Microsoft is developing in-house artificial intelligence models, including the 500 billion parameter MAI-1, to reduce its dependence on OpenAI. The company is working on AI reasoning systems that could potentially be made available to external developers as an API in the future. This strategic move is a key part of Microsoft’s effort to diversify its artificial intelligence offerings and optimise its technology-related expenditures.
The MAI-1 model, with approximately 500 billion parameters, sits between the GPT-3 (175 billion) and GPT-4 (presumed to be 1.76 trillion) in terms of performance. This configuration enables the model to deliver highly accurate responses while consuming significantly less energy, resulting in lower operational costs for Microsoft. The model’s training dataset includes text generated by GPT-4 as well as web content, and its development is being carried out on server clusters equipped with Nvidia graphics cards. "We are a full-stack solutions company, and we want to provide comprehensive system capabilities," said Satya Nadella.
The tech giant is already testing models from xAI, Meta, and DeepSeek as potential OpenAI alternatives for its Copilot service, while also developing the multimodal Phi-4 and Phi-4-mini models. Tensions between Microsoft and OpenAI are increasingly evident, particularly as OpenAI’s new Stargate Project cloud strategy aims to reduce its reliance on Azure. Additionally, OpenAI recently signed a $12 billion deal with CoreWeave, which was previously Microsoft’s largest customer, generating 62% of its revenue ($1.9 billion) in 2024.
Sources:
1.
Nileena Sunil
2.
3.
Maria Deutscher
