Microsoft and OpenAI's honeymoon period was once the most important alliance in the AI industry.
One party held the models, while the other held cloud services, office software, developer tools, and enterprise clients. They complemented each other, almost securing a first-class ticket for Microsoft in the AI era. But no matter how tight the alliance, Microsoft could not forever entrust its most critical AI prospects to another entity.
Especially as the relationship between the two sides begins to decouple.
The just-concluded Build 2026 conference thus turned into a special event. Microsoft needed a decisive AI victory more than ever to prove to the outside world whether it is the protagonist of the AI era, or still just OpenAI's cloud service provider.
From its MAI models, Azure AI Foundry, quantum computing, to local agent capabilities, plus appearances by Jensen Huang and the creator of OpenClaw, Microsoft showcased a complete ecosystem covering development, models, data, computing power, and governance. The goal was clear: to shift AI from the model premium dominated by OpenAI to a platform business dominated by Microsoft.
Microsoft's In-House Models Launched: MAI Fills the Most Critical Link in the AI Supply Chain
Compared to last year, Microsoft placed models in a more prominent position this time. CEO Satya Nadella stated that Microsoft Foundry now hosts over 11,000 models, covering OpenAI, Anthropic, and Microsoft's own MAI models.
Microsoft's assessment is that businesses and developers will not rely on a single model to complete all tasks. Different tasks will correspond to different models and will be constrained by latency, cost, and capability boundaries. Therefore, model catalogs, model selection, operating environments, and enterprise governance will together form new platform competition points.
Today, Microsoft officially launched seven new models in its in-house model family at once, covering areas such as reasoning, coding, imaging, voice, and transcription.
MAI Thinking 1 is the reasoning model among them. It adopts a sparse Mixture-of-Experts architecture, with 35 billion active parameters and a total parameter scale of about 1 trillion, supporting a 256K token context window, enough to hold roughly 600 pages of documents.
Mustafa Suleyman, head of Microsoft AI, emphasized that this model was not distilled from third-party models. Its training data comes from clean and compliant licensed data, and AI-generated content was excluded during pre-training. It is currently in private preview on Microsoft Foundry and will later enter public testing on the MAI Playground.
The code model MAI Code 1 Flash is designed for everyday development workflows. It is trained end-to-end by Microsoft using clean and compliant licensed data and is being rolled out to individual GitHub Copilot users in Visual Studio Code, accessible via the model selector and the default auto-selector.
Microsoft stated that this model is trained and adapted for the GitHub Copilot harness, supporting agentic coding and adaptive thinking. It keeps responses concise for simple requests while investing more reasoning budget into complex tasks.
Microsoft directly compared MAI Code 1 Flash with Claude Haiku 4.5.
MAI Code 1 Flash achieved 51.2% on SWE-Bench Pro, higher than Claude Haiku 4.5's 35.2%. It leads by 28.9 points on IF Bench for precise instruction following and by 14.5 points on Advanced IF. It will support common coding scenarios in Microsoft GitHub Copilot, especially code modification, multi-turn instructions, and agent tasks in real development environments.
Image and voice models were also incorporated into the MAI system.
Versions MAI Image 2.5 and Flash support text-to-image generation and image editing, already available in PowerPoint and expanding to OneDrive and Foundry.
MAI Transcribe 1.5 supports 43 languages. Microsoft claims its speed is five times faster than competitors, and it is being integrated into GitHub, Teams, Copilot, and Dynamics 365 Contact Center.
MAI Voice 2 supports 15 languages and can adapt voices using short samples, with built-in anti-abuse protection. A low-cost version, MAI Voice 2 Flash, is also in the works.
Microsoft also linked the MAI models to its own chips. MAI Thinking 1 has been optimized for Maia 200, and end-to-end operation of MAI models can achieve a 1.4x improvement in performance per watt.
Enterprise customization is also a significant direction for the MAI models. In the future, all enterprises will not only invoke models but also train their own processes into models.
