Independent benchmarks for sovereign AI.
SovBench compares providers that serve large language models from data centers in the European Union on speed, reliability, price and data sovereignty, and evaluates how well models handle German.
About SovBench
Many companies and public bodies in Europe may only use AI services that process data in the EU and comply with the GDPR. SovBench looks at this market independently: which providers offer LLM APIs from European data centers, what they cost, how fast and reliable they are, and who owns them.
SovBench is a project of Laps Ventures UG (haftungsbeschränkt) in Munich and is independent of the providers it covers.
Measured from Germany
Latency depends on where it is measured. Global benchmarks often run from data centers in the US, which adds a transatlantic round trip to every European provider. SovBench measures from a server in Germany, so the numbers reflect what users in Central Europe experience.
Metrics
- Latency
- Time to first token and end-to-end response time of streamed requests.
- Throughput
- Output speed in tokens per second.
- Reliability
- Availability and error rate of each API endpoint.
- Price
- Price per million input and output tokens, including how efficiently a model’s tokenizer handles German text.
- Data sovereignty
- Headquarters, ownership, data center location, certifications and data retention terms of each provider.
- German language
- How well open-weight and proprietary models handle German business documents such as contracts and official letters.
Providers
SovBench covers European providers of LLM inference, including:
- IONOS
- STACKIT
- Scaleway
- OVHcloud
- Mistral AI
- Cortecs
- EUrouter
It compares open-weight models such as gpt-oss, Llama and Qwen across the providers that offer them.