China AI token consumption: same tokens, different grid

China counts its AI tokens, the units a model reads and writes, about four characters of English text each. Its National Data Administration and its Ministry of Industry and Information Technology give a count of the tokens used across the country each day, and in March 2026 the administration’s head gave the token an official Chinese name, 词元. China’s phone companies also sell tokens to households in monthly packs. With the national count and the plans’ posted sizes, we can work out the carbon of a whole country’s AI use and of a single retail plan, and set each beside the same tokens priced on Google’s fleet.

China’s official token count increased 5,000-fold in two years

From early 2024 to June 2026 the daily count rose 5,000-fold. For the whole of 2025 the administration’s national survey puts the total at about 21,100 trillion tokens, 21.1 quadrillion. The official wording shifted from token consumption in August 2025 to token calls from March 2026. Neither agency has published how it counts, whether tokens served to users abroad are included, or what lies behind the jump from more than 140 trillion in March 2026 to more than 500 trillion in June 2026, 3.6 times in three months, so we report the daily counts as stated and build no growth rate on them.

The market research firm IDC counted far fewer, 1,944 trillion tokens called by businesses through public-cloud model platforms in China in 2025, up from 114 trillion in 2024, close to half of them on ByteDance’s Volcano Engine. IDC’s 2025 count is about 9% of the national count for the year. IDC counts only businesses’ calls to cloud model APIs, and the national count has no published definition.

China Telecom sells 10 million tokens for 9.9 yuan, with no carbon figure

On 17 May 2026 China Telecom launched a trial of token plans for individuals and households: 10 million tokens for 9.9 yuan a month, 40 million for 29.9 yuan and 80 million for 49.9 yuan. The plans run on DeepSeek V3.2 and China Telecom’s own Xingchen model, last a month from activation, and stop when the quota is used up. China Unicom sells 6 million tokens for 15 yuan a month. China Mobile’s national plan, from 5 yuan a month, is counted in its own standard points, and its provincial arms sell tokens directly, such as 2.5 million for 5 yuan in Jiangsu.

AI providers elsewhere post prices per million tokens and bill against dollar credit, so anyone with an account can work out what a given number of tokens costs. Only in China did we find tokens sold to households the way mobile data is: by the phone company, in a monthly pack, at a posted price. We could not find a published carbon figure for any of the phone companies’ token plans.

One full 10-million-token plan is about 3 kg of CO2e on China’s grid

Crbon Labs methodology v1.2, which adopts Watershed’s open framework for measuring emissions from AI usage, counts three layers of energy and carbon: the electricity the chips draw; the host machines and idle capacity around them; and the hardware’s manufacture and the models’ training. Every factor is on the methodology page. DeepSeek V3.2 is built on DeepSeek V3, which activates 37 billion of its 671 billion parameters for each token, a size that fits our medium tier. We assume the plan is used in full, with the methodology’s default split for an aggregate count, 75% input and 25% output, and nothing cached.

China’s grid factor is the Ministry of Ecology and Environment’s latest national average, 0.5306 kg of CO2 per kWh for 2023, published on 31 December 2025. For the data centre overhead we use a power usage effectiveness, or PUE, of 1.34, meaning the building draws 34% more power than its computers, the average for China’s super-large computing facilities at the end of 2025, with a PUE of 1.46, the average for all facilities in use at the end of 2024, as the higher case. For comparison, the methodology’s large tier covers models heavier than DeepSeek V3.2, its Gemini Flash tier covers Google’s lighter Gemini models, and Google’s fleet factors are a PUE of 1.09 and 0.345 kg of CO2e per kWh.

The same 10 million tokens cost about 3.0 kg of CO2e when they are served in China and about 1.7 kg when the same model runs on Google’s fleet factors, a difference that comes from the grid and not from the tokens. China’s average grid carries 1.54 times the carbon per kilowatt-hour of Google’s data centres, and China’s buildings draw more power around each chip, so the plan’s carbon is 1.75 times higher before the model changes at all. Choosing a heavier model moves it further, to 4.9 kg of CO2e on the large tier. For a subscriber, the central case is about 0.3 kg of CO2e per yuan spent, and 36 kg of CO2e for someone who uses the full plan every month for a year.

China’s ministry names one factor for location-based reporting and another for market-based

China’s environment ministry says its national average factor is for location-based reporting. For market-based reporting, green power a company buys counts at zero, and the rest of its electricity takes the factor that excludes market-traded non-fossil power, 0.6096 kg of CO2 per kWh. For the companies that serve the tokens, the electricity is Scope 2. A company that buys tokens from a Chinese provider or carrier reports them as a purchased service, in Scope 3 Category 1, on the grid that served them. Our post on market-based and location-based emissions explains the two methods, and our guide to Scope 3 emissions from AI covers the categories.

Chinese models also reach buyers through gateways such as OpenRouter, where models from Chinese labs carried 48% of the tokens routed in August 2026, as our post on LLM market share shows. The emissions of a Chinese model’s tokens follow the grid of whichever provider served them, which is why the provider, not the model’s country of origin, decides the factor.

