Microsoft tests first-party MAI-Realtime to reduce OpenAI dependence in Voice Live

Microsoft is testing MAI-Realtime, a first-party full-duplex speech-to-speech layer designed to replace its dependence on OpenAI's GPT-Realtime within Azure's Voice Live API. The effort is significant because it signals Microsoft's continued push to build in-house frontier capabilities under its MAI (Microsoft AI) banner and reduce reliance on OpenAI even as the two remain deeply intertwined commercially.

That intertwining was underscored by earnings: Microsoft stock surged 24.8% after reporting Q4 revenue of $90.0 billion and net income of $35.8 billion, with full-year revenue of $331.8 billion, driven heavily by Azure and Microsoft 365 Copilot AI demand. The company acknowledged that most of its current AI revenue still comes from OpenAI — a dependency MAI-Realtime and other first-party models are meant to gradually diversify.

Strategically, voice is a deliberate wedge: real-time speech-to-speech is a high-value, latency-sensitive capability where owning the model improves margins and control versus paying OpenAI per call. It fits a broader pattern of hyperscalers building first-party alternatives to their model-provider partners — mirrored by Perplexity CEO Aravind Srinivas noting his team tested GPT-5.6 Terra and made it the default for subagents, illustrating how quickly orchestration layers swap underlying models on cost and capability.

Azure also shipped infrastructure updates this week, including GA of its Virtual Network routing appliance (200+ Gbps private connections) and integration of GitHub Copilot Agent into the Microsoft Agent Framework for production .NET/Python agents with governance and approval workflows. The open question on MAI-Realtime is quality: matching GPT-Realtime's naturalness and latency is hard, and Microsoft hasn't published benchmarks or a GA timeline. Watch whether it ships as a true default or remains an opt-in fallback while OpenAI keeps powering the premium path.