20VC: Satya Nadella on Open vs Closed Models: Who Wins | Whether to Trust Frontier Model Providers | Why Companies Need to Own Their Own Intelligence | The Importance of Verticalisation: From Chips to Models and More · podgraph
20VC: Satya Nadella on Open vs Closed Models: Who Wins | Whether to Trust Frontier Model Providers | Why Companies Need to Own Their Own Intelligence | The Importance of Verticalisation: From Chips to Models and More
Satya Nadella says Microsoft will license its MAI models as weights rather than fully open because he wants a simple software business, not a network effects intelligence business.
Nadella argues companies must own their intelligence through context, fine-tuning on their own traces, and private evaluation benchmarks that every model must beat, rejecting a single vertical stack winner.
Nadella advises treating frontier models within a zero trust architecture that contains malfunctioning agents, since alignment proved harder than he assumed and chain of thought is unfaithful.
Nadella says companies must own their intelligence, not rent it
Satya Nadella argues that AI capability will spread across many competing models rather than concentrate in one vertically integrated winner. Because models can malfunction and most usage will become cheap commodity tokens, firms should contain models with zero trust controls and keep their own knowledge, data and evaluations. The value to chase is faster economic growth, not control of the stack.
Takeaways
Treat models as fallible and contain them. Nadella says alignment has not fallen out of scaling as he once expected, so the best posture is zero trust architecture. Microsoft's Agent 365 gives each agent an identity, logs its tool and model calls, and enforces policy through behavioural monitoring.
Most token usage will become commodity usage. Nadella expects a lasting market for frontier models, but most tokens will face price competition. Microsoft's Copilot already mixes many cheaper closed models with open ones. He declined to put a number on the open versus closed traffic split in two years.
Microsoft's MAI models will be licensed, not open. Nadella wants a simple software business rather than a network effects intelligence business. He is pleased with Western open weight progress. He thinks standardised access to chain of thought, which he calls unfaithful, may push the industry toward more open weights.
Companies must keep the surplus from intelligence. Drawing on Hayek's local knowledge and Arrow's information paradox, Nadella says firms should capture value through their own context, fine-tuning on their own traces, and private evals every model must beat. He expects AI native to sound outdated within five years, as cloud did.
No single company will own every layer. Nadella rejects one firm owning everything from silicon to application. He expects five or six closed and five or six open weight models, more competition at every layer, and routers that send each task to the best model.
Unilever shows implementation becoming the new services market. Unilever built an agentic order-to-cash and trade claims system, first on a frontier model and later on an MAI model at lower cost. Its IT team now manages it. Nadella says new services blend business and technical consulting.
The real test is lifting world GDP growth. Stebbings notes growth has stayed near 2 percent a year. Nadella cites Copilot in the NHS, elastic demand for software, and Rolls-Royce maintenance scheduling. Science and engineering may prove as elastic as coding, while knowledge work is harder to reinforcement-train. Huge training budgets are bounded by power and silicon shortages.
Tap a name to go there · tap a line to see why they are linked