Azure and Google Cloud all publish customer carbon dashboards. Use spend-based only for vendors that don't.
Restate when you change versions. If you move from v1.2 to v1.3, recalculate your base year or disclose the effect.

How this fits CSRD and ESRS E1-8

Under the revised ESRS, purchased AI emissions appear in your gross Scope 3 figures under E1-8. That's the disclosure previously labeled E1-6. See my guide to the E1-6 to E1-8 change.

Three points matter for spend-based AI figures:

Primary data share. You disclose what percentage of Scope 3 comes from supplier or value chain data. Spend-based estimates count against that percentage.
GWP values. EPA v1.3 uses AR5 GWPs. ESRS asks for the most recent IPCC values but allows older ones if you explain it. Note the basis in your methodology.
Refresh cycle. Significant Scope 3 categories are updated every year. The full inventory is refreshed at least every three years.

If you want to go beyond spend-based to token-level or compute-level calculations, my guide to calculating AI emissions for CSRD walks through the activity-based method.

The bottom line

The spend-based method is the fastest way to do carbon accounting for AI services, but the result depends on choices most teams never record: the factor version, the NAICS code and the dollar year. Pin all three, flag the figure as an estimate, and move your largest AI vendors to supplier data as soon as they offer it.

Sources

I help organizations build emissions figures that survive assurance, and write the methodology notes that make them defensible. If your AI spend is sitting in Scope 3 unmeasured, let's talk.