Microsoft has initiated public testing of its MAI-1-preview AI language model. This move signifies Microsoft’s intention to decrease its dependence on OpenAI for powering AI features in products like Copilot. The model is currently accessible for evaluation on LMArena, a platform designed for benchmarking AI models.
Microsoft has indicated a phased rollout of MAI-1-preview for specific Copilot text functionalities in the coming weeks. This gradual implementation aims to gather user feedback and refine the model’s performance. Developers interested in exploring MAI-1-preview can request early access through an online form provided by Microsoft. Initial assessments on LMArena positioned MAI-1-preview in 13th place for text-based tasks. This ranking places it behind models developed by OpenAI, Google, Anthropic, and xAI.
The development of MAI-1-preview involved substantial computational resources. Microsoft stated that the model was trained utilizing approximately 15,000 Nvidia H100 GPUs, alongside a cluster of Nvidia GB200 chips. These resources underscore the scale of investment and infrastructure required for training advanced AI models.
The evolution of strategic partnerships into competitive dynamics is evident through the changing relationship between Microsoft and OpenAI. In 2019, Microsoft made an initial investment of $1 billion in OpenAI, solidifying its position as OpenAI’s exclusive cloud provider via Azure. Over the subsequent five years, Microsoft’s total investment in OpenAI has exceeded $13 billion. Concurrently, Microsoft has been developing its own competing AI models, ultimately recognizing OpenAI as a competitor in its annual reports. This transition reflects OpenAI’s substantial growth, with a valuation now reaching $500 billion and ChatGPT boasting 700 million weekly users. This growth has transformed OpenAI from a collaborative research partner into a potential competitor to Microsoft’s AI ambitions.
The altered competitive landscape has resulted in practical adjustments to infrastructure strategies. OpenAI now distributes its cloud infrastructure needs across multiple providers, including CoreWeave, Google, and Oracle. This diversification reduces its exclusive reliance on Microsoft’s Azure platform. This pattern of initial collaboration leading to eventual competition is increasingly prevalent within the AI sector, where early partnerships facilitate capability development before companies ultimately compete for market share.
Microsoft’s strategy for building MAI-1 underscores the importance of talent acquisition in swiftly developing AI capabilities. The company recruited Mustafa Suleyman from Inflection, an AI startup, along with several of his colleagues. Additionally, Microsoft added approximately two dozen researchers from Google’s DeepMind in recent months. Suleyman’s background includes co-founding DeepMind prior to its acquisition by Google in 2014 and later leading Inflection as a competitor to OpenAI.
This experience provides Microsoft with established AI leadership and industry connections. This “acqui-hiring” strategy allows Microsoft to accelerate AI model development by integrating established teams rather than organically developing expertise. Microsoft has stated that MAI-1-preview represents its “first foundation model trained end to end in house,” suggesting that despite significant investment in OpenAI, the company recognized the need for independent AI capabilities.
