Don't Rent Your Learning Loop

The model is not the moat.

The learning loop might be.

That is the useful read on Satya Nadella’s warning about enterprises “paying twice” for AI: once with money, and again with the proprietary knowledge they reveal to make the model useful. Prompts, tool calls, corrections, retries, workflow traces; that exhaust is not just telemetry. It is institutional know-how.

This is a slightly uncomfortable point because everyone wants the smartest model today. But the durable value may be in the thing you produce while using the model: examples, evals, approvals, rejected answers, domain-specific corrections, and the tiny human edits that say “this is how our business actually works.”

That is also why Hugging Face’s CEO is arguing that companies are increasingly done renting their AI. Open models are not winning every benchmark, but they are good enough for many production workloads, cheaper to adapt, and easier to own. Meanwhile, Reflection’s $1 billion compute deal shows the open-model camp still needs serious infrastructure. Openness does not make compute free.

For SI and ISV partners, the practical move is not “use open models for everything.” That is too simple.

The move is to design for portability and retained learning from day one. Keep evals outside the vendor. Log corrections in a customer-owned store. Separate orchestration from model choice. Make the feedback loop exportable, auditable, and reusable.

Models will keep changing.

The question is whether your customer gets smarter every time they do.