Microsoft AI emissions: the tokens its report leaves out

On its quarterly earnings calls Microsoft counts tokens, the units its models read and write, about four characters of English each. In its annual sustainability report it counts emissions. The calls don’t mention emissions, the report doesn’t mention tokens, and no reporting rule asks Microsoft to connect them. The arithmetic that connects them uses public data only, and anyone can run it. We ran it on Microsoft’s own token count and emissions total for fiscal 2025, the year from July 2024 to June 2025.

What Microsoft said about tokens

Satya Nadella has given a token count on most of Microsoft’s earnings calls since April 2025, and what each count covers changes from call to call.

The 500 trillion tokens of July 2025 is the only full-year total Microsoft has given, and it covers the same twelve months as the latest sustainability report. It covers only the tokens served by Foundry APIs, the service companies use to build on hosted models, and Microsoft itself describes that scope as narrow. The calls do not say whether Microsoft 365 Copilot, GitHub Copilot, Bing or OpenAI’s own traffic on Azure sit inside it. Since July 2025 Microsoft has published only counts of customers.

Microsoft does connect tokens and energy elsewhere. On the call of 28 January 2026 Nadella said the key metric Microsoft is optimising for is tokens per watt per dollar. Microsoft has not published a tokens-per-watt figure.

What Microsoft’s 2026 sustainability report says

Microsoft published its 2026 Environmental Sustainability Report on 9 July 2026, under the subtitle “Responsibly building the AI future”. It covers fiscal 2025. Total emissions across Scopes 1, 2 and 3 were 20.29 million tonnes of CO2e, up from 16.22 million tonnes a year earlier. In Microsoft’s words, the increase was “driven primarily by the expansion of our datacenter infrastructure and pausing our use of non-additional, unbundled renewable energy certificates as we prioritize investments that bring net new power to grids.”

Against Microsoft’s 2020 baseline of 12.88 million tonnes, the fiscal 2025 total is 57.5% higher. The report does not print that percentage; it is our arithmetic on Microsoft’s own fact sheet, whose main totals Deloitte & Touche reviewed. Electricity use rose 24% in the year to 37.0 TWh, more than three times the fiscal 2020 level, at a datacenter power usage effectiveness of 1.17, meaning its buildings draw 17% more power than their computers.

The report discusses AI at length. It describes AI infrastructure driving demand for energy, water, land and materials, and it cites Microsoft research finding that optimised, large-scale AI inference can use less than one watt-hour per query. It gives no share of energy or emissions for AI, and the word token does not appear in it, in its data fact sheet or in the 2025 report before it. Microsoft’s seven earnings calls from January 2025 to July 2026 do not use the words emissions, carbon or climate.

Where Microsoft’s 25% increase came from

Two lines in the fact sheet account for more than the whole increase. Capital goods, the Scope 3 category that holds the construction and hardware of new data centres, rose 2.75 million tonnes, to 9.04 million tonnes. Market-based Scope 2 rose 2.45 million tonnes, to 2.71 million tonnes, because Microsoft stopped buying unbundled renewable energy certificates in February 2025, so more of its purchased electricity now carries grid emissions in the market-based total. Scope 2 went from under 2% of the total to 13%. Falls elsewhere, mainly in the use of sold products and in purchased goods and services, brought the net increase to 4.08 million tonnes.

Much of the Scope 2 jump reflects that decision. Location-based Scope 2, which counts the grid’s emissions whatever certificates a company buys, rose 2.08 million tonnes, 20.8%, close to the 24% increase in electricity use. Our post on market-based and location-based emissions explains the market-based and location-based methods.

Connecting Microsoft’s tokens to its emissions

The tokens to connect are the 500 trillion Microsoft counted through Foundry APIs in fiscal 2025. We run them through Crbon Labs methodology v1.2 exactly as we ran Meta’s leaked token count. The methodology works in three layers: the electricity the chips draw, then the host machines and idle capacity around them, then the manufacture of the hardware and the training of the models. Microsoft does not say which models served the tokens or how many were read from cache, so we use the default split for an aggregate count, 75% input and 25% output with no caching, on two tiers. The medium tier covers flagship models such as GPT-4o, and the large tier frontier models; Nadella said in July 2025 that the GPT-4o family had the highest volume of inference tokens. The factors themselves are on the methodology page.

