Microsoft’s signals shift from OpenAI with launch of first in-house AI models for Copilot

After years leaning on OpenAI’s foundation models, Microsoft’s in-house AI division has this week underlined the software giant’s determination to strike out on its own.

On Thursday, after months of speculation that something important was coming, Microsoft AI finally announced MAI-Voice-1 and MAI-1-preview, the first models developed and trained entirely in-house.

MAI-Voice-1 is a natural speech generation model that the company claims can generate a minute of audio in under a second while running on a single GPU. It is already being used in Microsoft’s Copilot Daily to generate AI news summaries, and with the Copilot Podcasts feature which lets users create on-the-fly podcasts from prompts, Microsoft said.

Microsoft has also shown off the ability of MAI-Voice-1 to generate “high-fidelity, expressive audio across both single and multi-speaker scenarios,” in a separate Copilot Labs demonstration.

Meanwhile, the company has begun testing of MAI-1-preview on LMArena, a community site that lets users evaluate different AI models against one another, famously used by Chinese startup DeepSeek to grab attention in January. Developers will also be able to get their hands on the API by applying to Microsoft.

“We will be rolling MAI-1-preview out for certain text use cases within Copilot over the coming weeks to learn and improve from user feedback,” Microsoft said.

Independence day

According to Microsoft, MAI-1-preview uses an in-house mixture-of-experts model that was pre-trained and post-trained on 15,000 Nvidia H100 GPUs, a more modest infrastructure than the 100,000 H100 cluster sizes reportedly used for model development by some rivals. However, with an eye to ramping up performance, Microsoft AI is now running MAI-1-preview on Nvidia’s more powerful GB200 cluster, the company said.

For Microsoft, and perhaps the entire AI sector, the models’ appearance marks an important moment. Hitherto, the company has relied on models from OpenAI, a company it has invested a reported $13 billion in since 2019 and which uses the Azure cloud platform to host its models and services.

However, in 2024 Microsoft surprised the industry by hiring former DeepMind founder Mustafa Suleyman and his Inflection AI startup team to develop AI models for Microsoft AI, independent of OpenAI. Inflection AI continues to develop its own enterprise AI with a new CEO, while the former founders and core team work on Microsoft’s Copilot products.

Microsoft and Suleyman have been tight-lipped about what this means for the relationship between the company and OpenAI, and the idea that the companies are moving apart remains mere speculation. Nevertheless, Microsoft clearly hired Suleyman and his team for a reason.

“What we’re committed to is that we have optionality,” Suleyman said in an interview with news site Semafor this week. “We may use models from third-party developers. We will certainly continue to use OpenAI models for a long time to come, and we will use open source models as we already do.”

Suleyman said he believed that the “orchestrator”,  an intermediate platform that routes queries to different AI models depending on capability, would be key to how the company uses AI going forward. This suggests that while Microsoft is not moving away from OpenAI, it sees the importance of having an alternative developed by its own engineers.

Another signal of the company’s direction in the announcement is that, in the short term, Microsoft AI model development is being directed towards Copilot in the consumer space rather than for enterprises.

“We have big ambitions for where we go next. Not only will we pursue further advances here, but we believe that orchestrating a range of specialized models serving different user intents and use cases will unlock immense value,” Microsoft’s announcement said.

Coming after a period of news speculation about where Microsoft AI was going with its in-house model development, the announcement does at least put to bed the idea that its dash for AI independence had somehow gone off track.