Skip to content
NuGenIT

Solution

We are weighing model APIs against running our own

Provider invoices have grown past the point where self-hosting on dedicated accelerators is worth costing properly.

There is a volume at which paying per token stops being the cheaper option and starts being the expensive one. Below it, a hosted API is almost always the right answer — no infrastructure to run, no capacity to hold, no on-call. Above it, the same workload served on dedicated accelerators using an open-weights model can cost a fraction of the invoice.

The difficult part is that the break-even is specific to your traffic. It moves with request volume, prompt and output length, how spiky demand is, whether requests can be batched, and how much engineering time you are prepared to spend keeping a serving stack healthy. Working it out properly is a modelling exercise, not a matter of opinion — and it is worth doing before the migration rather than during it.

What is usually going on

The comparison is made on headline token price alone

Self-hosting does not cost the accelerator rate — it costs the accelerator rate divided by how many requests you can actually serve per hour on it, plus the engineering time to keep it running. Both sides need the full number.

Idle hours are ignored in the model

A hosted API charges nothing when nobody is using it. A reserved accelerator charges the same at 3 AM as at peak. If demand is spiky, that gap can erase the saving entirely.

Quality is assumed rather than measured

Moving to an open-weights model is a quality decision as much as a cost one. Without an evaluation set that reflects your actual traffic, the comparison is not a comparison.

Serving throughput left at defaults

How many concurrent requests one accelerator handles varies enormously with the serving engine and its configuration. This single number decides whether self-hosting is cheaper at all.

The operational cost is left out

Somebody has to patch it, watch it, scale it and be paged for it. That is a real and recurring cost, and it belongs on the self-hosted side of the ledger.

What we look at, in order

  1. 01 Establish current spend and traffic shape — volume, token lengths, concurrency, how spiky demand actually is
  2. 02 Measure achievable throughput per accelerator for candidate models and serving configurations
  3. 03 Model both options end to end, including idle time, engineering effort and on-call
  4. 04 Identify which traffic should move and which is better left on a hosted API
  5. 05 Design the migration so it is reversible, and can run both paths side by side while you compare

What you receive

  • Break-even analysis with your traffic profile and the assumptions stated
  • Recommended serving architecture and expected throughput per accelerator
  • A traffic-split recommendation rather than an all-or-nothing answer
  • A reversible migration sequence, with the fallback path defined

Delivered as a written assessment. More on how we work.

Frequently the honest answer is a split — the high-volume, predictable traffic self-hosted and the rest left on a hosted API. If the numbers say stay where you are, that is what the assessment will say.

Tell us what you are trying to solve.

Describe your current setup and the problem in your own words. We reply within 2 working days with an honest assessment of whether we can help.