All may not be well between Microsoft and OpenAI. A new report suggests that Microsoft is building its own AI model to rival OpenAI.
In a recent podcast, Microsoft chief Satya Nadella said that Microsoft doesn’t need to build an LLM just to “prove a point”. When asked why Microsoft still hasn’t built its own foundational models, he said that Microsoft sees itself as a full-stack company and LLMs are just a part of it.
“We’re a full-stack systems company, and we want to have full-stack systems capability,” he said, adding that the company focuses on integrating models into broader systems and products.
He added that Microsoft has a “long-term stable relationship” with OpenAI and retains important IP rights through their partnership, and the company has built systems, tools, and products around OpenAI’s models rather than just relying on the models themselves. “I do believe the models are becoming commoditised, and in fact, OpenAI is not a model company, it is a product company,” he said.
To some extent, Nadella is right. He is worried that OpenAI, by releasing products, is creating direct competition. That explains why, whenever OpenAI releases new products, Microsoft integrates them into Copilot. For instance, the Copilot app recently announced unlimited access to Voice and Think Deeper, powered by OpenAI’s o1 model.
OpenAI’s revenue for 2024 was approximately $3.7 billion, with significant growth projected for 2025, reaching around $11.6 billion. Meanwhile, Microsoft recently informed shareholders that it is generating over $13 billion in annualised AI revenue.
Mustafa x Microsoft
Nadella asserted that under Microsoft’s AI chief Mustafa Suleyman’s leadership, Microsoft has built the Phi models. Since the tech giant has been trying to be less dependent on OpenAI, it recently launched Phi-4-multimodal and Phi-4-mini, the latest additions to its Phi family of small language models (SLMs).
The Phi-4 multimodal model supports applications such as document analysis and speech recognition. On multimodal audio and visual benchmarks, it surpasses Google Gemini 2 Flash and Gemini 1.5 Pro. Microsoft claims that it is comparable to OpenAI’s GPT-4o.
Suleyman reportedly clashed with OpenAI leadership over access to technical details, prompting the company to explore alternatives and invest in in-house innovation.
Last fall, during a video call, Suleyman pressed OpenAI to share documentation on its o1 model’s “chain of thought” reasoning. OpenAI’s refusal sparked a heated exchange with senior leaders, including then CTO Mira Murati, before the call abruptly ended, the report said.
Fast forward to today, AI researchers in Suleyman’s team believe they have made significant progress on the second of their two key priorities.
Under the leadership of Suleyman’s deputy, Karén Simonyan, the team has successfully trained a series of Microsoft models called MAI. These models reportedly perform at a level comparable to top models from OpenAI and Anthropic on widely recognised benchmarks.
First reported last year, MAI—internally referred to as MAI-1 (possibly Microsoft AI-1)—is being developed by the company and is around 500 billion parameters in size.
The company is also training reasoning models that use chain-of-thought techniques to rival OpenAI’s offerings. Microsoft is considering releasing the MAI models later this year as an API, allowing developers to integrate them into external applications, a move that would position Microsoft in direct competition with OpenAI’s API services.
Salesforce chief Marc Benioff recently commented that OpenAI chief Sam Altman and Suleyman are not exactly “best friends”. Notably, Microsoft first offered the AI chief role to Altman following his dramatic firing from OpenAI in 2023.
Microsoft has Moved On
To further hedge its bets, Microsoft has begun testing models from OpenAI competitors, including Anthropic, xAI, DeepSeek, and Meta, as potential replacements for OpenAI in its Copilot tools, which are integrated into products like Windows and Edge.
Notably, Microsoft recently announced that distilled versions of the DeepSeek-R1 models, the 7 billion and 14 billion parameter variants, will be available on the Copilot+ PCs.
His comments come after recent reports surfaced that OpenAI is planning to shift its entire workload to Project Stargate, leaving Microsoft Azure. Not to forget that Microsoft is no longer OpenAI’s exclusive cloud partner.
In a recent blog, OpenAI announced a new large-scale commitment to Azure, which will continue supporting all its products and model training. However, the agreement now allows for more flexibility.
Instead of exclusivity, Microsoft has a right of first refusal on any new capacity OpenAI wants to add. This means Microsoft gets the first chance to match any other cloud provider’s offer before OpenAI can move forward with them.
