AI carbon footprint per employee: the Meta token leak

Nearly every big AI number published in 2026 came from a seller: Google counts the tokens its models served, Microsoft the tokens its cloud processed, NVIDIA the tokens its chips could handle. Those are sales figures, and the emissions behind them belong to whoever bought the tokens. Until April 2026, nobody had published the other side of that transaction: how many tokens one company’s own people consumed. Then a Meta employee built a leaderboard.

Where the number comes from

In April 2026, The Information reported that an employee at Meta had built an internal dashboard called Claudeonomics, named after Anthropic’s Claude, that ranked colleagues by the number of tokens they consumed. It handed out titles like Token Legend, Model Connoisseur and Cache Wizard. In one 30-day window, the company’s staff had used more than 60 trillion tokens, and the top user averaged 281 billion. Two days after the story ran, the dashboard went dark with a note saying its data had been shared externally. Meta told Fortune that the employee took it down and that the company had not asked for it.

In June 2026, The Information reported a second document: a memo to about 6,000 employees saying that internal AI use had grown exponentially, that employees consumed 73.7 trillion tokens in just over 30 days, and that the cost was heading for billions of dollars in 2026. The memo announced token budgets from 2027, a central dashboard called AI Gateway to track spend, and a push to move people from third-party tools onto Meta’s own coding assistant. Meta’s chief technology officer, Andrew Bosworth, wrote separately that token usage alone is not a measure of impact of any kind.

So the two data points are more than 60 trillion and 73.7 trillion, in consecutive 30-day windows, roughly 22% apart. Both come from The Information’s reporting on internal documents, and we cite them that way. We will update the post the day Meta publishes its own figure.

What 73.7 trillion tokens means

A token is the unit a language model reads and writes. In English it is about four characters, so a short sentence is around 30 tokens and a page of text is a few hundred. The reason one person can reach billions is that coding agents do not read a sentence at a time. They read whole files, their own earlier output and the tool results in between, on every step, for hours. Much of that is the same text sent again and again, which is why the leaderboard had a title for cache use.

Spread evenly across Meta’s 78,865 employees, 73.7 trillion tokens is about 935 million tokens per person per month. The use is nowhere near even: the memo went to about 6,000 people, the leaderboard tracked the whole company, and the top user alone accounted for 281 billion. We do not know how many people sit behind the total, so the figure we can defend is the company-wide average, with the top user as the ceiling.

The arithmetic

Our methodology, version 1.2, turns tokens into energy and then into carbon in three layers. Layer one is the electricity the chips draw while they work: a joule figure per token for reading input, for writing output and for reading from cache, multiplied by a data centre overhead and a grid factor. Layer two adds the power the host machine draws around the chips and the fact that a cluster runs at about 30% utilisation, so idle capacity is paid for by the tokens that do run. Layer three adds the manufacturing of the hardware and the training of the model, spread over the tokens they serve. The methodology page has every factor and formula.

For an aggregate count with no breakdown, we use the worked example’s split, 75% input and 25% output, no caching, on the medium tier. The medium tier is 0.3 joules per input token and 0.5 per output token, a data centre overhead of 1.2 and a grid factor of 0.42 kg per kWh.

The total moves with two assumptions. On the large tier, which is closer to the frontier models a coding agent uses, the month is 26,414 tonnes, 62% more. If 70% of the input tokens were served from cache, which is plausible for agents that re-read the same files all day, the month drops to 10,515 tonnes. The leaked figures do not say which tokens were cached or which models served them, so we show the range and lead with the default.

One employee, one month

Divide the month by the headcount and the default comes to 206 kg of CO2e per employee, with the cached case at 133 kg and the large tier at 335 kg. Over a year at the same rate, the default is about 2.5 tonnes per employee.

The top user shows how uneven the use is: one person’s month on that leaderboard weighed as much as the default figure for 300 employees.

The same person’s commute and office

Meta publishes a detailed inventory every year, and its 2025 environmental data index covers fiscal 2024, when the company had 74,067 employees. Two lines describe the same people whose tokens we just counted.

Category 7, employee commuting, was 70,273 tonnes as the GHG Protocol counts it, or 52,299 tonnes after Meta applies energy certificates to the electricity of people working from home. Per employee that is 79 kg a month, or 59 kg. Office electricity was 361,853 MWh, which the grid supplied at 104,734 tonnes; per employee, 4.9 MWh and 118 kg a month. Meta matches all of its electricity with renewable certificates, so its market-based office figure is 1,223 tonnes for the whole company, about 1.4 kg a person. Our post on market-based and location-based accounting explains why we use the grid figure for a comparison like this.

On the default assumptions, one employee’s AI use for a month weighs 2.6 times their commute and 1.8 times their share of the office’s electricity. On the cached case it is still 1.7 times the commute and a little more than the office. There is no assumption in our range under which the tokens come out smaller than the drive to work.

Whose inventory it lands in

Tokens bought from Anthropic, OpenAI or any other vendor are a purchased service, and the GHG Protocol puts purchased services in Scope 3 Category 1 of the buyer. Tokens served on Meta’s own accelerators are electricity Meta bought, Scope 2. The leaked figures do not split the two, but the leaderboard’s name and the memo’s push away from third-party tools suggest a large share was purchased. Either way, the emissions belong in Meta’s inventory.

For scale, a year of months like this one is about 195,000 tonnes, close to 10% of the 1,976,448 tonnes Meta reported for Category 1 in fiscal 2024. The two figures are from different years and the tokens would not all land in Category 1, so treat it as a sense of scale only. Our guide to Scope 3 emissions from AI covers the categories in detail.

The assumptions in the range

The figure counts tokens. Whether they did useful work is a separate question; Bosworth’s memo made that point, and some of the leaderboard’s volume came from agents left running to pad a score. Which models served the tokens, how much was cached and which grid the data centres sat on are not in the reporting, so the tier is an assumption, the energy per token is a range, and we use the methodology’s default grid factor rather than Meta’s own, which is lower in the regions where Meta builds.

What holds across the whole range is the size: when one company’s people consume AI at the rate Meta’s did this spring, the emissions per person exceed the two lines every sustainability report already tracks for them.

What to do with your own number

If your company buys AI, the number is available. Every provider meters tokens per key, per model and per day, and the same arithmetic applies. Connect the accounts, keep the provider’s own counts of input, output and cached tokens, apply published factors, and report the result beside the categories you already publish. That is what CrbonFree does, and the OpenRouter case study shows the same method on a public dataset. Per employee is one useful cut; per team and per model tell you where the tokens go.

Meta’s figure reached the public without the models or the cache share. A company that meters its own usage can publish the same number on purpose, with both filled in.