Owning your AI:
costs and returns.

Comparisons of this kind are almost always unfair in one direction or the other, because each side counts only the costs the other side has. A defensible model counts all of them.

Hosted API, fully loaded. Per-token charges at your real volume including the input tokens you re-send on every turn; the retry and failure traffic that never reaches a user; the engineering time spent on rate limits and provider incidents; the migration work every time a model is deprecated; and the periodic cost of the transfer assessments and vendor reviews that a third-party processor obliges you to maintain. The last two are real and are almost never in the spreadsheet.

Self-hosted, fully loaded. Hardware amortised over a realistic life — three years is the usual assumption; power and cooling; facility or rack costs if colocated; the fraction of an engineer needed to keep it healthy, which is the largest and most frequently understated line; the initial build; and the periodic cost of upgrading the model, which does not go away just because nobody is invoicing you for it.

Put both on the same three-year footing, and the comparison becomes a straightforward one about volume. It also becomes clear that the operator-time line, not the hardware line, is what usually decides it.

The saving on existing usage is the easiest benefit to quantify and, in our experience, rarely the largest. Three others matter more and are harder to put in a spreadsheet, which is not the same as being soft.

Work that becomes possible. When marginal cost approaches zero, the calculus of what is worth processing changes completely. Reviewing every contract instead of a sample. Embedding the whole archive rather than the recent part. Running an agent across the entire backlog because there is no per-unit charge to justify. In our experience organisations discover more value here than they forecast, because a per-token budget trains people not to ask.

Data that becomes usable. The corpus that could not be sent to a third party — legal, HR, M&A, R&D, patient records, board material — is frequently the corpus where AI would be most valuable. Self-hosting does not make that data cheaper to process; it makes it processable at all. That is a category change rather than a cost reduction.

Risk that stops recurring. No third-party processor in the AI path means no annual vendor review for it, no sub-processor list, no transfer assessment, and one fewer supplier whose incident becomes your outage. Each is a modest recurring cost, and they are the ones finance never sees because they are paid in other people's time.

When volume is low or unpredictable, because idle hardware is pure cost. When there is nobody to operate it and no budget for managed operation, because an unmaintained deployment degrades and takes user trust with it. When the use case is still unproven, because you would be buying infrastructure for a system that might not survive its own pilot. And when every request genuinely needs frontier capability and none of the data is sensitive, because then you are paying for control that buys you nothing.

We say this in first meetings regularly. A client who bought the wrong infrastructure on our advice is a problem we own for years, and the arithmetic is not hard enough to justify getting it wrong.