Microsoft is developing its own advanced AI reasoning models, known internally as ‘MAI,’ to reduce dependency on OpenAI and enhance its AI capabilities, according to a report by The Information.
Microsoft develops its own AI models to reduce OpenAI reliance
The MAI models are designed to perform complex reasoning and problem-solving tasks, with the goal of matching or surpassing the performance of leading models from OpenAI. Microsoft is also experimenting with integrating these MAI models into its Microsoft 365 Copilot product.
Additionally, the company is considering releasing the MAI models as an application programming interface (API) for external developers later this year, allowing other developers to incorporate Microsoft’s AI capabilities into their applications.
This move comes after Microsoft invested $13.75 billion in OpenAI since 2019, while exploring alternatives to OpenAI’s AI model, particularly for its 365 Copilot product. In the initial launch of Microsoft 365 Copilot in 2023, the product prominently featured OpenAI’s GPT-4 model.
The shift in usage is motivated by several factors: avoiding monopoly or exclusive dependency, reducing costs, and improving speed. Microsoft’s AI division, led by Mustafa Suleyman, has completed training a family of models that reportedly perform nearly as well as top models from OpenAI and Anthropic, based on established benchmarks.
Testing and replacements
Microsoft is currently testing AI models from xAI, Meta, and DeepSeek as potential replacements for OpenAI’s technology within the Microsoft 365 Copilot framework. These developments coincide with Microsoft introducing AI features within the Microsoft 365 suite for Personal and Family subscription plans after previously restricting these capabilities to a separate Copilot Pro subscription at an additional cost of $20 per month.
Earlier this year, Microsoft announced a shift in its partnership with OpenAI, moving away from its exclusive arrangement as the cloud provider for OpenAI’s computing needs. Under the revised agreement, Microsoft retains a ‘right of first refusal’ (ROFR) to provide cloud resources to OpenAI while allowing OpenAI to seek services from other cloud providers, including Oracle.
The timing of this change aligns with the reveal of the Stargate Project, a joint initiative involving OpenAI, Oracle, SoftBank, and Nvidia, expected to inject up to $500 billion to develop infrastructure for next-generation AI models in the United States. Simultaneously, Microsoft plans to invest $80 billion in AI-focused data centers and cloud infrastructure in fiscal year 2025, including $3 billion allocated for India.
Microsoft has expressed a desire to decrease its reliance on OpenAI while maintaining a competitive edge among major technology firms in the AI landscape. Reports indicate that the company is actively working to integrate both internal and third-party AI models into Microsoft 365 Copilot, diversifying the current technology sourced from OpenAI and reducing associated costs.
According to The Information, Mustafa Suleyman’s team is engaged in training reasoning models utilizing chain-of-thought techniques, which generate answers through intermediate reasoning processes to tackle complex problems. The MAI models are reported to be larger than a previous family of Microsoft models referred to as Phi, and the team is already testing the potential of replacing OpenAI’s models in Copilot.
The company’s exploration of MAI models and alternative external AI technologies reflects a strategic shift towards greater independence in AI development and deployment.
Featured image credit: Demure Storyteller/Unsplash
