Microsoft (MS) announced on the 28th (local time) that it has begun publicly testing its self-developed artificial intelligence (AI) model.
Mustafa Suleyman, CEO of MS's AI division, explained that this model, called "MAI-1-Preview," is the "first foundation model trained entirely in-house from start to finish" by MS. The text-based model MAI-1-Preview serves as a supplementary model to enhance the performance of MS's flagship AI service, "Copilot," focusing on expense efficiency.
The model was trained using approximately 15,000 NVIDIA H100 chips. This is compared to "Grok," developed by Elon Musk's AI corporation xAI, which utilized more than 100,000 chips. Suleyman noted, "We maximized performance with minimal resources using open-source technology." MS plans to apply this model to some functions of Copilot in the coming weeks, reflecting user feedback and making improvements.
MS has relied on OpenAI models to integrate AI capabilities into its major products, but the release of this self-developed model is interpreted as a strategic move to reduce dependency. MS remains a key investor and partner of OpenAI, having invested over $13 billion, and OpenAI utilizes MS's cloud infrastructure. However, recently, MS included OpenAI in its official list of competitors in its annual report, and OpenAI is also utilizing other cloud corporations such as CoreWeave, Google, and Oracle, thus evolving their relationship into one of both cooperation and competition.
MS announced that it is developing the next model at the world's largest data center equipped with NVIDIA's latest chip, Blackwell-based GB-200. Additionally, it revealed its voice AI model, "MAI-Voice-1," which MS claims is capable of generating one minute of audio in less than one second.