The model is not your moat. The orchestration around it is.
Everyone has the same weights. The difference is what you wrap them in — and whether you can prove it works.
The compounding cost of treating model selection as the strategic decision: you rebuild your evaluation every three months and never accumulate anything.
Your competitors have access to the same models you do, on the same day, at the same price. That has been true for about two years and most strategy decks have not caught up.
What actually differs between two teams on the same model
The innermost layer is the one everyone argues about and the only one you cannot differentiate on.
Three columns
Orchestration decisions dominate unit cost. On a run where every model got identical work, spend ranged 338× — and cost did not correlate with time, so optimising for one actively works against the other.
Why: Routing policy, retry behaviour and concurrency are worth more to your margin than which model you picked.
Orchestration is where failures are contained. A provider outage mid-run, a rate limit, a model that silently changed behind a stable id — all of it is handled here or nowhere.
Why: We lost three full runs to provider behaviour during one census. The harness discarded them rather than publishing rates computed on survivors, and that was an orchestration decision, not a model one.
Evaluation is the part that compounds. A named failure taxonomy survives every model generation; a benchmark result survives about one quarter.
Why: It is the only asset in the stack whose value increases as the models change faster.
The two objections
"We'll just use whichever model is best." Best at what? On our panel the spread on a single failure mode runs 0% to 100% across models, and where a model lands is not recoverable from its general standing. There is no 'best' without naming the failure you care about.
"Orchestration is plumbing, not strategy." Plumbing that determines your cost, your failure containment and your ability to make a claim about your own product. If that is not strategy, nothing in the stack is.
Where I would start
Name the three failures that hurt you most. Add one control arm to one eval. Record which model actually served each request. That is a fortnight and it gives you a baseline nobody else has.
What in your AI stack would still be a differentiator if every competitor woke up tomorrow with your exact model?
Every figure here describes something measured and committed. See the measurements · read the method