Microsoft’s Build 2026 conference opened with a clear message: the company is no longer content to rely on a single AI partner. CEO Satya Nadella and Microsoft AI chief Mustafa Suleyman took the stage in early June to unveil MAI, a new family of internally developed foundation models that will power the next generation of Copilot experiences across Windows, Microsoft 365, and Azure. The move signals a fundamental reorientation of Microsoft’s AI strategy, placing its own models at the center while maintaining OpenAI’s technology as just one of many available tools.

For years, Microsoft’s AI push was synonymous with its multibillion-dollar bet on OpenAI and the GPT model family. With MAI—short for Microsoft AI models—the company is plotting a course toward greater independence, control, and differentiation. The first MAI models, already in preview, include a flagship large language model (MAI-1) and a series of smaller, specialized variants optimized for on-device and enterprise workloads.

Mustafa Suleyman takes the lead

Mustafa Suleyman, who joined Microsoft in 2024 to head the newly formed Microsoft AI division, has quietly built an organization capable of competing with the very partner that helped elevate Copilot into a household name. His keynote at Build 2026 was part manifesto, part technical deep dive. “We are building AI that understands the world in a Microsoft way—deeply integrated, responsible, and designed for productivity,” Suleyman said, before demonstrating MAI-1’s multimodal capabilities.

The demonstration showed MAI-1 ingesting a complex Excel spreadsheet, a PowerPoint deck, and a real-time video feed from a factory floor, then generating a unified risk report in seconds. Unlike previous Copilot demos that relied on stitching together separate models, MAI handled the entire workflow natively. Attendees noted the speed and fluency, but also the subtle branding shift: the new Copilot experience, powered by MAI, will be simply “Copilot,” while the existing GPT-powered version will be rebranded as “Copilot Classic” for enterprise customers who want to stick with OpenAI’s stack.

What MAI models bring to the table

Microsoft is positioning MAI as a family of models rather than a single monolithic offering. At the top end, MAI-1 competes squarely with GPT-5 and Google’s Gemini Ultra on reasoning, coding, and multimodal tasks. Microsoft claims it achieves parity on several industry benchmarks while using fewer compute resources, thanks to novel sparsity techniques and a custom training stack refined on Azure’s AI supercomputers.

More interesting for Windows users are the on-device variants: MAI-Nano and MAI-Edge. These compact models run directly on NPUs—neural processing units found in Copilot+ PCs—and enable features like real-time language translation, local semantic search, and background noise removal in Teams without an internet connection. At Build, Microsoft showed a Surface Pro running a fully local Copilot assistant that could summarize a meeting and draft follow-up emails even in airplane mode. This addresses one of the longest-standing complaints about Copilot: its reliance on the cloud made it sluggish and unreliable in patchy network conditions.

For enterprise IT, MAI brings a new level of control. Organizations can fine-tune MAI models on their own data within a secured Azure environment, with strict data residency guarantees. Microsoft calls this “Copilot in a box,” and it’s clearly aimed at regulated industries like healthcare and banking that have been hesitant to send sensitive data to third-party API endpoints. The MAI family also includes specialized models for code generation (MAI-Code), scientific reasoning (MAI-Science), and a remarkably efficient small model for text summarization that sips power, extending battery life on laptops.

The end of an exclusive partnership?

Microsoft’s relationship with OpenAI has never been simple. The two companies have a complex investment structure, and despite Microsoft’s $13 billion infusion, OpenAI retains the right to pursue commercial deals independently. Recent reports of friction—over product roadmaps, safety concerns, and the release cadence of new models—paint a picture of a partnership under strain. By developing MAI, Microsoft is hedging its bets.

Analysts at the conference noted that Nadella’s mention of OpenAI was notably brief. “We continue to work closely with OpenAI and offer their models as a choice,” he said during the keynote Q&A. It was a deliberate framing: OpenAI as one option among many, not the default. Then Suleyman demonstrated Copilot’s model switching feature, where users can pick from MAI, GPT, or even open-source models like Llama and Mistral, all from the same interface.

This “model-agnostic” approach is a clever move. It lets Microsoft serve enterprise customers who have already standardized on OpenAI while gradually migrating others to MAI. It also insulates Microsoft from the reputational risks that come with any single AI provider’s missteps. For Windows users, the impact is immediate: Copilot will soon default to MAI on new devices, and Microsoft is phasing out the requirement for a Microsoft 365 subscription to access the most advanced local AI features. Basic on-device Copilot powered by MAI-Nano will be free and built into Windows 11 and the upcoming Windows 12.

Windows AI gets a brain upgrade

Windows integration was the star of the show for PC enthusiasts. Panos Panay, Windows chief, showed off a new AI “System Engine” that continuously profiles hardware, apps, and user behavior to optimize MAI model placement. It can decide on the fly whether to run an inference on the NPU, GPU, or CPU, or to offload to the cloud—balancing speed, privacy, and battery life. This is a significant departure from the current setup, where most AI tasks are cloud-dependent.