To this end, Microsoft also released Microsoft Frontier Tuning, with reinforcement learning environments at its core. Businesses can turn real work trajectories, task steps, decisions, tool calls, and evaluation criteria into training environments, allowing models to learn the internal working methods of the organization.
PC Becomes Agent Workstation, Your Desktop is a Data Center
Beyond models, Microsoft also shifted its focus to local computing power.
The Surface RTX Spark Dev Box was the most noteworthy product in this segment. Nadella called it the 'dream machine' for developers. This device offers 1 petaflop of AI computing power, 20 CPU cores, and 128GB of unified memory, planned for launch this fall.
The Surface RTX Spark Dev Box is based on the Nvidia RTX Spark platform. As APPSO reported a few days ago, RTX Spark is the next-generation SoC for PCs, integrating CPU, GPU, and AI capabilities into a single chip and supporting a unified memory architecture and integrated DRTM.
Nvidia CEO Jensen Huang said during a video link that the PC is moving from a personal computer to personal AI. He gave an example: when users are out, they can message their PC and let the local agent invoke tools, modify code, advance designs, and then iterate further with the user.
The PC is no longer just a tool for human operation; it is becoming an AI assistant capable of continuously running tasks.
Furthermore, Microsoft pre-installed a developer-optimized Windows 11 Pro on the Surface RTX Spark Dev Box, including tools like VS Code, WSL, PowerShell 7, GitHub Copilot, and Coreutils for Windows.
In a live demo, this device showed no news feeds, widget pop-ups, or notifications by default, using a dark mode. The Windows Insider build also featured a vertical taskbar. Not only are the development tools further systematized, but the command line and container experience is also closer to Linux.
For hardware, it features a unibody design made from 3D-printed anodized aluminum with 1,000 ventilation holes. It has a 100W TDP and ports including USB-C, USB-A, HDMI, Ethernet, and a headphone jack.
Windows is poised to make significant strides in the AI era. Local AI aims to make the PC part of the agent workflow: developers can debug locally, run models, invoke tools, check logs, start containers, run sub-agents, and then offload larger-scale tasks to the cloud.
Agents Need New Gateways, Microsoft Explores Next-Gen AI Terminals
While the Surface RTX Spark Dev Box targets developers, Project Solara seems more like Microsoft's early exploration of what agent-capable devices might look like. The next computer won't be just one device, but a set of devices working together.
Microsoft showcased two reference device types.
The first is a workstation terminal fixed on a desk, based on a MediaTek chip.
When a user approaches, the system securely identifies them and lets them enter their personal agent work environment, accessing Microsoft 365 Copilot based on Work IQ.
It can display important matters for the day and supports delegating tasks to an Agent via tap or voice. It can also serve as a Windows PC companion or access Cloud PC via Windows 365. It functions more like an Agent control terminal on the corporate desk, responsible for identity recognition, task reminders, voice interaction, Copilot invocation, and Cloud PC access.
The second is a wearable digital badge, using a Qualcomm wearable chip, targeting mobile work scenarios.
In the demo, after unlocking with a fingerprint, the user asks Copilot to gather on-site materials for a social media post. The badge captures footage, and the agent selects shots, cleans up the image, and sends it to the individual and their team for review. The presentation also showed a medical scenario: a nurse can use it for hands-free voice recording, speaker differentiation, vital sign verification, medication scanning, and care process validation.
These two device types are merely reference designs.
Phones and PCs remain important, but some work scenarios require hardware that is closer to people, spaces, and sensors. Facing the future agent era, enterprises can swap agents and adjust appearance, screens, sensors, and input methods to adapt to different vertical industries on the same hardware and software foundation.
OpenClaw Creator on Stage, Microsoft Adds Enterprise Guardrails to Personal Agents
While the Surface RTX Spark Dev Box addressed local computing and Project Solara explored new device forms, OpenClaw on Windows shifted the focus to how personal agents can securely enter the enterprise.
Microsoft demonstrated the Windows suite for OpenClaw, which can help users set up their own instance or connect to one already hosted on Windows and WSL.
