Same AI Request, 8.5 Times the Emissions Depending on Where It Runs

On October 21 local time, Berget AI, a Swedish AI service provider, published the method and code it uses to calculate the greenhouse gas emissions generated by each AI request. Customers can see an estimate for every request in the company’s management dashboard. The calculation code has been released under terms that let anyone modify and reuse it (the MIT license), so other service operators can plug in their own hardware and regional data.

Why an 8.5x gap for the same request?

Because of where the request is processed. The comparison Berget AI released uses the same AI model (Kimi K3), the same server hardware (B300), and the same amount of text going in and out. The company’s calculated emissions per request came to roughly 0.38g in Sweden, 1.6g in Ireland, 2.5g in Germany, and 3.2g in Texas, measured in carbon dioxide equivalent (CO₂e), which converts the warming impact of various greenhouse gases into a single figure.

The biggest driver of the gap is the emissions produced when electricity is generated. Making the same amount of power releases different quantities of greenhouse gases depending on the region. On top of that comes the electricity consumed by data center facilities themselves, including cooling. This method assumes a data center’s total power draw is 1.15 times the servers’ own consumption in Sweden, and 1.80 times in Texas. Those two factors compound, which is how Texas ends up at 8.5 times Sweden’s emissions. The company said it re-examined the sources behind these regional figures and revised the pre-release numbers of 3.9g for Texas and a 10x gap down to 3.2g and 8.5x.

How do you calculate a single request?

The calculation takes in the AI model, the length of the text exchanged, and the processing time. It also factors in the type and utilization rate of the server hardware, the emissions from generating electricity in the region where the data center sits, and the power needed to cool the equipment. Emissions from manufacturing the hardware are also included, divided across its expected service life. In other words, the approach captures not just the electricity consumed while generating an answer but also part of the footprint of manufacturing the hardware.

Berget AI said it has been reporting per-request emissions in its developer-facing service since 2025. The company explained that existing open tools such as EcoLogits and CLEER estimate emissions from outside the service provider, whereas its own approach shows customers figures drawn directly from requests processed on hardware it operates itself. With this release, other operators can now inspect and use the calculation method.

The estimates come with gaps and margins of error. They rely on regional grid averages, and the default comparison excludes the emissions generated in training the models. The current version does not account for the water and energy used to manufacture chips. The company puts the uncertainty on its hardware manufacturing emissions estimates at ±30–50%.

Can it be used in Korea?

The published code and the web calculator are open to anyone. Korea is not among the 15 regions included by default. Instead, the documentation explains how users can add their own regional values. To run AI on hardware located in Korea and calculate results under Korean conditions, you would need to enter your hardware details along with the emissions from generating electricity in Korea.

This release makes it possible to examine the environmental impact of an AI service one request at a time. The same request yields different emissions depending on where it is processed and what kind of electricity powers it. Two things are worth watching from here: how faithfully users can enter their own hardware and regional conditions, and whether the scope of the calculation expands to cover the items currently left out, such as the water and energy that go into chip manufacturing.