There are now over 69 providers offering LLM inference. Prices range from 14 cents per million input tokens at the cheap end to 180 dollars per million output tokens at the frontier. That is a 1,280-times gap.
Most founders are somewhere in the middle without a clear reason to be there.
The calculation is somewhat straightforward once you do it.
lets take your daily active users, multiply by interactions per user per day, multiply by tokens per interaction - split into input and output, because output tokens are typically two to four times more expensive, and run that number through the per-token prices of three or four providers.
The result is often a shock.
A product with ten thousand daily active users and five AI interactions each can easily reach twenty thousand pounds per month on a frontier model, and under five hundred pounds per month on a comparable open-weight alternative.
Three things determine whether the API route makes sense.
Volume: below fifteen thousand pounds per month in API spend, the engineering overhead of self-hosting is not worth it.
Predictability: APIs absorb traffic spikes without operational overhead, and that flexibility has real value for early-stage products.
Model quality requirements: not every task needs a frontier model. Routing simple queries to smaller, cheaper models can cut inference costs by 40 to 70 percent with no user-visible quality loss.
This is how the vendor lock-in risk gets real and underestimated.
When you build on a single API provider, you are locked in not just contractually but through accumulated prompt engineering work, evaluation datasets tuned to that model’s behaviour, and observability tooling built around that provider’s response format. The exit cost from a fully-loaded enterprise AI stack has been estimated at between two hundred thousand and one million dollars in re-engineering.
The mitigation is simple: use an abstraction layer between your application code and the vendor API.
LiteLLM and Portkey both do this. It adds minimal overhead and lets you switch providers with a configuration change.
The cheapest route from day one is not the cheapest route at scale. Start tracking your cost per inference from launch.
Listen to the full podcast episode for the full breakdown.
This episode is second in the series Build vs Buy vs Rent.
see you in the next episode.
Note : APIs are sometime pronouced APees, apologies for that, I let Elevenlabs run wild.
