Inference provider

When a strong open-weights model is released, anyone with accelerators can serve it — and a competitive market of specialists does: companies like Groq, Together AI, Fireworks, Cerebras, and DeepInfra offer the same popular models at widely varying prices and speeds, some on custom hardware advertising output rates far beyond GPU serving.

Identical weights do not guarantee identical service. Providers differ in quantization level, context length actually supported, sampling defaults, and reliability, so the cheapest listing on a price table can under-deliver on quality or uptime — worth verifying with a quick eval before routing production traffic.

This market is where AI price competition is most direct, since the model itself is not the differentiator. Prices for popular open models erode quickly after release, and aggregator routers make switching providers nearly frictionless, which keeps margins thin and per-token prices honest.