Vercel adds Microsoft AI speech and transcription models to AI Gateway with zero data retention

Vercel has struck a partnership with Microsoft AI (MAI) to make the MAI model family available through AI Gateway, its unified routing layer for AI models. According to the Vercel blog, AI Gateway is now one of only a limited number of platforms offering access to MAI models at all, and the deal is framed around a shared emphasis on security and zero data retention (ZDR).

The launch lineup

The first wave of MAI models on AI Gateway is audio-focused, with three entries:

• microsoft/mai-voice-2.1 generates speech with expressive delivery across 23 languages and holds the same speaker voice over long passages. Vercel points to narration, audiobooks, podcasts and lessons as cases where consistency from start to finish matters.
• microsoft/mai-voice-2.1-flash offers the same multilingual speech generation but tuned for lower latency. It is aimed at interactive use cases such as voice agents, assistants and spoken replies where fast turnaround is the priority.
• microsoft/mai-transcribe-2-streaming covers the reverse direction, emitting partial transcripts while audio is still coming in so applications can render live captions or track a conversation as it happens, rather than waiting for the full recording. The changelog notes that partials can be revised as more audio lands, so builders should replace the displayed text whenever a new partial arrives.

Wiring it into the AI SDK

Both capabilities are reachable from AI SDK 7 through two experimental helpers, generateSpeech and streamTranscribe (exported as experimental_generateSpeech and experimental_streamTranscribe). Vercel's sample code invokes the Flash model with Microsoft's Harper voice — passed as the string en-US-Harper:MAI-Voice-2.1-Flash — to render a short pickup notification into an MP3 file. A second example feeds a 16 kHz, 16-bit PCM audio stream into the transcription model and prints each transcript partial to the console as it arrives.

Pricing and data handling

Vercel says the MAI models are billed at their listed rates, with no platform fee or markup added on inference. On the data side, the company argues that MAI's zero data retention posture aligns with its own goal of giving AI Gateway users control over their data and transparency into how it is used for training.

The gateway itself supplies a single API for calling models, tracking usage and cost, and inspecting request traces, and it supports routing, retries and failover across the providers it fronts.

Why it matters

Two things lift this above a routine catalogue addition. First, access to MAI models remains scarce: by Vercel's own account, AI Gateway is one of only a few platforms carrying them, which gives developers a concrete reason to route through the gateway rather than negotiate directly with each provider. Second, the zero data retention commitment speaks to the compliance question that increasingly decides where enterprise AI traffic flows — prompts routed through these models are, by design, not retained for the provider's purposes.

For teams already building on Vercel's AI SDK, expressive speech generation across 23 languages and live streaming transcription become a provider swap rather than a new integration. And because the gateway handles routing, retries and failover, MAI can sit alongside other providers as one option among several instead of becoming another lock-in.