Three decision levers
Whether a self-hosted model pays off comes down to three factors: cost (TCO), performance and control.
1. Total cost of ownership
Cloud APIs bill per token, ideal for sporadic use and prototypes. At high, steady volume the maths flips: an owned or rented GPU has predictable fixed costs that, above a certain utilisation, sit well below token prices. We model this for your concrete load.
2. Performance: are open models enough?
For most enterprise workloads: yes. With vLLM, open models from 7B to 70B achieve high throughput thanks to PagedAttention and continuous batching; tensor parallelism scales across GPUs. Not every task needs a frontier model.
3. Control and sovereignty
The often-underrated lever: a self-operated model runs in your sovereign environment, no subscription lock-in to US providers, no overnight price or model change, no US jurisdiction, no data egress. For regulated industries this isn't a nice-to-have but a prerequisite.