One AI request can carry 8.5 times more carbon, depending on where it runs
CO₂e
A common way to express different climate-warming gases as carbon dioxide.
PUE
A measure of extra data-center power beyond the computers.
marginal-emissions factors
Estimates of emissions caused by one more unit of electricity demand.
What happened
On October 21, 2026, Berget AI, a Swedish company that runs AI services, released an open-source method for estimating carbon dioxide equivalent, or CO₂e, for each AI request. Its release says the method was developed with the Stockholm Environment Institute, WattTime, and Klimatkollen.
The comparison used Kimi K3, an open model, on identical B300 hardware. A typical request produced an estimate of 0.38 grams in Sweden and 3.2 grams in Texas. Berget AI described the difference as 8.5 times. Both cases used about 4 watt-hours of electricity. The Texas case also used about 6 milliliters of cooling water. In the estimate, hardware manufacturing made up 89% of Sweden’s total. Electricity made up 90% of Texas’s total.
Why location changes the result
The method adds GPU work, server power, idle power, cooling, and hardware manufacturing. It excludes model training by default. The release uses average grid factors of about 8 grams of CO₂e per kilowatt-hour for Sweden and 333 for Texas’s ERCOT grid. Cooling adds another difference. The reported PUE values are 1.15 for Sweden and 1.80 for Texas. PUE describes the extra facility power needed beyond the computers themselves.
So the model and hardware can stay the same. The electricity system and the climate around the data center can still change the estimate.
Why it matters
AI’s footprint is not determined only by model size. It can change with the model, prompt length, cached input, time of day, hardware use, and server location. Open models can run in different data centers. If providers report request-level numbers, buyers can compare more than price and speed.
That does not mean every AI service will show an 8.5-fold gap. It means location belongs in the question.
What is confirmed
The calculator and code are published under an MIT license. The project follows the Green Software Foundation’s SCI for AI specification. Berget AI says customers can view emissions for their own requests. Its public method also lists what it includes and excludes.
The story drew attention on Hacker News, a technology news forum. The supplied cluster data lists 116 points and 120 comments. Those figures measure community attention. They do not prove that the estimate is correct.
What remains uncertain
These are estimates, not readings from a meter attached to every request. Results can change with the model, token counts, cache hits, traffic, hardware use, and accounting choices. Average grid intensity is not the same as the emissions caused by one additional unit of demand. Training emissions are also absent from the default result.
The release corrected some regional figures on the day it appeared. The public repository says SEI provided targeted comments, not a full institute review or endorsement. That makes the 8.5-fold result useful as a transparent case study, not a universal law.
What to watch next
Berget AI plans to add marginal-emissions factors, which estimate the effect of extra electricity demand. The larger test will be whether other providers publish comparable, request-level data. Independent calculations can then check the hardware, cooling, grid, and manufacturing assumptions.
Sources: Berget AI’s release and the open calculator methodology.
The same AI question can have different carbon costs
📰 Full story: One AI request can carry 8.5 times more carbon, depending on where it runs
Where an AI runs can change how much carbon pollution it creates.
carbon intensity
How much carbon pollution comes from each unit of electricity.
PUE
A measure comparing total facility power with computer power.
request-level
Measured or estimated separately for each request.
💡 The gist
- The same AI request can create different carbon amounts.
- Texas produced 8.5 times Sweden’s estimated amount.
- The result depends on location and calculation choices.
Berget AI, a Swedish company that runs AI services, studied this question. It tested Kimi K3, an AI model, on the same hardware. The test used one typical request. Sweden’s estimate was 0.38 grams of CO₂e. Texas’s estimate was 3.2 grams of CO₂e. Berget AI called this an 8.5-fold difference.
Why did the same model have different results? AI computers use electricity while making answers. Electricity can come from different sources. Some sources create more carbon dioxide than others. Sweden’s grid has a lower reported carbon intensity. Texas’s grid has a higher reported carbon intensity. Texas is also hotter. Hot data centers need more cooling.
The calculation included computer work, idle power, cooling, and hardware manufacturing. It did not include the electricity used to train the model. It also estimated about 4 watt-hours in both places. The Texas case included about 6 milliliters of cooling water.
