Public transcript

Untitled recording

Length: 57 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Satya Nadella, CEO of Microsoft, speaks to a Stanford class during finals week about the thinking behind Microsoft's early OpenAI bet and the company's broader AI strategy. He covers the Build announcements including the MAI model family, the frontier intelligence ecosystem, Scout agents, custom silicon, quantum computing, and new agent-era hardware form factors. He also takes student questions on culture, growth mindset, career advice, and open versus licensed models.

Key points

Microsoft's OpenAI bet came from a 'prepared mind' shaped by a long obsession with natural language, plus the appeal of the scaling laws paper and OpenAI's ambition on transformers.
The MAI models were trained on clean data lineage so they can be licensed with weights, letting every company build its own 'hill-climbing machine' with RL environments and private evals.
Scout is described as the third Copilot form factor, an autopilot-style long-running agent with its own identity, sandbox, and heartbeat, like an enterprise OpenClaw.
Microsoft is working with the OpenClaw Foundation on secure containment, including an MXC sandbox container on Windows and options like Windows 365 cloud instances for isolation.
New silicon partnerships with NVIDIA include a Surface Laptop on the new RTX SoC, a dev box with a petaflop of AI compute and 128GB unified memory, and Windows on a GB300 DGX workstation.
Maia 200 is running GPT 5.5 in multiple data centers powering Copilot, and Cobalt is an ARM processor optimized using GitHub Copilot traces for the agentic loop.
On quantum, Microsoft has Majorana 2 built at industrial scale and claims a quantum computer solving real problems by the end of the decade.
Nadella says growth mindset is about having the courage to confront your own fixed mindset, influenced by nonviolent communication and Carol Dweck's work.
The MAI lineage models will be licensed rather than open weight, while smaller models like Ion Instruct and Ion Plan for Windows will be open weight.

Questions it answers

00:00The OpenAI bet

Why did Microsoft invest a billion dollars in OpenAI in 2019?

Nadella describes a prepared mind: Microsoft had a long obsession with natural language, and the scaling laws paper plus OpenAI's ambition on transformers made the bet appealing. Microsoft also bought companies and invested broadly in anyone with an ambitious natural language angle, and there was no internal uprising because Microsoft is conditioned to ecosystem partnerships.

The big decision was compute concentration on one effort, not just capital.
The capability graph has stayed on the scaling law Pareto frontier.

05:00Frontier ecosystem and MAI models

Why did Microsoft launch seven new models and what is the frontier intelligence ecosystem?

The vision is that every company can operate at the frontier with its own IP compounding, not just human capital but token capital. The models were trained on a clean data lineage with no synth data so reasoning emerged, and they can be licensed with weights so companies can build their own hill-climbing machines using their own traces and private evals.

Companies just consuming a foundation model may not retain enterprise value.
Microsoft 365 can bootstrap RL environments and generate evals from observed business processes, with data owned by the company.

11:00Scout, agents, and security

What is Scout and how will Microsoft secure long-running agents?

Scout is the third Copilot form factor after chat and Cowork: a long-running autopilot agent with a heartbeat that can take a delegated Entra ID and act as a digital twin, and users can mint more autopilots with their own identities and sandboxes. Microsoft is working with the OpenClaw Foundation on secure containment, including an MXC sandbox container on Windows and options like Windows 365 cloud instances.

Agents will need process, session, and container isolation boundaries like we once thought about processes.
Scout is described as an enterprise OpenClaw that fits with the Copilot system.

14:54Edge compute and new form factors

How is Microsoft bringing AI to consumer devices?

Microsoft wants to tap the large dGPU install base for unmetered intelligence, with a Surface Laptop on NVIDIA's new RTX SoC coming in the fall, a dev box with a petaflop of AI compute, 20 CPU cores, and 128GB unified memory running a trillion-parameter model locally, plus Windows on a GB300 DGX workstation. Project Solara showed two reference designs, a badge and a desk companion, as endpoints for long-running agents.

Running agents 24/7 without billing is best done on your own laptop or desktop.
Windows will stay an open platform with new platform rules for the agent era.

20:00AI's value to the world

What is AI's equivalent of selling light rather than electricity?

Nadella says the tech bubble overhypes the technology for its own sake, but the world will evaluate it by value created one community at a time, like AI changing the cost equation of healthcare for someone in your family. He argues humans are the most adaptive species and will create new value on top of commoditized intelligence, and that the ecosystem must be positive-sum or society will lose social permission.

If a few firms capture all the returns, social permission will be lost.
Displacement is real, but new economic activity with human agency and wages will emerge.

