AI & Automation

Build, Fine-tune, or Buy: the Model Decision as a Capital Decision

Treat the model decision the way you would treat buying a machine: an asset with a cost, a life, and a payback period. The life is set by someone else's release schedule, and it is much shorter than anyone plans for.

  • Advanced
  • 15 min total
  • 15 chapters

What decision this helps you make: Whether to call an API, engineer the context around a general model, fine-tune your own, or run inference on hardware you control, and what evidence would have to exist before each step is worth the capital.

What this topic is

There are four positions on one ladder, not two options. Call a general model through an API. Engineer around it (retrieval, structure, validation) without touching the weights. Adapt a model to your task by fine-tuning. Or run inference on capacity you control. Each rung costs more capital, takes longer to reverse, and buys a different thing. Framing it as a capital decision means asking the questions you would ask of any asset: what does it cost, what does it save, how long does it last, and what is the alternative doing while we own it?

Why it matters

The alternative to building is not static. It gets cheaper and better every quarter, on a schedule you do not control, which means a model asset depreciates against a moving benchmark. The depreciation is often faster than the payback. That single asymmetry explains most of the disappointing model investments of the last few years: they were evaluated as though the buy option would stay where it was, and it did not. Get the life estimate right and the rest of the arithmetic is ordinary capital budgeting.

Who should learn it

Founders and engineering leaders deciding where to spend a scarce quarter, finance leaders being asked to approve model spend, and operators weighing a vendor's proposal to build something bespoke.

What you will understand

  • The four rungs of the ladder, what each buys, and the evidence that justifies stepping up
  • How to compute the payback on a fine-tune, and the life estimate that usually kills it
  • The break-even utilisation that decides whether self-hosting beats an API
  • How to read a provider agreement for the five clauses that decide your exposure

Prerequisites

Common misconception

"Building our own model is the way to own our AI capability." It is one way to spend capital on an asset with an unusually short and externally determined life. Ownership of a model is not the same as ownership of an advantage: the thing customers pay for is a job done reliably, and the assets that deliver that sit above the model rather than in it: the data, the workflow, the evidence that the output is right. The right question is never "should we own a model?" but "what does this rung buy that the rung below it cannot, and how long will that remain true?"