Within the application, you can view gateways, other machines participating in OpenClaw, sessions, and usage statistics, and quickly access chat, canvas, and the main console.
The security demo revolved around file permissions.
The OpenClaw Windows Companion app allows users to control which folders the agent can access and whether those folders are read-only, writable, or hidden. It can also configure clipboard access, network permissions, and other fine-grained options.
On stage, Microsoft instructed OpenClaw to delete all files on the desktop, temporarily turning off OpenClaw's own security layer and leaving only the MXC's system-level restrictions. Because the desktop folder was set to read-only, OpenClaw attempted multiple times to delete and check the directory but ultimately could not delete the files, leaving the 94 JPGs on the desktop intact.
Peter Steinberger, creator of OpenClaw, also revealed that over the past few months, OpenClaw has collaborated with teams at Microsoft, GitHub, OpenAI, Nvidia, and others, adding observability, automatic permission modes, and a redesigned access control system. Now, permissions are no longer just all allow or all deny; users can specify which folders are read-only, which are writable, and which are hidden from the agent.
He also announced that OpenClaw can run within a company, and the harness itself is now plug-in enabled. Businesses can integrate their trusted Copilot, Codex, or other systems, bringing their existing rules into OpenClaw and gaining continuous memory, heartbeat, and the ability to use OpenClaw within Slack or Teams.
Into the Second Half of AI, Microsoft Eyes the Enterprise Platform Gateway
In addition to the hardware and Windows updates mentioned above, Microsoft announced more products.
On the development tools side, Microsoft launched the new GitHub Copilot app. It functions more like an agent coding session manager, allowing developers to start multiple issue sessions simultaneously and isolate them using Git worktrees, enabling multiple agents to work in parallel.
Agent Merge is responsible for tracking CI checks, code reviews, and merge conflicts for pull requests. Microsoft also released Raven, an agent-first SDK designed to connect to backend as a service, handling issues like identity, storage, and database schema.
Contextual capabilities are handled by Web IQ.
For agents to enter enterprise workflows, they need to connect new web information, enterprise business objects, real-time operational statuses, personnel relationships, and organizational processes. Web IQ is responsible for external web information, supporting web pages, news, images, and videos. It is model-agnostic and MCP native, and can be plugged into any agent runtime, allowing agent responses to be based on updated and verifiable content.
Copilot is also being upgraded into a more complex work entry point.
Nadella stated that this summer, chat, cowork, and code will be brought together into a single Copilot, and Autopilots will be released. The first Autopilot, named Scout, will be available to Copilot Frontier users and can work within Teams group chats and Outlook threads.
For enterprise governance, Microsoft released Agent 365. It provides agents with identity, permissions, access controls, and compliance management, integrating with Entra, Defender, and Purview. Agent 365 can manage agents hosted on Azure, AWS, GCP, or other environments and supports agents built with different frameworks.
On the research front, there is Microsoft Discovery. Nadella defined it as an agent platform for scientific discovery, aiming to connect paper research, candidate solution generation, simulation computation, experimental design, and automated labs into a continuous flow.
At the end of the conference, Microsoft officially launched its next-generation quantum chip, Majorana 2. Its average qubit lifetime can reach 20 seconds, with a peak close to 1 minute, roughly 1,000 times higher than Majorana 1. The operation time is 1 microsecond, with dimensions still in the 0.01 millimeter range, and it uses all-digital control.
With that, this sprawling and ambitious conference puzzle was finally complete. The first phase of AI saw model companies dominate the industry narrative; the second phase may see platform companies dominate industrial implementation.
Whoever selects the models, allocates the tasks, manages the agents, and defines the permissions and audits will be closer to the core gateway of enterprise AI. As models gradually become a standard capability, the systems that host their operation will truly determine the attribution of value.
If that first-class ticket for the AI era was largely seized with the help of ally OpenAI back then, now Microsoft has moved into the cockpit and intends to take over the course of this aircraft itself.