Most of the LLM traffic is handled by the hyperscalers (GCP, Azure, AWS) and we can only guess where they host their models.
Based on strategic investments and agreements, we assume that:
- Anthropic runs mainly on AWS: Anthropic has established AWS as its "primary cloud and training partner" following Amazon's $8 billion investment
- OpenAI runs mainly on Azure: While historically Azure was "the exclusive cloud provider for all OpenAI workloads", Microsoft is no longer OpenAI's exclusive cloud provider as of January 2025, though the OpenAI API remains "exclusive to Azure"
Let's take a look the hypescalers and their respective AI hosting services with a focus on european regions.
Google (GCP): Vertex
- Anthropic Claude Sonnet 4
- 🌿 europe-west1 (Belgium, Brussels, 82% CFE%)
- Anthropic Claude Opus 4
⚠️ us-east5 (USA, Columbus, 52% CFE)
- Google Gemini Pro 2.5 (same for other Gemini models)
- 🌿 europe-north1 (Finland, Helsinki, 98% CFE%)
- 🌿 europe-west9 (France, Paris, 94% CFE%, ⚛️)
- 🌿 europe-west1 (Belgium, Brussels, 82% CFE%, ⚛️)
- 🌿 europe-west4 (Netherlands, Amsterdam, 80% CFE%)
- europe-southwest1 (Spain, Madrid, 76% CFE%)
- europe-west8 (Italy, Milan, 52% CFE%)
- europe-central2 (Poland, Warsaw, 31% CFE%)
- Mistral Small, Large, Codestral, OCR
- 🌿 europe-west4 (Netherlands, Amsterdam, 80% CFE%)
CFE based on Google data: Carbon data across Google Cloud regions
Amazon (AWS): Bedrock
-
- 🇸🇪 Sweden (Stockholm), 🌿 35.8 gCO2/kWh
- 🇪🇸 Spain 🌿 146.2 gCO2/kWh
- 🇮🇪 Ireland, 279.7 gCO2/kWh
- 🇮🇹 Italy (Milan), 287.5 gCO2/kWh
- 🇩🇪 Germany (Frankfurt), 344.1 gCO2/kWh
-
US East (Ohio, N. Virginia)US West (Oregon)
All European are only available with Cross-Region-Inference. Which means that requests might be routed to another region while making sure they stay in the same geography (i.e. EU).
All European Antropic endpoints are only available with Cross-Region-Inference. Which means that requests might be routed to another region while making sure they stay in the same geography (i.e. EU).
- Some models are accessible in some Regions only through cross-Region inference. Cross-Region inference allows you to seamlessly manage unplanned traffic bursts by utilizing compute across different AWS Regions. With cross-Region inference, you can distribute traffic across multiple AWS Regions. To learn more about cross-Region inference, see Increase throughput with cross-Region inference and Supported Regions and models for inference profiles.
Cross-Region inference requests are kept within the AWS Regions that are part of the geography where the data originally resides. For example, a request made within the US is kept within the AWS Regions in the US. Although the data remains stored only in the source Region, your input prompts and output results might move outside of your source Region during cross-Region inference. All data will be transmitted encrypted across Amazon’s secure network.
Microsoft (Azure): AI Foundry
- Mistal Small, Large, Codestral, OCR
- 🇸🇪 Sweden (swedencentral), 🌿 ⚛️ 35.8 gCO2/kWh
- OpenAI GPT-4.1, GPT-4o, o4-mini, o3
- 🇳🇴 Norway (norwayeast), 🌿 30.7 gCO2/kWh
- 🇸🇪 Sweden (swedencentral), 🌿 ⚛️ 35.8 gCO2/kWh
- 🇨🇭 Switzerland (switzerlandnorth), 🌿 ⚛️ 36.6 gCO2/kWh
- 🇫🇷 France (francecentral) 🌿 ⚛️ , 44.2 gCO2/kWh
- 🇪🇸 Spain (spaincentral), 🌿 146.2 gCO2/kWh
- 🇬🇧 UK (uksouth), 210.9 gCO2/kWh
- 🇳🇱 Netherlands (westeurope), 253.3 gCO2/kWh
- 🇮🇹 Italy (italynorth), 287.5 gCO2/kWh
- 🇩🇪 Germany (germanywestcentral), 344.1 gCO2/kWh
- 🇵🇱 Poland (polandcentral), 615.0 gCO2/kWh
Sources:
- Serverless Model Regions (Mistral and others)
- Azure OpenAI Model Regions
- Choosing the right model
- Region list
Data Source & Methodology: Carbon intensity values are sourced from Carbon intensity in electricity generation: 2024 representing grams of CO₂ equivalent per kilowatt-hour (gCO2eq/kWh) for 2024. The ⚛️ symbol indicates countries with high nuclear power share (>25% of electricity generation).
🌿 = Sustainable regions (<200 gCO2/kWh)
⚛️ = High nuclear power share (>25%)