Union Square Ventures published a map of the AI stack this summer and named its thesis after a rebellion: value escapes the integrated giants and settles into layers, the way it escaped IBM, then Microsoft, then the carriers. Whatever the outcome of that bet, one layer is already commoditizing on schedule — the model.
Frontier models leapfrog each other quarterly. Prices fall by orders of magnitude on a cadence no procurement cycle can track. Open-weight models close the gap from below. This is what commoditization looks like from the inside — and every operator who hard-wired a single provider into their operation inherits a migration each time the leaderboard turns over.
The layer that appreciates
While models churn, one layer compounds: the harness — the interface agents speak, the policies that bound them, the evaluations that admit a model to production, the budgets that stop it, and the ledger that remembers everything it did. None of that belongs to a model provider. All of it is where an operation’s actual intelligence accumulates.
Rent the model. Own the harness.
When the harness is yours, a model is a component: admitted by evaluation, replaced by configuration, mixed by workload. The best model for drafting is not the best model for classification, and neither should be a marriage. When the harness belongs to your provider, the relationship inverts — every capability you build deepens their moat, not yours.
The tell
Ask one question of any AI platform: what happens the day we switch model providers? If the answer is a migration, you do not own your harness. Your provider does — and the switching costs you are accruing are the product they are actually selling.