In November 2026 three factors matter:
1. Models
2. Data
3. Harness
Earlier this year models seemed like the only thing that mattered. But now it seems the vast majority of usecases will involve commoditized models, both in terms of cost as well as access. On some tasks model quality still matter, first and foremost coding.
Andrew Ng was right that data is the new oil. But like the world itself, the internet has shifted to be more closed-off. It's unclear what businesses will be able to continually generate useful proprietary data.
Harnesses face pressures from model providers both in terms of cost (hence inference/inhouse training booms) as well as direct competition (why use a wrapper company when I can use Claude for X?). Key question is: what % of business processes can be captured in closed-form?
(GTM is yet another completely separate factor.)