China’s tokens would carry 43% less carbon on Google’s fleet

Google does not serve China’s tokens, so this comparison is hypothetical. It keeps the tokens and the joules per token and changes only the factors, to the ones our methodology uses for Google’s fleet: a PUE of 1.09 and 345 grams of CO2e per kWh, Google’s own location-based figure for 2024.

China’s grid factor alone is 1.54 times Google’s: 0.5306 kg of CO2 per kWh against 0.345 kg of CO2e per kWh. With the data centre overhead included, the same tokens carry 1.75 times the emissions at a PUE of 1.34 and 1.89 times at a PUE of 1.46. China’s 2025 tokens need about 10.8 TWh at the data centre on the medium tier at a PUE of 1.34, about 5.5% of the 196 TWh China’s computing infrastructure used in 2025.

Two thirds of the gap with Google’s fleet is the grid, and China’s provinces differ by more than that gap

With the model held constant, the grid accounts for 67% of the gap between China and Google’s fleet at a PUE of 1.34, and the data centre overhead for the other 33%. At a PUE of 1.46 the split is 59% and 41%. Against the Gemini Flash tier, a lighter model on Google’s factors, the model accounts for 49% of the gap, more than the grid’s 34%, and the overhead for 17%.

Inside China, the provincial grids move a plan’s carbon more than the comparison with Google does. On the ministry’s provincial factors for 2023, at the same PUE of 1.34, the same plan is about 1.1 kg of CO2e in Sichuan, where hydropower dominates, and 3.6 kg of CO2e in Anhui. Both provinces host national data centre clusters, Tianfu in Sichuan and Wuhu in Anhui.

Alberta’s grid fell from 907 grams of CO2e per kWh in 2005 to 335 grams of CO2e per kWh in 2024, the year Alberta finished phasing out coal, and China’s 2023 average, 531 grams of CO2 per kWh, sits where Alberta’s grid stood in 2020 and 2021.

What China’s token count means for a company that buys Chinese tokens

Through Crbon Labs methodology v1.2, China’s official count of more than 500 trillion tokens a day in June 2026 comes to about 150,000 tonnes of CO2e a day on China’s average grid, and the 21.1 quadrillion tokens China counted for 2025 come to about 6.3 million tonnes of CO2e. Google’s own count, over 3.2 quadrillion tokens in May 2026, is about 103 trillion tokens a day, and priced on the Gemini Flash tier, as in our post on Gemini’s energy, it comes to about 11,000 tonnes of CO2e a day.

Google has not published a token count for a full year. Twelve months at Google’s May 2026 count would be over 38 quadrillion tokens and about 4.0 million tonnes of CO2e. China’s daily CO2e is about 14 times Google’s because China counts 4.8 times as many tokens a day, a medium-tier token carries about 1.65 times the CO2e of a Gemini Flash token on the same factors, and China’s grid and data centres add 1.75 times the CO2e per token.

For an individual subscriber, a full 9.9-yuan plan of 10 million tokens comes to about 3.0 kg of CO2e on China’s average grid and about 1.7 kg of CO2e with the same model on Google’s fleet. Two thirds of the gap between 3.0 kg of CO2e and 1.7 kg of CO2e is the grid, and the rest is in the buildings the chips sit in.

For a company that buys tokens from a Chinese provider, a carrier or a gateway, the grid that served the tokens sets the emissions, and the model’s country of origin does not. Tokens bought as a service belong in Scope 3, Category 1, on the factor of the grid behind the provider, and inside China that factor runs from 0.1564 kg of CO2 per kWh in Sichuan to 0.6553 kg in Anhui, a wider spread than the difference between China’s average and Google’s fleet. When a provider does not say where it served the tokens, the national average of 0.5306 kg of CO2 per kWh is the location-based factor the environment ministry itself assigns, and it should be named in the report as an assumption.

On OpenRouter, a model from a Chinese lab is served by whichever provider OpenRouter routes the request to, and most of those providers are headquartered outside China. On 6 October 2026, DeepSeek V3.2 had 13 providers on OpenRouter. Nine of the 13 providers were headquartered in the United States, and of the five that listed the countries of their data centres, four listed the United States and one, Alibaba, listed Singapore and China. The Chinese labs Z-AI, Tencent and Xiaomi listed Singapore for the data centres behind their own endpoints, and Alibaba was the only one of OpenRouter’s 112 providers to list a data centre in China. OpenRouter records the provider that served each request, so a company that calls a Chinese model through OpenRouter can find which provider answered and apply the grid factor of the country where that provider runs its data centres.

CrbonFree meters the tokens a company uses per call, per model and per provider, including the tokens it buys through OpenRouter. Crbon Labs methodology v1.2 integrates Watershed’s open framework for measuring emissions from AI usage, and it prices Google’s models on Google’s fleet factors and every other model on a global average grid of 0.42 kg of CO2e per kWh at a PUE of 1.2. Every factor is published on the methodology page, so a company whose tokens were served in China can restate them on China’s 0.5306 kg of CO2 per kWh, as this post does. A company that connects its accounts to CrbonFree sees its own tokens, energy and CO2e month by month on Crbon Labs methodology v1.2, and can set them beside the figures we publish for China, for Google and for the other large providers and users of AI.