Berget AI estimates Texas request emissions at 8.5 times Sweden’s
The Swedish AI provider has published an open-source carbon calculator, saying a matched request produces sharply different emissions depending on electricity supply, cooling and hardware assumptions.
Swedish provider Berget AI says its open-source calculator estimates that the same specified AI request produces about 8.5 times as much carbon pollution in Texas as in Sweden. The comparison is an estimate under defined operating assumptions, rather than a finding that applies to every AI interaction. Berget AI’s announcement
The company’s release, dated October 21, 2026, puts the request’s footprint at 0.38 grams of carbon dioxide equivalent in Sweden and 3.2 grams in Texas. It specifies the Kimi K3 model, B300 hardware, 5,000 input tokens and 387 output tokens in a shared deployment, weighted across a typical day’s traffic.
Berget also discloses a correction to an earlier version distributed under embargo, which had claimed a tenfold difference. The published comparison uses regional average electricity emissions factors, and the company says a planned addition of marginal emissions estimates could substantially change Sweden’s measured intensity. Release and correction
The calculator’s public repository describes a tool intended to let providers configure their own models, hardware and data-centre locations. Its calculations separate GPU operation, server infrastructure, cooling and hardware manufacturing, rather than treating electricity consumption as the entire footprint. That distinction helps explain why changing location does not eliminate all estimated emissions. Calculator repository
Hardware manufacturing is allocated across the equipment’s operating lifetime, so that component does not simply disappear when a server uses lower-carbon electricity. The repository explains that this location-independent contribution narrows the difference between regions when total emissions are compared, even where the operational electricity footprints differ much more sharply.
Cooling assumptions also vary by location, while equipment utilisation affects how fixed energy costs are spread across requests. The repository explicitly describes its methods as vendor-neutral but its default settings as reflecting Berget’s own infrastructure. Buyers therefore need matched configurations, not just headline ratios, when comparing providers. Repository documentation
The detailed methodology adopts the core Software Carbon Intensity formula: energy consumption multiplied by grid emissions intensity, plus an allocation of embodied hardware emissions. However, it says it does not implement the full SCI-AI lifecycle analysis. Its focus is the operation and monitoring phase, with a request used as the reporting unit. Published methodology
That boundary matters for interpreting the result. Model training is excluded by default, as are end-user devices and network transmission. The methodology also describes input from a Stockholm Environment Institute researcher as targeted comments on specific sections, not a full institutional review or endorsement of the document.
For organisations comparing AI services, the practical implication is that a per-request figure needs an accompanying account of what was measured, allocated or excluded. The open documentation makes those choices inspectable; the next test is whether additional provider measurements and revised electricity-accounting assumptions produce comparable results. Methodology and limitations
Sources
Primary source: Hacker News. Newzly rewrites and structures source reporting for clarity; verify critical details at the original.