Berget AI, a Swedish inference provider, has released an open-source methodology for calculating carbon emissions at the individual AI request level — not annual averages, not estimates from outside, but per-query figures tied to specific hardware, grid intensity, and usage patterns. Developed with Stockholm Environment Institute, WattTime, and Klimatkollen, the method follows the Green Software Foundation's SCI for AI spec and ships as an MIT-licensed npm package. The headline number is stark: a typical request to Kimi K3 (5,000 input tokens, 387 output tokens) emits 0.38 grams of CO₂ in Sweden versus 3.2 grams in Texas — an 8.5× difference on identical hardware drawing roughly the same 4 Wh of energy. The gap is almost entirely explained by grid carbon intensity: Sweden's grid runs at ~8 g CO₂e/kWh (overwhelmingly hydro and nuclear), while Texas ERCOT sits at ~333 g CO₂e/kWh. Ireland lands at 1.6g per request (~224 g/kWh grid), Germany at 2.5g (~337 g/kWh). In Sweden, 89% of the per-request footprint comes from embodied hardware emissions; in Texas, 90% comes from electricity. The timing aligns with a structural shift in AI deployment. Open-weight models now handle 29% of tokens on Vercel's AI Gateway (up from 11% in April 2026) and roughly 69% of named-model tokens on OpenRouter. Unlike proprietary APIs locked to specific data centers, open models can run anywhere — which means the buyer, not the model vendor, determines the carbon intensity. The method makes that choice visible for the first time at request granularity. The methodology includes details most corporate sustainability reports omit. Embodied hardware emissions are estimated at ~1,000 kg CO₂ per data center GPU (refurbished counted as zero). Idle power draw is measured, not assumed: an NVIDIA B300 draws 236W idle, measured over seven production days. Cooling overhead varies dramatically by climate — Nordic free-air cooling achieves PUE ~1.15 with no water consumption, while hot-climate facilities push toward PUE 1.8 and consume up to 2 liters of cooling water per kWh. Texas requests carry roughly 6 mL of water per query on top of the carbon. The honest limitations are clearly stated. Training emissions are excluded because self-reported figures vary by ±50% and no reliable per-request allocation exists. The current version uses average grid intensity (location-based accounting), not marginal emissions — a distinction that matters because marginal rates, which estimate the impact of additional load on the grid, may significantly change the picture for low-carbon grids like Sweden's. Marginal emissions will be added next month. Water and energy costs of chip fabrication are also excluded from this version. The competitive logic is transparent. Berget AI runs Nordic data centers and benefits directly from Sweden's clean grid. Publishing the methodology is both a public good and a competitive weapon: if buyers can see per-request emissions, Nordic providers win on carbon metrics without trying. The IEA projects data center electricity consumption roughly doubling to 945 TWh by 2030 with AI as the principal driver. SEI's Somya Joshi notes bluntly that efficiency gains are likely to be outweighed by higher demand, making location and grid mix the controllable variables. The correction history is worth noting as a credibility signal. The original embargo release overstated the Texas-Sweden gap at 10× based on incorrect grid intensities; Berget AI published corrected numbers (8.5×) on the same day with a transparent correction note. The underlying grid intensity sources — IEA for Sweden, SEAI for Ireland, Ember for Germany, EPA eGRID for Texas — are named and checkable. This is what honest methodology release looks like: version-numbered, source-cited, error-corrected in public.
