Microsoft has unveiled its first proprietary AI models, MAI-Voice-1, marking a significant step in reshaping its position in the ongoing AI race. MAI-Voice-1, a high-speed speech model, and MAI-1-preview, a text-based model that Microsoft calls “a glimpse of what’s coming next” inside Copilot.
MAI-Voice-1 is optimized for speed, capable of generating a full minute of audio in less than a second using just one GPU. It’s already powering features like Copilot Daily, which delivers short news summaries via an AI voice host, and helps create podcast-style conversations that simplify complex topics.
The second release, MAI-1-preview, is designed for text-based tasks. Trained on about 15,000 Nvidia H100 GPUs, it supports instruction-following and natural Q&A at scale. Users can already test it through Copilot Labs, and Microsoft says the model will soon be integrated more deeply into Copilot for everyday use.
These launches arrive as Microsoft remains closely tied to OpenAI, having invested over $13 billion into the ChatGPT maker, now valued at around $500 billion. OpenAI still depends on Microsoft’s cloud infrastructure, while Microsoft continues to integrate OpenAI’s models into Bing, Windows, and other products.
Yet, the two companies are increasingly becoming rivals. Microsoft now lists OpenAI alongside Amazon, Apple, Google, and Meta as competitors, while OpenAI has started diversifying its infrastructure providers, including CoreWeave, Google, and Oracle, to meet surging ChatGPT demand, currently drawing 700 million weekly users.
Early benchmarks show MAI-1-preview ranks 13th for text workloads on LMArena, trailing models from Anthropic, DeepSeek, Google, Mistral, OpenAI, and xAI.
While performance lags some peers, Microsoft frames this as the first step toward building its own foundation models at scale.
