For the last three years, Microsoft’s AI strategy has been inseparable from OpenAI. That dependency was not subtle. It was baked into a relationship cemented by Microsoft’s cumulative investment exceeding $13 billion, giving Microsoft early access to advanced AI models and helping drive Copilot deeper into the enterprise. But Mustafa Suleyman, CEO of Microsoft AI, now says the story changed roughly six months ago when Microsoft’s contract with OpenAI was renegotiated, giving his division formal authority to pursue what he openly calls “superintelligence.”

In an exclusive sit-down interview with VentureBeat at Microsoft Build 2026, Suleyman said, “We were only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence.” And then he added the part that makes this more than a wording tweak: “So this is very early days.” The “early days” framing is important because it tells you Microsoft is not claiming it has completed a full-stack replacement for partners. It is signaling that the constraints are no longer there, so building can scale internally. Translation for executives: Microsoft wants optionality at the frontier level, not just an ecosystem share of it.

The same Build day also delivered tangible proof. Microsoft announced a family of seven new AI models developed entirely in-house by its AI Superintelligence Team, branded under the “MAI” family name. The lineup spans reasoning, code generation, image creation, transcription, and voice synthesis. The flagship is MAI-Thinking-1, a 35-billion-active-parameter reasoning model. Microsoft says it matches leading models in its weight class on key software engineering benchmarks and shows advanced mathematical reasoning. Suleyman emphasized one differentiator repeatedly: the model was trained from scratch on clean, commercially licensed data, without distillation from third-party frontier models, writing in a blog post that “We train our reasoning models from scratch,” and that “We don’t distill from other labs and we don’t rely on unlicensed or opaque data.”

Other MAI releases broaden the multimodal enterprise portfolio: MAI-Code-1-Flash, a lightweight coding model built specifically for GitHub Copilot and VS Code; MAI-Image-2.5, supporting both text-to-image and image editing; MAI-Transcribe-1.5, which Microsoft claims is the most accurate transcription model available, operating across 43 languages; and MAI-Voice-2, a multilingual speech-generation system. All of these ship through Microsoft Foundry, Microsoft’s model-hosting and deployment infrastructure. And for the first time, developers can tune model weights themselves through third-party platforms including OpenRouter, Fireworks, and Baseten.

But Suleyman also made clear these seven models are a proof of concept, not the finish line. The real project is the lab and what it can build over time. He said Microsoft’s job is to ensure that by looking out to 2030 and beyond, it has the capacity not just to buy models from third parties, but to build the absolute frontier, the best models in the world. “That’s a long transition,” he said. This matters because it reframes what to watch. The near-term headlines are the seven MAI models. The strategic bet is the internal ability to iterate toward frontier-scale capability without being locked to a single external provider.

So what does “set free” actually mean, contractually? The source lays out the earlier structure from Microsoft’s 2019 investment into OpenAI: OpenAI would build the frontier models, while Microsoft would be the exclusive cloud provider, integrating those models into its products and reselling them through Azure. That arrangement gave Microsoft commercial leverage and speed, but it also created dependency. Microsoft was explicitly barred from pursuing its own AGI research, and the agreement even capped how large a model the company could train, restricting it from building systems beyond a certain computing threshold measured in FLOPS. The source notes that this architecture was renegotiated. As Fortune and Axios reported in November, a revised deal with OpenAI removed those restrictions, clearing the way for Suleyman to launch the MAI Superintelligence Team and pursue what he calls “humanist superintelligence.”

In Suleyman’s telling, this is not a rupture with OpenAI. At Build 2026, he framed the current posture as one of abundance, not scarcity: “There’s no immediate urgent need to fill a gap in three months' time or six months' time,” he said. “We have OpenAI, we have Anthropic, we have thousands of models inside Foundry. So there's already a huge amount of optionality available to us.” That’s the board-level subtext. Microsoft is spreading risk across multiple frontier sources (including OpenAI and Anthropic), while simultaneously building internal capability so the company is not hostage to partner constraints in the next AI wave.

Finally, Microsoft is pushing beyond chat and into agent-like enterprise action. Suleyman said the shift from chatbots to autonomous AI agents has already begun. At Build, Microsoft announced “Frontier Tuning,” which lets enterprise customers customize MAI models using their own proprietary data, workflows, and domain terminology within their secure compliance boundary. The system uses reinforcement learning environments, what Microsoft calls “training gyms for AI,” letting agents learn directly from real workplace tasks without affecting production systems. Microsoft’s shared results are the kind that enterprise buyers care about: an MAI model tuned for Excel reportedly matches GPT 5.4 performance while operating at up to ten times greater efficiency. The source also reports early enterprise adopters saw gains, including a model achieving the highest win rate of any model tested at roughly one-tenth the cost under one organization’s exacting standards.

Suleyman ties it to a progression from intelligence to action. “We’ve basically moved beyond just conversation,” he told VentureBeat. “Now we’re moving to action.” He introduced a framework shifting from IQ (factual intelligence) to EQ (emotional intelligence, or following tone and style instructions) to AQ, the “Actions Quotient.” Future AI agents, in his telling, won’t just answer questions. They will log into enterprise software, navigate complex multi-application workflows, and execute tasks across tools like Excel, Word, Teams, Jira, and Adobe InDesign. For leaders running enterprise AI strategies, this is the strategic stake: Microsoft is not only building models. It is positioning its stack so customization and action can happen faster, inside compliance, with less dependence on whoever owns the latest frontier checkpoint.