The per-token factors land inside the range of Microsoft’s own research. A paper by Microsoft researchers, published in Joule in April 2026 with Melanie Nakagawa, the company’s chief sustainability officer, among its authors, estimates a median of 0.31 watt-hours per query for frontier-scale models on H100 servers, with an interquartile range of 0.16 to 0.60 watt-hours. For the paper’s standard query of 500 input and 300 output tokens, methodology v1.2 gives 0.39 watt-hours at the data centre on the medium tier, inside that range. Google’s published median of 0.24 watt-hours per Gemini prompt sits in the same range, and our post on how much energy Gemini uses checks that figure against Google’s own token count the same way.

The tokens Microsoft chose to count explain about one two-hundredth of its emissions and about a fortieth of the year’s increase. On the medium tier they would need to reach about 18.5 quadrillion tokens in a year, 37 times the published 500 trillion, to account for the whole increase. Comparing like with like, the tokens added in fiscal 2025 (the 500 trillion less the fiscal 2024 count implied by “up over 7X”) come to about 95,000 tonnes of CO2e on the medium tier and 154,000 tonnes on the large tier, set against 2.75 million tonnes for capital goods and 2.45 million tonnes for Scope 2. Microsoft’s 25% increase came from building the data centres that will serve tokens and from counting grid electricity that certificates used to cover, and the tokens the report leaves out are small next to the buildings.

What the 110,000-tonne estimate leaves out

The tokens behind the 110,000-tonne estimate are the ones Microsoft chose to count, and Nadella called that count narrow. The Copilot products, Bing and OpenAI’s traffic on Azure are not stated either way, and the customer counts since then set a floor at best: over 300 customers each on track for more than a trillion tokens means more than 300 trillion tokens from those customers alone. The tokens Microsoft’s data centres serve in total are more than 500 trillion, by an amount Microsoft has not published.

The 110,000-tonne estimate also adds no emissions to Microsoft’s inventory. The electricity behind tokens served in its own data centres is already inside its Scope 2, and the servers are inside capital goods. It attributes part of those lines to the counted tokens. Microsoft has the data to make that attribution itself, since it bills its API customers by the token and optimises for tokens per watt per dollar, and it has not published one.

What the reporting frameworks ask for

No reporting framework requires a company to report its AI usage or the emissions from it. The GHG Protocol’s standards, the basis for most emissions reporting, do not mention AI. The International Sustainability Standards Board’s climate standard does not mention AI either, and its industry guidance for software companies asks about data centre energy and planning. California’s SB 253 requires emissions reports in line with the GHG Protocol, with the first due, on the regulator’s current proposal, on 10 November 2026. The US Securities and Exchange Commission’s climate rule has been stayed since April 2024, and the Commission proposed rescinding it on 29 May 2026.

Two European rules come closest. The European Sustainability Reporting Standards ask a company, where material, to disclose the emissions from purchased cloud computing and data centre services as a subset of Scope 3 Category 1: the buyer’s side, and cloud in general. The EU AI Act requires providers of general-purpose AI models to document a model’s known or estimated energy consumption, for regulators on request, and not for the public. The GHG Protocol is revising its corporate standard jointly with ISO, with publication planned for late 2028, and one of its working groups has discussed how to account for digital products, AI among them.

What Microsoft’s token count and emissions total mean for an Azure customer

Through our methodology, the 500 trillion tokens Microsoft’s Foundry APIs served in fiscal 2025 come to about 110,000 tonnes of CO2e, 0.54% of its emissions total of 20.29 million tonnes for the same year, and its 25% increase came from data centre construction and a change in how electricity is counted.

For a company that buys Azure OpenAI or Foundry tokens, nobody will connect the tokens to the emissions for you. No reporting framework requires it, Microsoft’s report gives no AI line, and Watershed’s open framework for measuring emissions from AI usage, which our methodology adopts, notes that Microsoft offers no AI-specific breakout of emissions for services such as Azure OpenAI. Microsoft’s own documentation says the emissions of Azure services become customers’ Scope 3 Category 1, and its Azure tools allocate them by billing and usage, so the AI line in your report comes from your own token counts. Our guide to Scope 3 emissions from AI and our post on the GHG Protocol and AI cover how to report it.

Every call to Azure OpenAI or a Foundry model returns its input and output token counts, with cached input where the model supports it, and a published factor set such as methodology v1.2 turns a year of them into tonnes of CO2e with the boundary and factor version stated. CrbonFree does this for Microsoft Azure Foundry and the other providers it connects.