Microsoft’s AI division breaks free from OpenAI to build superintelligence

Microsoft is no longer tethered to OpenAI’s roadmap. In a strategic pivot announced at Microsoft Build 2026, the company’s AI division—led by CEO Mustafa Suleyman—has formally broken away from its long-standing dependency on its San Francisco-based partner, opening the door to pursue what Suleyman calls "superintelligence." The shift, triggered by a contractual renegotiation roughly six months prior, grants Microsoft’s AI Superintelligence Team full autonomy to design, train, and deploy its own models using proprietary data pipelines and custom silicon.

"We were effectively set free from our arrangement with OpenAI about half a year ago, allowing us to formally pursue superintelligent systems," Suleyman explained during an exclusive interview with VentureBeat. "This is still the very beginning of the journey."

The announcement marks a critical inflection point for Microsoft, which has spent over $13 billion accumulating early access to OpenAI’s cutting-edge models. While the partnership remains intact for commercial products like Copilot, the company is now laying the groundwork for an independent AI ecosystem—one that could eventually transcend its reliance on third-party providers.

Microsoft’s first homegrown AI models debut under the MAI family

Alongside Suleyman’s remarks, Microsoft unveiled its first in-house AI model family, codenamed MAI, signaling a new phase of ambition. The suite consists of seven specialized models spanning reasoning, coding, image generation, transcription, and speech synthesis. Unlike previous generations, these models were developed entirely in-house, from scratch, using commercially licensed data—no distillation from rival systems was involved.

The flagship, MAI-Thinking-1, is a 35-billion-active-parameter reasoning model that Microsoft claims competes with leading peers in its class for software engineering benchmarks. In a follow-up blog post, Suleyman emphasized the model’s independence: "We train our reasoning models from scratch. We don’t distill from other labs, and we don’t rely on unlicensed or opaque data."

The remaining models round out a multimodal portfolio tailored for enterprise use:

• MAI-Code-1-Flash: A lightweight model optimized for GitHub Copilot and Visual Studio Code, designed for real-time code assistance.
• MAI-Image-2.5: Supports both text-to-image generation and image editing, catering to creative and marketing workflows.
• MAI-Transcribe-1.5: Boasts the highest accuracy among transcription models, supporting 43 languages for global deployment.
• MAI-Voice-2: A multilingual speech-generation system enabling natural, context-aware interactions.

All models are accessible through Microsoft Foundry, the company’s model-hosting infrastructure, and for the first time, developers can fine-tune weights via third-party platforms such as OpenRouter, Fireworks, and Baseten.

What the "set free" moment really means for Microsoft’s AI future

The contractual shift that enabled this independence was years in the making. When Microsoft first invested billions into OpenAI in 2019, the deal positioned the tech giant as OpenAI’s exclusive cloud provider, granting it early access to frontier models like GPT-4. While the arrangement accelerated Microsoft’s AI integration into products like Azure and Copilot, it also imposed restrictions: the company was barred from pursuing its own artificial general intelligence (AGI) research and capped in the size of models it could train, measured in floating-point operations per second (FLOPS).

Those constraints were lifted in a renegotiated agreement last year, according to reporting from Fortune and Axios. The revised terms removed model-size limitations and explicitly permitted Microsoft to launch the MAI Superintelligence Team, dedicated to advancing "humanist superintelligence"—a concept Suleyman has described as AI aligned with human values and societal benefit.

The implications are profound. No longer restricted to adopting OpenAI’s models wholesale, Microsoft can now compete at the frontier of AI development. Suleyman framed the transition as a long-term play: "Our mission is to ensure that by 2030 and beyond, we’re not just purchasing models from others—we’re building the absolute best in the world. This is a gradual but essential evolution."

A dual-track strategy: OpenAI and in-house innovation

Despite the newfound independence, Microsoft has made clear it isn’t walking away from OpenAI. The partnership remains central to its commercial AI strategy, particularly for enterprise customers who rely on Copilot and Azure AI services. However, the MAI family represents a parallel track—one focused on innovation that could eventually stand alone.

For developers and businesses, this dual approach offers both stability and choice. Existing Copilot users will continue to benefit from OpenAI’s models, while early adopters can experiment with Microsoft’s first-party alternatives. The company’s decision to open fine-tuning capabilities to third-party platforms further signals an intent to foster an ecosystem around its models.

As Suleyman noted, the journey has just begun. "We’re building the infrastructure and talent to ensure Microsoft is not just a follower, but a leader in the next era of AI." With the MAI family as its first milestone, the company is turning a new page—one where independence and ambition go hand in hand.