Watch Microsoft Build 2026 from San Francisco

The Big Theme

• Satya Nadella opens Build 2026 in San Francisco with one key takeaway: how developers participate fully in the frontier intelligence ecosystem.
• The AI stack unpacked at the conference: compute fabric (edge + cloud), models/context/tools, agent runtime, tooling, and security/compliance/governance.

Edge: Windows and Local AI

• Goal of "unmetered intelligence to every desk and every home" — local onboard AI already powers Outlook summarize, PowerPoint, Teams super resolution, and Adobe apps via Windows ML.
• Windows ML expands to the full GPU install base; two new inbox models: ion instruct (efficient reasoning SLM) and ion plan (local agentic loop).
• Hardware partners: AMD Ryzen, Intel Panther Lake, Qualcomm Snapdragon X2 Elite and Snapdragon C (sub-$500 PCs), and Nvidia RTX Spark SOC (CPU+GPU+AI, unified memory).
• Surface Ultra (128 GB unified memory) and the Surface RTX Spark Dev Box: one petaflop AI compute, 20 CPU cores, 128 GB unified memory, coming this fall.
• Windows 365 gets a developer distribution; dev updates include distraction-free mode, intelligent terminal with Copilot, 70+ Linux utilities (grep etc.), Starship/ZSH/Homebrew native, and WSL containers.
• Kayla demos: vertical taskbar, winget configure repo, PowerToys Grab and Move, end task, Git-aware File Explorer, running a 120B-parameter model locally (~3.4M tokens used), 90 GB RAM utilized by GPU, three local models simultaneously.

Cloud: Azure Infrastructure

• Driving equation: tokens per dollar per watt. Azure: 500+ data centers, 80 regions; more capacity added in last 18 months than Azure's first decade.
• Three dominant AI workloads: training, inference, agent runtime. Fairwater (Georgia, Wisconsin): two-story AI super factory, closed-loop cooling with near-zero water use.
• First cloud to validate Nvidia Vera Rubin; Maya 200 live in Iowa and Arizona, 30% better tokens per dollar, powering Microsoft 365 Copilot. Cobalt 200 Arm VMs preview: 33% lower agent latency, 23% higher throughput.

Jensen Huang Conversation

• RTX Spark: petaflop NVFP4, 128 GB memory fits ~200B-parameter model; PC becomes "personal AI" that can be texted to run tasks autonomously.
• Vera Rubin designed for agents; Vera CPU designed for low latency; Grace Blackwell deployments at Microsoft are the largest in the world; token generation cost cut ~30x vs Hopper.
• GitHub commits up 3x in recent months; all Azure data tools (SQL, Spark, vector, graph) being GPU-accelerated for impatient agents.

Project Solara

• New chip-to-cloud platform for agent-first devices — "the next computer is not one device."
• Two preview categories: stationary desk device (MediaTek, Hello for Business sign-in, ambient Microsoft 365 Copilot with Work IQ) and portable digital badge (Qualcomm silicon, fingerprint unlock, camera capture, agent-driven workflows).
• Verticals: healthcare (nurse documentation, patient vitals), retail, industrial, hospitality, financial services. Explorers: AccuWeather, Best Buy, CVS Health, Levi's, Target.

Microsoft IQ and Data

• Horizon DB: fully managed PostgreSQL on Azure — zone redundant, 128 TB per cluster, 15 read replicas, ~3x throughput vs self-managed.
• GPU-accelerated Fabric data warehouse: 7x performance gain.
• Web IQ: model-agnostic, MCP-native web grounding on infrastructure serving 1B+ users; best in quality, speed, cost.
• Demo: power utility agent combining Web IQ (current prices), Fabric IQ (grid ontology from Power BI semantic models + live telemetry), Work IQ (SharePoint response procedures), running via Foundry routines with a Teams incident brief.

Agent Runtime, Security, Tools

• MXC (Microsoft Execution Containers): OS-native policy isolation for agents — process, session, VM/WSL, or Windows 365 for full isolation.
• Open Claw on Windows: companion app (WinUI 3), sandboxed tool calls, granular folder permissions; demo of read-only sandbox blocking file deletion; Open Claw Foundation non-profit announced; harness itself now a plugin (Copilot, Codex).
• GitHub Copilot app: home base with parallel sessions via Git worktrees, agent merge babysitting PRs, model picker (OpenAI, Anthropic, Google), rubber duck review across models, agent-built canvases for custom UI.
• Rayfin: agent-first SDK connecting agents to a backend-as-a-service on Fabric; partnership with Replit.
• Agent 365: agent control plane — Entra identities, Defender real-time defense, Purview data protection; works across AWS/GCP and any framework; GA of SDK; Foundry hosted agents with micro VMs, persistent file systems, generated rubrics, and an agent optimizer (tunes model, instructions, tools, skills).
• M-Dash: multi-model agentic security system, 100+ agents, topped Cyber Gym benchmark; found a real Wasmtime bug spanning three code areas; coming to CLI and Defender portal.

Copilot and Autopilots

• Copilot evolves into a super app: chat, Co-work, and code by summer.
• Autopilots: enterprise-grade claws — autonomous long-running agents with names, personalities, memory, running in your tenant. First: Scout, available today for Copilot Frontier users; a full digital team of autopilots coming.

MAI Models (Mustafa Suleyman)

• Philosophy: humanist superintelligence — AI that serves people, not replaces them.
• Seven new models: MAI Image 2.5 (+Flash, #2 on leaderboard, beating Nano Banana 2), MAI Transcribe 1.5 (best in world, 43 languages, 5x faster), MAI Voice 2 (+Flash, 15 languages), MAI Thinking 1 (35B active param MoE, 256K context, 53% SWE-Bench Pro, on par with Opus 4.6, zero distillation, clean data lineage), MAI Code 1 Flash (5B params, 51% SWE-Bench Pro, in VS Code today).
• Available on Foundry, OpenRouter, Fireworks, Baseten; co-designed with Maia 200 silicon (further 1.4x perf/watt).
• Mayo Clinic partnership: jointly develop a frontier health model on Mayo's platform data (100M people, four continents).

Frontier Tuning

• Every company should build its own "hill-climbing machine": private evals, RL environments (RLEs), enterprise knowledge.
• Demo with Land O'Lakes: RL gym from M365 signals, skills with rubrics, virtualized tools, hill-climbed to >90% accuracy, ~10x more efficient than baseline.
• McKinsey tuning: highest win rate, outperforming GPT-5.5 at 10x cost efficiency. "Your RLEs and models become your moat."

Discovery and Quantum

• Microsoft Discovery GA: agentic scientific loop — models + HPC + knowledge graphs + automated labs. Demo: PET plastic recycling proteins, 80 candidate proteins, DNA sequences sent to Cambridge Consultants' automated lab.
• Majorana 2: topological qubits with 20-second to one-minute lifetimes (1000x Majorana 1), 1-microsecond operations, million qubits possible on a chip smaller than a credit card.

Closing

• Two possible stories about this moment: tech concentrates power, or tech unlocks opportunity for developers, scientists, enterprises and communities. "Our job is to make the second story true."