Public transcript

Microsoft Build 2026 | Satya Nadella Opening Keynote

By: Microsoft
Length: 2 h 23 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 opens Microsoft Build 2026 with a sweeping tour of the AI stack, from Windows edge devices like the Surface RTX DevBox and new local models, through Azure data center and silicon innovations including Fairwater, Cobalt 100 and Maia 200, to Project Solara agent-first devices. He then moves up the stack with Microsoft IQ (Web IQ, Fabric IQ, Work IQ), MXC execution containers, Open Claw on Windows, the new GitHub Copilot app, Rayfin SDK, Agent 365 governance and MDASH security. Mustafa Suleyman announces seven new MAI models and frontier tuning with reinforcement learning environments, including a Mayo Clinic healthcare partnership, before Nadella closes with Microsoft Discovery and the Majorana 2 quantum announcement, framing everything around building a frontier ecosystem that spreads opportunity rather than concentrating power.

Key points

Satya Nadella frames the conference around participating fully in the frontier intelligence ecosystem, spanning a ubiquitous compute fabric, a models/context/tools layer, agent runtimes, tooling, and security and governance.
Microsoft announces the Surface RTX DevBox with 1 petaflop of AI compute and 128 GB of unified memory, plus new local models Ion Instruct and Ion Plan for fully local agentic loops on Windows.
Azure now spans more than 500 data centers, with Fairwater as the first AI super factory designed ground-up for AI, featuring a two-story architecture, closed-loop cooling, and near-zero water consumption.
Project Solara previews agent-first devices: a stationary desk device on MediaTek silicon and a portable wearable badge on Qualcomm silicon, built on an AOSP-based Microsoft device ecosystem platform.
Microsoft IQ unifies Web IQ, Fabric IQ, and Work IQ to ground agents in fresh web data, live operational ontologies, and enterprise procedures.
Microsoft Execution Containers (MXC) bring policy-based isolation and containment for agents to Windows, and Open Claw is announced as running natively on Windows with MXC sandboxing.
The new GitHub Copilot app manages parallel agentic coding sessions with Git work trees, agent merge, canvases, and multi-model access, paired with Rayfin, an agent-first SDK for back end as a service deployed into Microsoft Fabric.
Mustafa Suleyman announces seven new MAI models across image, voice, transcription, and coding, including MAI Thinking 1 and MAI Code 1 Flash, and describes frontier tuning with reinforcement learning environments (RLEs) as each company's hill-climbing machine.
Microsoft announces a partnership with Mayo Clinic to build a frontier model for healthcare, and a Land O'Lakes demo shows frontier tuning reaching over 90% accuracy with roughly 10X efficiency versus baseline models.
Microsoft Discovery reaches GA as an agentic scientific discovery loop, and Majorana 2 qubits deliver lifetimes of 20 seconds to a minute, about a thousand times Majorana 1, in the same qubit form factor.

Questions it answers

00:50Opening framing

What is the core theme of Satya Nadella's Build 2026 keynote?

He describes an AI stack of ubiquitous compute fabric (edge and cloud), a models/context/tools layer, agent runtimes, plus tooling and security/compliance/governance, and says the key takeaway is how developers participate fully in the frontier intelligence ecosystem.

The conference is about the frontier intelligence ecosystem, not one technology
Value comes from what you build and compound on top of the platform

02:47Windows AI and DevBox

What did Microsoft announce for AI at the edge?

Windows leverages the full install base of NPUs, GPUs and CPUs, with new local models Ion Instruct and Ion Plan enabling fully local agentic applications, and the Surface RTX DevBox offering 1 petaflop of AI compute and 128 GB unified memory.

Ion Instruct and Ion Plan run fully local agentic loops
Surface RTX DevBox: 1 petaflop, 20 CPU cores, 128 GB unified memory
Kayla demos vertical taskbar, PowerToys, WSL containers, and 75+ command line utilities

17:58Azure infrastructure

How is Azure scaling for AI workloads?

Azure spans more than 500 data centers, with Fairwater as the first AI super factory designed ground-up for AI, closed-loop cooling with near-zero water use, Cobalt 100 VMs previewing for agent workloads, and Maia 200 delivering 30% more tokens per dollar.

Fairwater: two-story architecture, 100 kW per row, near-zero water consumption
Cobalt 100: 33% lower latency for agent calls, 23% higher throughput
Maia 200 validated with GPT-5 and powering Microsoft 365 Copilot

25:29Jensen Huang conversation

What did Jensen Huang say about the edge and cloud AI opportunity?

Jensen Huang traces the three-year collaboration from DirectX to RTX Spark, describes the PC evolving from personal computer to personal AI, and previews Vera Rubin as a system designed for agentic workloads, with Microsoft deploying the largest number of Grace Blackwells in the world.

RTX Spark fits a couple hundred billion parameter model locally
Grace Blackwell increased token generation rate ~30x over Hopper
Vera Rubin and Vera CPU are designed for agents and low latency

37:36Project Solara devices

What is Project Solara?