One detail changes the story. Sweden’s estimate included 89% from hardware manufacturing. Texas’s estimate included 90% from electricity. Cleaner electricity lowers operation emissions. It cannot erase hardware-making emissions. A measure called PUE also matters. It compares total facility power with computer power. The reported values were 1.15 in Sweden and 1.80 in Texas.
This matters because AI impact is not only about model size. It can change with the model, request, cache, time, and server location. Open models can run in different data centers. Request-level numbers could help buyers compare services.
The number is still an estimate. It is not a meter reading from every request. Different assumptions can change the answer. Berget AI corrected some regional numbers on release day. Its public code lets people inspect the method. The project says it follows the SCI for AI standard.
The story attracted attention on Hacker News, a technology news forum. The supplied data lists 116 points and 120 comments. Those numbers show attention, not truth. They do not confirm the calculation.
Next, Berget AI plans to add marginal emissions factors. Those factors estimate what extra electricity demand causes. Other AI providers may publish similar request-level numbers. Independent checks can compare the assumptions.
💬 How should we think about the AI emissions gap?
The comments discuss the article’s claim that the same AI work can produce different CO2 emissions depending on where it runs, and they question how meaningful that comparison is. The numbers are commenters’ self-reported calculations or figures taken from the article.
- Commenters say the article uses a public calculation model. The article’s reported Kimi K3 estimate is 0.4 g in Sweden, 1.6 g in Ireland, 2.5 g in Germany, and 3.2 g in Texas, making Texas 8.5 times Sweden in that example. One commenter suggests adding the result to CodeCarbon.
- A request is not always the same amount of work. Changing a button and building a whole website are very different tasks. A commenter claims that deeper reasoning and tool use can make the computer do far more work.
- Some people say clean electricity used by AI could otherwise heat homes or run factories. Others say Sweden and Texas are not connected, and electricity cannot simply be treated as one shared global pool.
- Data centers need power, network connections, roads, workers, and sometimes nearby heat customers. Commenters disagree about how useful waste heat is and about whether some Danish examples are already operating.
- Some favor carbon taxes and clean-energy requirements instead of bans. A commenter also compared a burger with thousands of AI requests, but that was a personal calculation; other commenters say total usage matters and different sources of emissions can be addressed together.
initial digest at 120 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
The same AI question can make different amounts of warming gas
📰 Full story: One AI request can carry 8.5 times more carbon, depending on where it runs
AI computers use electricity to make answers.
CO₂
A gas that can make Earth warmer.
estimate
A number calculated to be close to the real amount.
Berget AI, a company that runs AI helpers, tested one question. Kimi K3, an AI helper, answered it. The same computer setup ran the test.
In Sweden, the estimate was 0.38 grams. In Texas, the estimate was 3.2 grams. The report called Texas’s amount 8.5 times larger.
CO₂ is a gas that warms Earth. Making an answer uses electricity. Electricity is made in different ways. Some ways make more CO₂. Texas is hotter, too. Hot computers need more cooling.
The number includes making computer parts. It does not include teaching the AI. It is an estimate from a calculation. It is not a number caught by a tiny scale.
Berget AI showed its calculation rules. Hacker News, a website for tech talk, discussed the story. The supplied count shows 116 points and 120 comments. Those numbers show attention. They do not show truth.
People should compare more tests later.
💬 How much pollution can one AI question make?
The commenters are trying to understand why the same AI question may use dirtier or cleaner electricity in different places.
- A calculation copied from the article says one Kimi K3 request produces about 0.4 g in Sweden and 3.2 g in Texas. The thread presents this as a reported estimate, not a separately checked fact.
- One question can be tiny or huge. Asking for a small change is different from asking the computer to build something big, especially when the AI thinks longer or uses tools.
- People disagree about clean electricity. Some say AI may take power from homes or factories; others say Sweden and Texas are not connected, so their electricity is not one shared bucket. Data centers are placed where power and connections are available, and people disagree about using their waste heat.
- One commenter estimated that a burger could equal thousands of AI requests, but that was a personal calculation. The commenters also say we must consider how much AI is used, not just the pollution from one request.
initial digest at 120 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 What HN commenters debated about per-request AI emissions
The discussion starts from the article’s regional comparison and questions what a request means, whether clean power is actually displaced, why data centers are sited where they are, and how AI use should be compared with other consumption. The figures, performance claims, and operational examples below are commenters’ self-reported claims, calculations, or figures copied from the article; the comments do not independently verify them.
initial digest at 120 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.