The Swedish inference provider's measurements show an identical AI request emits 8.5x more CO2 when served from Texas's ERCOT grid (400+ gCO2/kWh) versus Sweden's hydro/nuclear mix (40-50 gCO2/kWh). Their argument is that the biggest lever on per-request carbon cost right now is where the inference physically runs, not model efficiency or architecture.
By submitting the Berget.ai post to Hacker News, jlundberg amplifies the position that regional grid choice dwarfs other sustainability variables in AI inference. The 111-point score suggests the framing resonated with the HN audience.
The editorial argues that training-only numbers, fleet-wide averages, and renewable energy certificates purchased elsewhere are defensible for annual reports but useless for understanding a single request's true cost. A call routed to a Texas gas peaker at 7pm isn't actually canceled out by a wind PPA signed somewhere else — per-request measurement exposes that gap.
The editorial highlights that OpenAI, Anthropic, Microsoft, Google, and Meta have signed multi-gigawatt deals disproportionately in Texas and the US Southeast, chosen for permitting speed and gas availability rather than grid cleanliness. The implication is that the industry's build-out trajectory is baking in exactly the high-intensity grids Berget's data flags as worst-case.
Swedish inference provider Berget.ai published measurements comparing the per-request CO2 emissions of running the same AI workload across different regional grids. Their headline number: an identical request emits roughly 8.5 times more CO2 when served from Texas (ERCOT) than from Sweden's largely hydro- and nuclear-powered grid. The post pins the delta on grid carbon intensity — Sweden's electricity mix hovers around 40–50 gCO2/kWh, while ERCOT in Texas regularly runs over 400 gCO2/kWh depending on time of day and gas-vs-wind dispatch.
The comparison isn't limited to those two regions. Inference served from coal-heavy grids (parts of Poland, India, and some US regions) looks even worse per joule; grids dominated by hydro, nuclear, or wind (Sweden, Norway, Quebec, France) look dramatically better. The provider's point is narrower and sharper than a general sustainability pitch: the biggest variable in a model's per-request carbon cost right now has nothing to do with the model.
The data arrives against a backdrop of hyperscalers aggressively scaling datacenter capacity for inference. OpenAI, Anthropic, Microsoft, Google, and Meta have all signed multi-gigawatt power deals in the last 18 months, with Texas and the US Southeast disproportionately represented because of permitting speed and gas availability. The implicit assumption in most capacity planning has been that "24/7 carbon-free energy" PPAs clean up the accounting. Per-request measurement cuts through that assumption.
Most public emissions numbers from AI labs are either (a) training-only, (b) averaged over an entire fleet, or (c) netted against renewable energy certificates purchased somewhere else on the planet. That accounting is defensible for annual sustainability reports and almost useless for anyone trying to understand the marginal cost of a single API call. A request routed to a datacenter running on gas peakers at 7pm Texas time is not canceled out by a wind farm in Oklahoma selling RECs at 3am.
There's a genuine debate here, and the strongest counter-argument is worth stating fairly: hyperscalers argue that location-based accounting penalizes them for building where land, water, and permits are available, and that market-based accounting (PPAs, RECs) is what actually drives new renewable buildout. That's not nothing. New gigawatt-scale solar and wind in Texas exists in part because of exactly these deals. But the gap between "my long-term PPA finances a new wind farm" and "the electron powering your 2pm request came from gas" is real, and the 8.5x number is what that gap looks like when you measure the electron instead of the contract.
The practical consequence for engineers is that region is now a first-class carbon lever, not a rounding error. The industry has spent two years obsessing over tokens-per-dollar and tokens-per-watt at the model level — quantization, speculative decoding, MoE routing, smaller distilled models. All of that matters. But the Berget data suggests the headline optimization — the one you can apply today without touching your model, your prompts, or your infra — is picking a region. An unquantized Llama 70B served from Stockholm is cleaner per request than a heavily optimized 7B served from Dallas at peak demand.
It's also worth noting what the measurement doesn't capture. Embodied carbon of GPUs, water usage (a separate and increasingly contentious number in drought-prone regions), and the carbon cost of training amortized across inference are all excluded. If anything, those additions make the regional story more lopsided, not less — a GPU in Sweden runs cooler, uses less water, and sits on a cleaner grid for its entire 5-year useful life.
If you're building on OpenAI, Anthropic, or Google's APIs, you currently have very little direct control over which region serves your request — routing is opaque and optimized for latency and capacity, not carbon. The most actionable step right now is simply measurement: ask your provider for per-request or per-region emissions data, and treat vague "we're carbon neutral" responses as the non-answer they are. Microsoft and Google publish some region-level data; OpenAI and Anthropic do not, and that's a reasonable thing to pressure them on in procurement conversations.
If you're self-hosting inference or using a provider that lets you pin region (Together, Fireworks, Replicate, Modal, and the smaller EU providers like Berget and Mistral's La Plateforme), region selection becomes an actual lever. Nordic and French regions are the obvious wins; Quebec is a strong North American option. The latency cost from the US to Stockholm is real (~100–120ms RTT from US East) but for most non-interactive workloads — batch summarization, embedding generation, async agent runs — it's irrelevant. Interactive chat is where it hurts.
For teams building agent systems where a single user request fans out into dozens of model calls, the regional multiplier compounds hard. A 50-call agent trace at 8.5x per call isn't 8.5x the carbon of a single-call baseline — it's an order-of-magnitude different operational footprint, and it shows up on whatever ESG dashboard your enterprise customers are starting to ask about. If you sell to European buyers in particular, expect this question in your next security review. It's already in some.
The larger shift this foreshadows is carbon-aware routing as a product feature, not a sustainability footnote. Expect inference providers over the next 12 months to start offering explicit "low-carbon region" or "follow-the-sun green routing" options the way CDNs offer edge selection today — partly from genuine conviction, partly because EU regulation (CSRD, the AI Act's energy disclosure provisions) is going to force the disclosure anyway. The 8.5x number is uncomfortable precisely because it's measurable, reproducible, and hard to offset-wash. That's exactly the kind of number that ends up on a compliance form.
This is true of ~everything that uses electricity, not just AI https://app.electricitymaps.com
That's why Mistral is even more important than just being the only competitive european LLM builder : it's on the French grid.
The reasoning effort knob is becoming a really important variable regarding how much impact AI models will have on energy consumption. Medium and max can differ by as much as two orders of magnitude in their token consumption. We also have to consider the amplification of compute demand that occurs
Great writeup! I've done some work on Codecarbon [1] in the past, it's a pretty standard tool for measuring ML/AI workload emissions, it would be awesome to have your findings contributed to that project as well.[1] https://github.com/mlco2/codecarbon
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
For 2025 the grid carbon intensity in Texas was 363gCO2eq/kWh, while it was 22gCO2eq/kWh in Sweden, so 16,5x more in Texas. Right now it's 427 in Texas vs. 50 in Finland ie. 8,54x more, like in the article (yes I looked the wrong country at first).The reason is half of Texas electrici