23:40Custom silicon

What is Microsoft doing in custom silicon versus other hyperscalers?

Microsoft is motivated by three new workloads: training, inference, and the long-running agent loop that uses inference plus regular compute. Maia 200 is codesigned with Microsoft and OpenAI models and is running GPT 5.5 in multiple data centers powering Copilot, while Cobalt is an ARM processor optimized using GitHub Copilot traces; old GPUs in the fleet accelerate Fabric with 7x-plus gains.

The fleet is heterogeneous, with software doing smart workload placement.
Nadella calls it a great time to be in computer architecture, from data center physical design to electron efficiency.

27:55Quantum computing

What is the state of Microsoft's quantum program?

Microsoft works with partners on ion trap, photonics, and natural atom machines, including a quantum computer powered by Atom Computing within the year, and quantum traces already train material science models. For fault tolerance, Majorana 1 proved the physics of the Majorana state of matter and Majorana 2 builds it at industrial scale with perfected digital control; Microsoft claims a quantum computer solving real problems by the end of the decade.

Quantum is an accelerator that must be married with classical computing.
100 logical qubits with good error correction could generate synth data for science models in the shorter term.

32:40Culture, growth mindset, and advice

How does Nadella think about culture and what advice does he give students?

Growth mindset is not instilled but invoked, and the key is having the courage to confront your own fixed mindset, influenced by nonviolent communication and Carol Dweck's work. For students, learning has become easier with personalized tutors, so he would use agents for cognitive coverage across many subjects while still understanding what they produce, and he would consider going into hardware given the moment in computer architecture.

Growth mindset is practiced for yourself, not as a corporate mandate.
Somebody will break a new pedagogy around agent-assisted learning.

Notes

The OpenAI Bet

Nadella describes a "prepared mind": Microsoft was long obsessed with natural language, though before 2017-2018 it assumed symbolic logic plus machine learning would be needed.
Beyond OpenAI, Microsoft bought companies and invested broadly in anyone with an ambitious natural language angle.
The scaling laws paper and OpenAI's ambition on transformers with more compute and data made the bet appealing; the capability graph has stayed on the scaling law Pareto frontier.
There was no internal uprising: Microsoft is conditioned to ecosystem partnerships (Gates-Grove model, SAP and SQL Server), and the big decision was compute concentration.

Build Announcements and MAI Models

Seven new models were launched with a clean data lineage, no synth data, so reasoning emerged and the weights can be licensed.
The vision is a frontier ecosystem where every company operates at the frontier with its own IP compounding, including "token capital."
Every company should think strategically about its RL environment, private evals, and letting models hill-climb on its own traces without leaking value.
Microsoft 365 customers can bootstrap RL environments and even generate evals, e.g., for an HR onboarding process, with data owned by the company.

Scout and Agent Form Factors

Copilot evolution: chat (thinking assistant), Cowork (delegated tasks via multi-step agentic loops), and Scout (autopilot, long-running agents with a heartbeat).
Scout can take a delegated Entra ID, act as a digital twin, and mint more autopilots with their own identities and sandboxes.
Windows will offer an out-of-box OpenClaw experience sandboxed in MXC containers, with process, session, or VM isolation boundaries.

Silicon, Edge, and New Form Factors

Maia 200 is codesigned with Microsoft and OpenAI models and is running GPT 5.5 in multiple data centers powering Copilot; Cobalt is an ARM CPU optimized with GitHub Copilot traces.
NVIDIA partnership: Surface Laptop on the new RTX SoC in the fall, a dev box with a petaflop of AI compute and 128GB unified memory running a trillion-parameter model locally, and Windows on a GB300 DGX workstation.
Project Solara showed two reference designs: a badge with fingerprint reader, camera, and MediaTek processor, and a desk companion, as endpoints for long-running agents.

Quantum

Majorana 2 builds on the Majorana state of matter physics breakthrough at industrial scale, with perfected digital control.
Near term: natural atom machines with partners, including a quantum computer powered by Atom Computing within the year; quantum traces can train material science models.
Claim: by the end of the decade, a quantum computer solving real problems; 100 logical qubits could generate synth data for science models sooner.

Culture and Advice

Growth mindset means confronting your own fixed mindset; influenced by nonviolent communication and Carol Dweck, and Herbert Simon's bounded rationality.
For students: use agents for "cognitive coverage" across many classes while still understanding what they produce; grades and assignments may be relitigated.
MAI lineage models will be licensed with weights (available via Base10, Fireworks), not open weight, for safety and inspection; smaller Ion Instruct and Ion Plan models will be open weight on Windows.

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