Microsoft Build 2026: After Cutting Ties with OpenAI, Unveils 20+ New AI Models and Hardware Updates

Packing Compute into a Small Box

During a remote conversation with Nvidia CEO Jensen Huang, Microsoft CEO Satya Nadella announced the Surface RTX Spark Dev Box, a desktop workstation powered by Nvidia RTX Spark, delivering 1 PFLOPS of AI compute, 20 CPU cores and 128 GB unified memory. It can run 120‑billion‑parameter models locally, supports fine‑tuning, and ships in an aluminum chassis this fall. The device ships with a developer‑focused Windows 11 configuration that pre‑installs VS Code, an integrated GitHub Copilot terminal, WSL, PowerShell 7, and disables news feeds, pop‑ups and notifications; the setup can be deployed via the Windows Developer Config project on GitHub. Security features include the Secure PC architecture, BitLocker, Microsoft Defender, and enterprise‑grade management via Entra ID and Intune.

Project Solara, presented by Microsoft VP Steven Bathiche, showcases two reference Agent‑first devices: a fixed desktop terminal built on a MediaTek platform with Windows Hello enterprise authentication, and a lightweight wearable badge built on a Qualcomm Snapdragon platform. Both devices surface contextual information, automate tasks, and integrate seamlessly with Windows PCs and Windows 365 cloud PCs. Enterprises such as Best Buy, CVS Health, Levi’s and Target are exploring these reference designs.

Seven Models Debut

Microsoft AI research (MAI) released seven new models after ending its partnership with OpenAI. MAI‑Thinking‑1, the first inference model, uses a Mixture‑of‑Experts architecture with 350 billion active parameters (≈1 trillion total) and a 256 K context window. According to AI CEO Mustafa Suleyman, it leads industry benchmarks in human‑rated preference and multiple reasoning tests, scoring 53 % on SWE‑bench Pro, comparable to Claude Opus 4.6. The model was trained from scratch without benchmark‑specific tuning, distillation, or proprietary data, and carries a commercial licence suitable for enterprise deployment.

MAI‑Image‑2.5 and its lightweight variant MAI‑Image‑2.5‑Flash achieve top‑ranked image quality and editing capabilities, placing second on the Large Model Arena image‑editing leaderboard, surpassing Google’s Nano Banana 2. Two voice models were also announced: MAI‑Transcribe‑1.5 supports 43 languages with industry‑leading transcription accuracy, and MAI‑Voice 2 (and MAI‑Voice‑2‑Flash for real‑time Agent scenarios) cover 15 languages with enhanced prosody, emotion and fine‑grained control.

The programming model MAI‑Code‑1‑Flash, with 5 billion parameters, scores 51 % on SWE‑bench Pro and is heavily optimized for VS Code and GitHub Copilot CLI, delivering near‑large‑model code generation at lower cost. All MAI models are being integrated into PowerPoint, rolled out to OneDrive, and made available through Azure Foundry, OpenRouter, Fireworks AI, Baseten and other third‑party platforms, with weights open for fine‑tuning.

Model‑chip co‑design advances include MAI‑Thinking‑1’s specialization for Microsoft’s in‑house Maia 200 AI accelerator, delivering 1.4× the performance‑per‑watt of Nvidia’s GB200. Deployed in Iowa, Maia 200 offers ~30 % better token‑per‑dollar efficiency than mainstream GPUs and has been validated on GPT‑5.5 workloads.

Microsoft also introduced Frontier Fine‑Tuning, enabling enterprises to train proprietary models on their own data, workflows and knowledge assets via reinforcement‑learning‑environment (RLE) pipelines, achieving up to 10× cost‑efficiency improvements and matching flagship model performance on select tasks.

Agents Must Follow Rules

The first locally‑operating Agent, Microsoft Scout, built on the open‑source OpenClaw framework, runs on Windows and macOS. Scout possesses a distinct identity, can create, edit and search documents, join Teams chats, process Outlook mail, and launch sub‑Agents for parallel research, code review and complex tasks. Its permission system lets administrators enable or disable specific skill categories and protect sensitive files. Scout serves as the default Agent in the Autopilots platform, with users able to create additional Agents.

To mitigate the security risks of increasingly autonomous Agents, Microsoft released the MXC (Microsoft Execution Container) SDK preview, a cross‑platform, policy‑driven execution layer that abstracts isolation primitives for developers. MXC currently supports process and session isolation—suitable for programming Agents and long‑running automation workflows respectively—and will later add micro‑VM and Linux‑container isolation.

Microsoft collaborates with partners such as Hermes, Manus, Nvidia, OpenAI and OpenClaw to ensure compliance with real‑world requirements. The OpenClaw Windows Companion app, built with WinUI 3, demonstrates policy enforcement: a command to delete a desktop file was blocked because MXC was set to read‑only mode.

Windows 365 for Agents extends isolation to cloud PCs managed by Intune, while the Agent 365 control plane integrates Entra identity, Defender protection and Purview compliance to enforce identity‑bound, fully protected Agent execution. The MDASH multi‑model Agent security system monitors over 100 Agents, detects exploitable vulnerabilities, and surfaces remediation suggestions in the Defender portal.

Cloud to Quantum

Satya Nadella emphasized that the key metric for AI‑centric cloud infrastructure is tokens generated per watt‑per‑dollar. Azure now spans 80 regions with over 500 data centers, adding more capacity in the past 18 months than in its first decade. New data centers are purpose‑built for AI training, inference and Agent runtimes.

The Fairwater AI super‑factory, co‑designed with Nvidia, employs a dual‑layer architecture and three‑dimensional rack layout to maximize GPU density, bandwidth and low‑latency interconnects, while achieving near‑zero water consumption comparable to a typical restaurant’s daily usage.

Microsoft is the first cloud provider to validate Nvidia’s next‑gen Vera Rubin platform and continues its partnership with AMD on MI300 accelerators. The in‑house Maia 200 AI accelerator is also progressing, with a projected 1:1 CPU‑to‑GPU ratio in future Agent systems.

Cobalt 200, an Arm‑based cloud CPU preview, delivers >50 % performance uplift for cloud‑native workloads, 33 % lower Agent latency, 14 % faster execution and 23 % higher throughput compared to Cobalt 100.

On the networking side, Microsoft’s MRC architecture re‑architects Azure traffic scheduling for massive synchronous data parallelism required by AI training, and inter‑regional AI‑dedicated networks connect Fairwater factories into a unified compute pool.

In quantum, Microsoft announced the Majorana 2 processor, whose qubits are 1,000× more reliable than those in Majorana 1, operate on microsecond timescales with an average lifetime of 20 seconds (occasionally >1 minute). The chip uses a new material stack—lead replacing aluminum superconductors and a mixed arsenic/antimony semiconductor—to create a larger topological gap, improving protection against noise. Microsoft projects a scalable quantum computer by 2029, driven by Agentic AI.

Finally, Microsoft Discovery, an enterprise‑grade Agent AI platform for end‑to‑end research workflows, is now available for drug discovery, semiconductor design and other scientific domains, with a free local preview.

Overall, Build 2026 presented a comprehensive stack—from chips to models, from edge devices to cloud services, from security to quantum—aimed at establishing a robust foundation for the Agent era.