In 2025, Satya Nadella said models were commodities and Microsoft owned the toll booth. In 2026, he says the models are eating everyone's moat. What changed wasn't his mind. It was contract.
If models are interchangeable commodities, knowledge leakage to a model provider is a non-event — switch providers, nothing lost.
The Reverse Information Paradox, the centerpiece of Nadella's July essay, only exists if frontier models accumulate durable, non-substitutable advantage from what you feed them. We have seen lots of example of this, i say this should be surprise also!
Aggregator View(2025)
2025 position was the confident aggregator's position. Frontier labs build the engine; Microsoft owns the car, the road, and the toll booth. Nadella dismissed AGI milestones as benchmark theater and said the real test was GDP growth. Even the admission that Microsoft's own MAI models sat "off-frontier" was framed as strategy: being slightly behind is cheaper, and sufficient, when you control distribution.
key thing is OpenAI was contractually locked to Azure. Commoditization talk is natural when you hold exclusive distribution of the thing being commoditized.
Realization ( 2026)
Three essays in seven months. December's "Looking Ahead to 2026" — spectacle versus substance, models to systems.
June's "A frontier without an ecosystem is not stable" — human capital and token capital, the warning that a handful of models could hollow out entire industries the way globalization hollowed out manufacturing towns.
July's "The Reverse Information Paradox" — Arrow inverted: the buyer of intelligence, not the seller, now bears the information-leakage risk.
Suggestion is private evals, private RL environments on your own traces, queryable institutional memory, and an orchestration layer decoupled from any single model.
This should sound familiar. Every one of those prescriptions was standard engineering 101 practice before the philosophy arrived.
Models are the commodity
