Microsoft plans to reduce OpenAI dependency

- Microsoft's AI chief, Mustafa Suleyman, stated the company is pursuing "true AI self-sufficiency" by developing its own frontier-level foundation models. This strategic shift follows a restructuring of their partnership with OpenAI in October 2025. - The company is developing an in-house family of AI models internally referred to as MAI. The text-based model, MAI-1-preview, was trained on approximately 15,000 Nvidia H100 GPUs and is intended to be integrated into Copilot. - Microsoft has also been developing a series of small language models (SLMs) called Phi. The Phi-3 family of open models are designed to be more cost-effective and outperform models of a similar size. For example, the Phi-3-mini has 3.8 billion parameters and can support a context window of up to 128K tokens. - Despite the move toward independence, Microsoft maintains a significant partnership with OpenAI, owning a 27% stake in the company, which is valued at around $135 billion. Their revised agreement extends Microsoft's access to OpenAI's intellectual property through 2032. - Microsoft has invested over $13 billion in OpenAI. As part of their agreement, OpenAI is contracted to purchase an additional $250 billion in Azure cloud services from Microsoft. - The move to reduce dependency is also driven by cost and performance concerns, with Microsoft looking to diversify the models used in products like Microsoft 365 Copilot. The company has already started hosting models from competitors like Anthropic, Meta, and Mistral on its Azure platform. - Following the announcement of its reduced reliance on OpenAI, Melius Research and Stifel downgraded Microsoft's stock from "buy" to "hold," citing concerns over the loss of AI exclusivity and the increased capital expenditures required to compete with Google and Amazon in the AI space. - Enterprise adoption of open-source models is a growing trend, with 76% of companies using large language models choosing open-source options, often alongside proprietary ones. This shift is driven by the desire for more control, customization, and cost savings, with open-source models being up to 10 times cheaper than comparable closed-source APIs.