A live demo pitted a Copilot+ PC running MAI-Nano against a MacBook Air with Apple Intelligence. Both were asked to generate a 10-slide presentation from a raw text outline. The Surface Pro completed the task entirely on-device in 14 seconds, while the MacBook required an internet connection and took 22 seconds, during which it stalled twice. The audience erupted in applause. It was a carefully orchestrated comparison, but the point was clear: Microsoft believes its local AI can beat Apple’s.

Other Windows AI upgrades revealed at Build include a MAI-powered “Recall” successor that actually respects privacy by processing all data locally and encrypting semantic indexes, and a universal “Copilot Vision” feature that can see and analyze anything on the screen—from PDFs to video streams—without needing cloud uploads. Developers can access MAI models through a new Windows AI API, and Microsoft promised that all MAI-Nano features would be available to third-party apps through the Microsoft Store, with strict quality and security reviews.

Enterprise IT takes notice

IT administrators at Build 2026 were buzzing about the control and compliance features. A new Copilot admin console, built on top of Microsoft Intune, lets IT set policies that mandate on-device processing for certain data classifications. “You can literally tick a box that says ‘never send HR data to the cloud,’ and it’s enforced at the model level,” said one IT manager from a large financial firm during a breakout session.

Microsoft also announced a dedicated MAI model for cybersecurity, trained on trillions of signals from Microsoft Defender and Azure Sentinel. This model can detect and respond to threats in milliseconds, without ever sending logs outside the organization’s perimeter. For the 85% of Fortune 500 firms that use Azure, this is a compelling reason to double down on the Microsoft ecosystem rather than cobbling together AI components from various vendors.

Pricing was another hot topic. Microsoft said MAI inference will be included in existing Microsoft 365 E3 and E5 plans at no extra cost, at least for the on-device models. Cloud-based MAI usage will be metered through Azure, but the company teased a novel “MAI credits” system that gives each enterprise a monthly pool of tokens for custom fine-tuning and heavy workloads. This bundled approach contrasts sharply with OpenAI’s per-seat pricing and could further lock in Microsoft’s enterprise base.

The developer story: MAI Studio and Azure AI

For the developer community, Build 2026 introduced MAI Studio, a web-based playground akin to OpenAI’s playground but with deeper Azure integration. Developers can compare MAI, GPT, and open models side by side, test prompts, and deploy optimized endpoints with a few clicks. Microsoft also open-sourced the MAI-Nano architecture, allowing researchers and hardware makers to adapt it for IoT and edge devices.

GitHub Copilot, the company’s flagship AI coding tool, will soon support MAI-Code as the default backend for enterprise customers, with a private preview showing a 23% reduction in cost per suggestion compared to GPT-4. GitHub CEO Thomas Dohmke said during a session that the switch would be seamless for users, as the model will be trained on the same data and follow the same coding styles that developers expect.

A key technical detail shared in breakout sessions: MAI-1 was trained using a novel “scaffolded” approach, where smaller expert models were trained first on specific domains, then merged into a single large model. This allowed Microsoft to bypass some of the stability issues that have plagued rivals’ large-scale training runs. The result is a model that exhibits fewer hallucinations and better adherence to safety guidelines—claims Microsoft supported with internal benchmark data, though independent verification is still pending.

What this means for the AI arms race

The arrival of MAI reshuffles the competitive landscape. Microsoft is now a direct competitor to OpenAI, Anthropic, and Google in the foundation model space, while also being the platform on which many of those models run via Azure. This dual role will invite scrutiny from regulators, who already have their eyes on Microsoft’s AI investments. At a press briefing, Microsoft’s chief legal officer said the company is “fully committed to fair competition and will continue to offer interoperability with all major model providers.” Still, antitrust lawyers at the conference noted the potential for conflicts of interest, especially if Azure prioritizes MAI performance over third-party models.

For consumers, the shift to MAI will likely be invisible at first. The Copilot icon won’t change, and prompts will feel snappier. But over the next 12 months, as MAI models improve and become the default across Windows, Office, and beyond, users may notice a distinctly Microsoft-flavored intelligence—one that’s deeply woven into the OS, understands Windows-specific tasks better, and costs nothing to use on-device. It’s a bold bet that the AI future belongs to those who control the entire stack, from silicon to software.

Microsoft’s Build 2026 was, above all, a declaration of confidence. The company that once looked like an AI follower now wants to lead. MAI models are the engine of that ambition. Whether they can truly outshine the best from OpenAI remains to be proven, but the message is unmistakable: Copilot is no longer just a wrapper around someone else’s models. It’s becoming a platform in its own right, and Microsoft is holding the keys.