Public transcript

Microsoft CEO Satya Nadella on AI's Business Revolution: What Happens to SaaS, OpenAI, and Micros...

By: Zenvora Productions
Length: 32 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This fireside chat at Davos features Microsoft CEO Satya Nadella talking with David Sacks about how AI is reshaping knowledge work, Microsoft's Copilot strategy, and the company's relationship with OpenAI. Nadella argues that AI form factors will compose together rather than converge on one, that models will proliferate like the database market, and that AI's real payoff depends on diffusion and intense use across the economy. He also discusses restructuring at LinkedIn, local models on PCs, and why Microsoft remains committed to college hiring.

Key points

Nadella traces coding's evolution from next-edit suggestions to chat, actions, and autonomous agents, and says all form factors get used in parallel rather than one winning.
He favors the metaphor of AI as a 'manager of infinite minds' — you macro-delegate and micro-steer, with identity and permissions being the key things to get right for digital employees.
Microsoft introduced Agent 365 to give agents identities, extending the identity and endpoint protection it has for humans and their devices.
At LinkedIn, Microsoft combined product managers, designers, and front-end and backend engineers into 'full stack builders,' a structural change Nadella calls the biggest in knowledge work since PCs.
Nadella says Microsoft's strategy is to build 'token factories' in Azure and an app-server business with Foundry, and that any application will orchestrate multiple models rather than use just one.
He predicts as many models as there are firms in the world, with companies embedding tacit knowledge into weights they control, and compares the model market to the database market's proliferation including open source.
On OpenAI, he confirms Microsoft has the IP and pushes back on the idea that it created a competitor, framing the deal around Azure's token-factory ambitions.
He cites economist Diego Comin's work on the industrial revolution to argue that countries get ahead by adopting the latest technology and building on top of it, which is why diffusion matters more than invention.
Microsoft is committed to making the PC a great place for local models, with a Phi Silica model resident on NPUs and the possibility of distributed model architectures changing hybrid AI.
Nadella remains a believer in college recruiting, arguing AI steepens the productivity curve of new hires and that Microsoft is experimenting with apprenticeship cohorts attached to senior developers.

Questions it answers

00:00Opening and immigration story

How did Satya Nadella end up in the United States?

He came for grad school, married his college sweetheart from India, and had to give up his green card and get an H1 visa so she could join him, since she couldn't come after they married.

Nadella gave up his green card at the US embassy in Delhi to get an H1 for his wife.
The conversation then turns to Copilot's lukewarm early reception on Windows.

02:22Copilot form factors

How does Nadella think about the different AI form factors for knowledge work?

Using coding as the example, he traces the journey from next-edit suggestions to chat, actions, and autonomous agents, and notes that all form factors get used and composed in parallel rather than one winning.

Coding evolved from next-edit to chat to actions to autonomous agents.
Foreground, background, cloud and local agents all compose together.
The metaphor for the AI age is a 'manager of infinite minds' — macro-delegate, micro-steer.

04:52Agent identity and Agent 365

How will digital employees get credentials and accountability?

Microsoft introduced Agent 365 to give agents identities, extending the identity and endpoint protection humans and their devices already have, so organizations can trace who did what to whom.

Agent 365 extends human identity and endpoint protection to agents.
Provenance — who did what to whom — is the most important organizational query.
Two modalities: infinite minds attached to a human's identity, or agents with separate identities.

07:43Flat headcount and restructuring

How did Microsoft add $90 billion of revenue with the same headcount?

Nadella points to structural change: at LinkedIn, PMs, designers, front-end and backend engineers were combined into full stack builders with larger scope, changing both the work and the workflow, alongside a new eval-to-science-to-infrastructure loop for AI products.

LinkedIn merged four roles into full stack builders.
AI product development follows a new eval, science, infrastructure loop.
Hot-patching Windows quality and building evals must both be first class.

11:16Competition and brand permission

How does Nadella view the intense new competition?

He welcomes a completely new set of competitors every decade as it keeps Microsoft fit, notes tech's share of GDP will be higher in five years, and stresses understanding what customers uniquely expect from Microsoft's brand rather than treating everyone as a competitor.

