LLM gateways sit between an application and its model providers. They give developers a common API for accessing different models, selecting a route, handling provider failures and controlling usage. This category includes managed gateways and locally deployed routers; it excludes products whose main purpose is hosting inference, running models on a local GPU or providing an end-user chatbot.
For European teams, a gateway can provide a central place to enforce model-access policies and review how prompts move between services. European incorporation, gateway hosting and model inference are separate considerations. A European gateway can still send requests to providers outside Europe.
Teams considering services such as OpenRouter or Vercel AI Gateway can compare European alternatives by API compatibility, routing behaviour, operational control and contractual terms. Some gateways provide consolidated model access and billing; others use your existing provider accounts or run in your own environment.
Neither deployment model guarantees privacy or compliance. A hosted gateway becomes another party handling requests, while a self-hosted router still forwards prompts to the configured model providers. Review the full request path rather than relying on the location of the gateway alone.
A gateway routes requests to model services. An inference provider runs the model computation. A business may offer both, but a routing endpoint alone does not make it an inference host.
Not necessarily. The routing software can run locally while model inference happens at external providers. Local model execution requires a separate runtime and suitable hardware.
Only when the selected model endpoints, fallback policies and supporting data services meet that requirement. Check inference locations and logging destinations for the complete route.
No. Your organisation remains responsible for a lawful basis, data minimisation, access controls, retention and suitable contractual safeguards. Assess the gateway and its upstream providers against your actual use case.
A common API can reduce migration work, but provider-specific parameters and model behaviour still vary. Test representative requests, authentication, streaming and failure handling before moving production traffic.