A Swedish AI host says the same request emits 8.5 times more CO2 in Texas than in Sweden
Berget AI counts about 0.4 grams of CO2e for a typical request in Sweden and 3.2 grams in Texas, on identical hardware. In Sweden, 89% of that is the chips' manufacturing.
The figures come from an open-source per-request method that Berget AI built with SEI, WattTime and Klimatkollen, for a 5,000-token-in, 387-token-out request on a B300 node. Energy use is about 4 Wh in both places.
The company notes that an embargoed version had overstated the gap, and the final numbers were revised down after a grid-intensity correction.
What the method counts
Berget AI's release estimates emissions per request for Kimi K3 on a B300 node in shared deployment: Sweden about 0.4 g CO2e, Ireland about 1.6 g, Germany about 2.5 g and Texas about 3.2 g. Grid intensity assumptions run from roughly 8 g/kWh in Sweden to 333 in Texas, and embodied emissions are about 1,000 kg of CO2 per data-centre GPU. In Sweden about 89% of emissions come from hardware manufacturing; in Texas about 90% come from electricity. The company is the source of all of these figures, and it excludes training.
The objections
A commenter based in Paris checked the grid numbers and concluded: "So it's not really about ai, the models or the hardware. Fossil fuels make CO2, that's it." A quant developer argued the comparison ignores energy trade: "a Watt spent in Sweden increases the marginal cost of energy by the same amount as a Watt spent in Texas."
A developer from Sweden went after the calculator itself: you cannot know which hardware or inference backend a provider uses, and the tool uses a response-time metric that cannot separate compute from rate limiting. The company says marginal emission rates will be added next month and may significantly change the Swedish figures. For anyone choosing a region today, the practical lever is where the electricity comes from, because the two places differ by far less than the grid numbers imply once manufacturing is counted.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
UUIDv7 only promises uniqueness, and Python's version can run ahead of your clock
The blog post shows Python's uuid.uuid7() drifting ahead of the system clock after a backward clock jump, with the gap growing without bound at high load.