New competitors each decade keep you fit.
The TAM is getting much bigger, so it's not zero sum.
Brand permission defines what customers expect from Microsoft.

12:47Diffusion and global south

Why does Nadella emphasize diffusion of AI rather than just invention?

Citing Diego Comin's study of the industrial revolution, he argues countries get ahead by bringing in the latest technology and building value-add on top, so AI must spread into healthcare, financial services and the public sector, with the global south gaining GDP points from public-sector efficiency.

Benefits come only from intense use, not creation alone.
Cloud and mobile rails make diffusion faster than past general-purpose technologies.
Global south public sectors could gain a couple points of GDP growth.

20:03OpenAI, token factories and models

Did the OpenAI deal create a competitor, and where is Microsoft's own model?

Nadella confirms Microsoft has the IP, but frames strategy around building token factories in Azure and an app-server business with Foundry, predicting applications will orchestrate multiple models and that there will be as many models as firms, like the database market.

Azure's TAM requires being fantastic at token factories with heterogeneous fleets.
Foundry is the app server for the agent era.
Orchestrating role-assigned models beats a single frontier model.
Models will proliferate like databases, including open-source frontier models.

24:22Local models and the PC

Is Microsoft building an LLM for the Windows desktop?

Yes — a Phi Silica model already runs resident on NPUs, the workstation form factor is back, and Microsoft is committed to local models doing prompt processing and calling into the cloud, with distributed model architectures a potential breakthrough for hybrid AI.

Phi Silica runs completely resident using NPUs and GPUs.
The workstation is back as a form factor.
One architecture tweak from distributed models could change hybrid AI.

Notes

Immigration story and early framing

Nadella recounts giving up his green card and getting an H1 so his wife could join him in the US, calling it a workaround for immigration policy.
The conversation opens with a question about Copilot's lukewarm early reception on Windows and the gap between playing with chatbots and getting business results.

Form factors for knowledge work

Coding shows the full journey: next-edit suggestions, then chat, then actions, and now autonomous agents that run foreground, background, in the cloud or locally.
All form factors compose — Nadella describes running a CLI agent, a background agent, and edits in VS Code in parallel.
His preferred metaphor for the AI age, borrowed from Notion's CEO, is a "manager of infinite minds": you macro-delegate and micro-steer.
Agent 365 gives agents identities, extending human identity and endpoint protection to them; provenance — who did what to whom — is the key organizational query.

Structural change in knowledge work

Nadella calls this the biggest change since PCs, when work artifacts and workflows both changed.
At LinkedIn, PMs, designers, front-end and backend engineers were combined into "full stack builders" with increased scope.
Building AI products creates a new loop: evals, science, infrastructure — while hot-patching Windows must remain first class.

Strategy: token factories and models

Microsoft's biggest business is Azure, so it must be fantastic at building "token factories" with a heterogeneous fleet for TCO and utilization.
Foundry is the app server for the agent era; every application will orchestrate multiple models, not one.
He compares models to the database market: closed frontier models, open-source frontier models, and eventually as many models as firms, each embedding tacit knowledge in weights.
On OpenAI: Microsoft has the IP, and the deal serves the token-factory strategy.

Diffusion and geopolitics

Citing Diego Comin's study of the industrial revolution, Nadella argues countries advance by adopting the latest tech and building on it.
Success is measured by market share and ecosystem effects — employment, channel partners, ISVs built around the platform country by country.
The global south has huge opportunity: efficiency gains in public-sector delivery could add points of GDP growth.

Local models and the PC

A Phi Silica model runs resident on NPUs; the workstation is back and Microsoft is committed to local models doing prompt processing and calling into the cloud.

Adoption and hiring

Adoption is both top-down (customer service, supply chain, HR self-service ROI projects) and bottom-up, like PCs when lawyers brought in Word.
Microsoft runs roughly 500 fiber operators in Azure; its network lead built digital employees to handle fiber-cut DevOps back-and-forth.
College hiring continues: AI steepens new-hire productivity curves, and Microsoft is experimenting with apprenticeship cohorts attached to senior ICs.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.