Berget AI Says a Texas Request Emits 8.5 Times More CO2 Than Sweden's. The Grid Gap Is 42 Times.
A.I. / news
Berget AI Says a Texas Request Emits 8.5 Times More CO2 Than Sweden's. The Grid Gap Is 42 Times.
The Swedish inference host's own calculator attributes 89% of a Swedish request's footprint to building the GPUs, which is why its headline ratio is a fifth of what grid data alone would give.
Berget AI, a Swedish inference provider, said on October 21 that one typical request to the Kimi K3 model emits 3.2 grams of CO2-equivalent in Texas and 0.38 grams in Sweden, a ratio it rounds to 8.5 times. The grid data it cites for the same two places differs by a factor of about 42.
The gap between the two ratios is the story. Berget's release explains it: in its Swedish case, about 89% of the footprint is hardware manufacturing, which does not change with location. In Texas, about 90% is electricity.
The numbers behind the headline
Berget published per-request figures for one request of 5,000 input and 387 output tokens on a B300 node in a shared deployment, using its own open-source calculator (version 1.1.1). It put the energy at about 4 Wh in both places.
- Sweden0.38 g CO2e
- Ireland1.6 g CO2e
- Germany2.5 g CO2e
- Texas3.2 g CO2e
Source: Berget AI release and blog post, Oct. 21, 2026
The grid intensities Berget used are about 8 g CO2e per kWh for Sweden (International Energy Agency, 2024) and about 333 for Texas's ERCOT grid (EPA eGRID, 2023).
| Place | Grid, g CO2e/kWh | Per request, g CO2e |
|---|---|---|
| Sweden | about 8 | 0.38 |
| Ireland | about 224 | 1.6 |
| Germany | about 337 | 2.5 |
| Texas | about 333 | 3.2 |
Checking the grid figures
Both ends of the comparison hold up against independent data. The EPA's eGRID summary tables for 2023 list the ERCOT subregion at 733.862 pounds of CO2 per megawatt-hour, which is about 333 grams per kilowatt-hour. Ember, the energy think tank, reports that 99% of Sweden's electricity is low-carbon and 1.2% fossil, for 2025.
So the electricity side is not where the 8.5 comes from. At 4 Wh, a Swedish request carries roughly 0.03 grams of operating emissions. The remaining 0.34 grams or so is manufacturing, according to Berget's 89% split.
What the method assumes
Berget listed its own caveats. It uses average grid intensity, not marginal, and said marginal estimates are planned for next month and "could change Swedish results significantly". Training is excluded. Refurbished GPUs are counted as zero embodied carbon, and new data-centre GPUs are assumed to carry about 1,000 kg of CO2. Water and energy used in chip manufacturing are not yet counted.
The per-request figures are calculated from an assumed request profile, not metered. Berget did measure one input: an idle B300 draws 236 W, over seven days in production.
Berget is a seller of inference in Sweden, and its release is written by the company, with SEI, WattTime and Klimatkollen named as partners. The post quotes Chief Executive Andreas Lundmark, Chief Technology Officer Christian Landgren and SEI Research Director Somya Joshi.
The release also carries a correction. An earlier embargoed version gave 1.2, 2.1 and 3.9 grams and a tenfold difference between regions, and was replaced by the figures above.
The Terminal's own arithmetic raises one question the post does not answer in the parts read. Multiplying 4 Wh by 333 g/kWh gives about 1.3 grams for Texas, not the roughly 2.9 grams that 90% of 3.2 grams implies. The post lists about 6 mL of cooling water as a Texas factor, but the numbers as read do not show how the difference is made up.
Related coverage: the grid-scale storage story in Rio Tinto and Rondo's 360 MWh heat battery, and the Muse chat exposure in Meta's shared-chat report.
Berget said it will publish marginal-emissions figures next month.
Sources
- 03Sweden
More in A.I.
- 01OpenAI's Reported $30 Billion Raise Now Prices the Company at $1.25 Trillion, Down From $1.4 TrillionNothing is signed and OpenAI declined to comment. The valuation in the reporting has moved three times in a month, and the March round is the only figure OpenAI has confirmed.
- 02Gemini 4 Argon's $2 Introductory Price Doubles on November 30Google's launch post and a newsletter written days earlier disagree on who can use the model, and the independent scores sit alongside a hallucination figure that only one of them reports.
- 03WorldTUI Is a 360M-Parameter Model That Fakes a Linux Terminal, and Its Weights Are Not Out YetSimon Edwardsson of V7 trained a trimmed Qwen2.5-Coder on 2 billion tokens of recorded shell sessions. It scores 80 of 106 on his own test and still forgets files it just wrote.
- 04Omnara, a 3,200-Star Apache-2.0 Rival to Claude Managed Agents, Ships With Defaults Its README Calls InsecureThe omnara-ai project runs agents on any model through a Postgres-backed event log and a daemon that dials out. Its README does not name the team behind it.