Refactoring Orgs

The argument for building internal AI infrastructure usually runs like this: we need control, we need customization, we need to not be dependent on a vendor. These are real concerns. They also apply to almost every piece of software a company uses, and most companies have correctly concluded that building their own database engine or their own payroll system is not the right answer.

The exception is when the capability is core to your competitive differentiation. Stripe builds its own fraud detection because payments fraud is their central problem - they have the scale, the proprietary data, and the domain expertise to justify it. For most companies at most scales, none of those conditions hold.

What makes frontier AI feel different is that it is new enough that nobody has figured out the build-vs-buy intuition yet. That uncertainty is producing expensive infrastructure bets in organizations where the real problem is a lot simpler: they have not decided which single workflow to fix first, and they haven't found out how to make money doing it.

That's a prioritization problem. It has nothing to do with compute.