UpTrajectory Review

A Swedish AI startup, Berget AI, working with the Stockholm Environment Institute and two climate-data organizations, has published what it calls a new method for measuring the carbon emissions of individual AI queries. The headline finding: a typical request to the Kimi K3 model produces roughly 0.4 grams of CO2 when processed in Sweden but 3.2 grams in Texas — an eightfold difference on identical hardware. The divergence comes entirely from the carbon intensity of the local electricity grid. Sweden runs heavily on hydro and nuclear, Ireland and Germany still burn meaningful gas and coal, and Texas relies on a dirtier mix despite its wind capacity. Berget has released the methodology as open source and paired it with a web calculator that lets users compare models and regions side by side.

For a small-business operator, this is the first practical tool that turns AI's climate cost from an abstraction into a procurement decision. Most owners we talk to have no idea that moving a workload from one data center region to another can cut its footprint by 75 percent or more without changing a single line of code. If your company has made any public sustainability commitments — or reports to larger clients who audit their supply chains — the location of your AI processing now matters in the same way the location of your manufacturing does. The calculator makes it possible to ask a vendor a pointed question: where exactly is my inference running, and can you show me the emissions per request?

What is genuinely new here is the granularity. Existing carbon estimates for AI tend to be annual, company-level figures — Google's or Microsoft's total data-center emissions — which are easy to dismiss as someone else's problem. Berget's method attributes emissions to individual requests and specific geographies, which makes the number actionable. We are somewhat skeptical of the precision: grid carbon intensity fluctuates hour by hour, and the method likely relies on averages rather than real-time marginal emissions data. WattTime's involvement suggests some sophistication on that front, but the calculator should be treated as a directional tool, not an audit-grade measurement. The choice of Kimi K3, a less-known model, also raises the question of how the numbers look for GPT-4-class systems that are far more computationally hungry.

The second-order effects cut in two directions. On one hand, transparent per-request accounting could accelerate a race among cloud providers to clean up their grids, because suddenly the dirtiest regions carry a visible penalty. On the other hand, it could concentrate AI workloads in a handful of low-carbon regions — Scandinavia, Quebec, parts of the Pacific Northwest — creating new capacity constraints and pushing up prices there. Small businesses in Europe may also find themselves at a structural advantage over US-based competitors if EU regulations begin requiring emissions disclosure per AI workload, which the EU's AI Act and corporate sustainability reporting rules are already moving toward. The open-source release means the methodology could become a de facto standard, which is either a public service or a competitive moat for Berget, depending on how you read it.

The practical next step is simple: before your next AI vendor renewal or infrastructure decision, run your current usage through Berget's calculator and see what your actual footprint looks like by region. If the number is material — and for anything beyond light chatbot use, it may be — ask your provider whether they offer region selection for inference and what their per-request emissions are. Watch also whether larger AI labs adopt or contest the methodology. If OpenAI, Anthropic, or Google begin publishing per-request, per-region figures using a compatible approach, that will signal the metric has teeth. If they ignore it, that silence will be telling too.

“a typical request to the AI ​​model Kimi K3 results in around 0.4 grams of carbon dioxide emissions in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany, and 3.2 grams in Texas” — CIO Magazine

Takeaway: Run your AI workloads through Berget's open-source calculator and ask vendors where your inference runs — the same query can cost eight times more carbon in one region than another.