Satya Nadella argues a platform wins by creating more value around it than it captures, so any company should be able to build its own AI on Microsoft's stack.
Nadella says each company will need private evaluations, which he calls potentially its most important intellectual property, since models learn from few samples and switching models is key.
Nadella says data center build-out must deliver tangible community benefits like jobs, tax base, energy prices and replenished water within 12 to 18 months, or the industry will not earn permission.
Nadella says AI winners will be ecosystems, not single models
Microsoft is betting that AI is an ecosystem, not one model or one vendor. A platform earns its place by creating more value around it than it captures, so any company should be able to build its own AI on the stack. Durable value then sits in what each company owns, such as private evals and traces, and in earning public permission for the build-out.
Takeaways
A platform must create more value for others. Nadella defines a platform as creating more value around it than it captures. Any company, AI native or traditional, should build its own AI on Microsoft's stack. A developer conference promoting one model would make no sense. Microsoft added more Azure capacity in fifteen months than in its first fifteen years.
Private evals may be a company's biggest IP. Models learn from few samples, so network effects around intelligence work differently. Each company's private evals, traces and freedom to switch models become the key assets. In the Land O'Lakes demo, GPT-5.5 traces trained a 5B reasoning model that scored higher.
Coding success forces new agent-centered interfaces. A user may run around 100 agent sessions, and the cognitive load demands a new coding interface. Open harnesses let companies combine their own tools, models and context. Nadella says Emdash found bugs that Mythos did not.
SaaS endures as data models get rebundled. Application data models and business logic stay valuable and can be unbundled and rebundled, as with Power BI semantic models and WorkIQ. Agent usage may exceed human usage, forcing mailbox rebuilds. Per-user pricing stays, with consumption metering added.
Generalists gain the most leverage in engineering. Elad Gil relays a view that engineering narrows to four roles within four or five years. Nadella points to LinkedIn's full stack builder discipline and hard infrastructure work such as reinforcement learning environments for Excel, and expects generalists to gain most.
Real ambition means asking for tokens, not headcount. Azure's networking team built an agentic system called Miles to manage its 500-plus fiber operators and asked for tokens instead of people. Nadella says organizations lack a new conceptual model of what they can build, echoing Kevin Scott on making the impossible possible.
Data centers must earn community permission. Nadella says the build-out needs real benefits such as jobs, tax base, lower energy prices and replenished water through closed loop systems. He predicts people must see these benefits within twelve to eighteen months, and that the world will doubt tech companies asking for trust.
Token capital and education
Tap a name to go there · tap a line to see why they are linked
On this episode 4
Show 1
Nadella says every company will hold both human capital and token capital, and the task is to compound the two. He and swyx discuss putting human capital on the balance sheet through agent traces, and swyx expects the SEC to need accounting standards for token expertise.
Nadella cites a Stanford AI class that made sure students learned to apply softmax appropriately, as evidence that concepts stay critical as AI changes education. He suggests the next big startup might build a new university or pedagogy.