It is a turnkey platform for agent-first devices, previewing a stationary desk device on MediaTek silicon and a portable wearable badge on Qualcomm silicon, built on an AOSP-based Microsoft device ecosystem platform with just-in-time UI and bring-your-own-agent extensibility.

Devices extend agents across a constellation of form factors
Healthcare, retail, industrial, hospitality and other verticals are targets
Partners like Best Buy and Target are exploring the devices

51:38Microsoft IQ and data

How does Microsoft IQ ground agents in enterprise context?

Microsoft IQ combines Web IQ (fresh, fast web grounding), Fabric IQ (live operational ontologies), and Work IQ (procedures in SharePoint), demonstrated by a power utility agent that produced an incident brief and sent a Teams notification.

Web IQ is model agnostic and MCP native
Verizon DB shows 3X throughput vs PostgreSQL; Fabric data warehouse shows 7X GPU-accelerated gains
Answers stay current with no reuploads or stale versions

1:04:08MXC and Open Claw

How does Microsoft secure agents running on Windows?

MXC is a new policy layer for isolation and containment of agents in Windows, and Open Claw is announced as running natively on Windows with MXC sandboxing, shown blocking a file deletion attempt in a demo.

Process, session, and Windows 365 isolation options
Open Claw companion app with granular permissions and sandboxing
Peter Steinberger announces the Open Claw Foundation

1:16:10Copilot app and Rayfin

What new developer tools were announced?

The new GitHub Copilot app runs parallel agent sessions with Git work trees, agent merge, canvases and multi-model access, while Rayfin is an agent-first SDK connecting agents to back end as a service deployed into enterprise-managed Fabric tenants.

Sessions run in isolated work trees and can merge via CI
Rubber duck review lets one model check another's work
Rayfin deploys agent-built apps with enterprise back ends

Notes

Overview

Satya Nadella's Build 2026 keynote centers on the frontier intelligence ecosystem: a ubiquitous compute fabric (edge and cloud), a models/context/tools layer, agent runtimes, tooling, and security/governance. The core message: participate fully and compound value on top of the platform.

Edge and Windows

Windows AI leverages the full install base of NPUs, GPUs and CPUs; new local models Ion Instruct (reasoning) and Ion Plan (planning) enable fully local agentic loops.
Surface RTX DevBox: 1 petaflop of AI compute, 20 CPU cores, 128 GB unified memory; NVIDIA's SOC and DGX Station bring "a data center on your desktop."
Windows dev updates: vertical taskbar, PowerToys Grab and Move, Dev Drive, intelligent terminal with Copilot, 70+ Linux utilities including grep, Homebrew, and native containers via WSL.

Cloud and Azure

Azure spans 500+ data centers; Fairwater is the first AI super factory (Georgia and Wisconsin), closed-loop cooled with near-zero water use.
Cobalt 100 VMs preview: 30% better on cloud-native workloads, 33% lower latency for agent calls. Maia 200 delivers 30% more tokens per dollar versus the leading GPU.
Jensen Huang joins to discuss RTX Spark, Grace Blackwell, and Vera Rubin, designed for agentic workloads.

Project Solara

Agent-first devices: a stationary desk device (MediaTek silicon, Hello for Business) and a portable wearable badge (Qualcomm silicon), built on an AOSP-based platform with just-in-time UI and bring-your-own-agent extensibility.

IQ Layer and Agent Platform

Web IQ, Fabric IQ, Work IQ unify grounding for agents; Verizon DB (managed SQL Server) shows 3X throughput vs PostgreSQL; GPU-accelerated Fabric data warehouse shows 7X gains.
MXC (Microsoft Execution Containers) enforce agent containment in the OS; Open Claw runs natively on Windows with MXC sandboxing.
New GitHub Copilot app (parallel sessions, work trees, agent merge, canvases) plus Rayfin SDK for agent-first back ends deployed to Fabric.
Agent 365 governs agents across clouds; MDASH is a multi-model agentic security harness.

Copilot and Autopilots

Copilot becomes a super app (chat, cowork, code); Autopilots are enterprise-grade autonomous long-running agents, starting with Scout on Copilot Frontier.

MAI Models and Frontier Tuning

Mustafa Suleyman announces seven models: MAI Image 2.5 (+flash), MAI Transcribe 1.5 (43 languages), MAI Voice 2 (+flash), MAI Thinking 1 (35B active parameters, 53% on bench pro), MAI Code 1 Flash. Frontier tuning with RLEs lets enterprises build their own hill-climbing machines; Land O'Lakes reached over 90% accuracy at roughly 10X efficiency. A Mayo Clinic partnership builds a frontier model for healthcare.

Discovery and Quantum

Microsoft Discovery hits GA as an agentic scientific loop (plastic recycling demo). Majorana 2 qubits achieve 20-second to one-minute lifetimes, ~1000x Majorana 1, enabling a million qubits on a chip smaller than a credit card.

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