No shared GPUs
On the shared API your requests are isolated logically, but the silicon is shared with other customers. Dedicated inference removes that question entirely. Your model, your hardware, nothing else on the machine.
Nobody else's peak is your problem.
A shared model can be busiest exactly when you need it most. With dedicated inference nobody else is on the model. Capacity is constant, latency is predictable, and you set the priorities.
When a dedicated GPU makes sense
At low volume, paying per token is your cheapest option. Past a few billion tokens a month, reserving the GPUs costs less than the metered API. Pick a model to see where the lines cross. Smaller models start around €2,500 per GPU per month.
Hardware is the least of it.
KV-cache tuning, quantisation, throughput optimisation, monitoring, and capacity planning are all included. You keep control, we carry the toil.
Any model, including yours
A legacy model we no longer host in the shared fleet, a fine-tune you own, or something in between. If our hardware can run it, we'll deploy it for you.
Guaranteed delivery, in writing
Public endpoints are best effort. Dedicated inference with an HA setup carries a 99.5% monthly uptime commitment — in writing, with compensation if we fall short.
Read the full SLA