Microsoft Reduces OpenAI Dependency With In-House Frontier Models
The first big event happened in late October 2025, when OpenAI formally completed its conversion into a Public Benefit Corporation, or PBC. Since then OpenAI Group PBC has been functioning under a nonprofit parent called the OpenAI Foundation. Microsoft formalized its position at the time as a roughly 27% stakeholder, valued at approximately $135 billion, up from a total investment of around $13.8 billion dating back to 2019.
During this time, OpenAI also committed to purchasing $250 billion worth of Azure services, while Microsoft locked in IP licensing rights through 2032, including access to post-AGI models, with that AGI milestone being determined by an independent expert panel rather than OpenAI's board. Under the old deal, Microsoft could have lost access to OpenAI's technology the moment OpenAI declared it had achieved Artificial General Intelligence, or AGI. Microsoft was able to defuse this risk under the new deal, however.
It is also important to note, Microsoft remained OpenAI's exclusive partner for frontier models and Azure API access.
That exclusivity, however, died in the second deal, announced April 27, 2026. OpenAI can now sell its products and APIs on any cloud it chooses. It has since signed a $38 billion multi-year compute agreement with AWS, a reported $300 billion arrangement with Oracle through the Stargate initiative, and deals with Google Cloud and CoreWeave. The IP license shifted from exclusive to non-exclusive through 2032, and the 20% revenue share OpenAI pays Microsoft was capped at a total of $38 billion, saving OpenAI an estimated $97 billion compared to what an uncapped arrangement would have cost at its current growth rate. Microsoft, in turn, stopped paying its own revenue share back to OpenAI entirely.
The new deal also freed Microsoft from the exclusivity constraint. It didn't take long for the software giant to flaunt its new found freedom, and share its new plans. At Build 2026, the company unveiled seven in-house AI models under the MAI brand. The headliner is MAI-Thinking-1, Microsoft's first reasoning model, trained from scratch on commercially licensed data with no refining from OpenAI or any other third-party model. That data origin matters quite a bit for enterprise customers with strict compliance requirements. The company explained the model uses a sparse Mixture-of-Experts architecture with about 35 billion active parameters and a 256,000-token context window. CEO Satya Nadella framed the moment as a shift from "consuming a frontier model to fully participating at the frontier."
The rest of the MAI family fills out the stack with models covering image generation, transcription, voice, and coding. MAI-Code-1-Flash is already rolling out to GitHub Copilot users. The models are also available through Azure AI Foundry and will appear on third-party platforms including Fireworks AI, Baseten, and Open Router.
By loosening the partnership terms that once kept OpenAI exclusive to Azure, Microsoft also bought itself the freedom to become a direct competitor. This also means OpenAI effectively helped train its biggest rival. For enterprise customers, that competition is probably good news, introducing pricing pressure and giving IT teams a Microsoft-native AI option that runs entirely within Azure without an OpenAI API call anywhere in the chain. Whether MAI-Thinking-1 can close the gap against GPT-5 and Anthropic's latest models in real-world workloads remains to be seen, but Microsoft has made it clear that it is done leasing the foundation of its own AI strategy and is ready to produce its own "humanist superintelligence."
