Microsoft deploys proprietary AI models to slash enterprise compute costs
Microsoft has launched new proprietary AI models that drastically reduce compute costs for enterprise applications, signaling a strategic shift away from its reliance on OpenAI for routine workloads.
Microsoft AI introduced two proprietary models into public preview on Wednesday, publishing deployment data that demonstrates significant reductions in compute expenses. The portfolio includes MAI-Image-2.5-Pro for high-fidelity graphics and MAI-Voice-2-Flash for high-volume enterprise speech workloads.
The accompanying production metrics reveal that these in-house tools can drastically lower operational expenses compared to third-party frontier models. In PowerPoint, the image model reduces GPU costs by up to 84 percent relative to OpenAI’s GPT-Image-2.
For high-volume voice workloads powering Dynamics 365 Contact Center, the new speech model cuts GPU costs by as much as 89 percent. Microsoft also reported a 26 percent increase in save rates and a 25 percent drop in latency for image editing in OneDrive.
The pricing structures reflect a tiered approach to enterprise AI spending. MAI-Image-2.5-Pro is priced at $106 per million image output tokens, while MAI-Voice-2-Flash costs $15 per million characters for high-volume applications.
This deployment marks a deliberate pivot in how Microsoft structures its artificial intelligence infrastructure. Chief Executive Satya Nadella stated the company is "beginning to route traffic" to its own models "whenever our models match or outperform frontier alternatives."
The move redefines Microsoft’s relationship with OpenAI, its primary AI partner. While Nadella noted that frontier models from OpenAI and Anthropic remain part of the orchestration system, the company is increasingly treating them as interchangeable components rather than exclusive foundations.
The strategy also alters hardware utilization across Microsoft's data centers. A proprietary coding model launched in GitHub Copilot runs efficiently on older Nvidia H100 and A100 GPUs, reserving the newest GB200 clusters for training rather than routine inference.
Industry leaders are taking note of the performance gains. Rob Reilly, global chief creative officer at WPP, called the image model "a strong leap forward for GenMedia tools" that shows Microsoft has "firmly established itself among the leaders in generative AI."
Market observers and developers have largely welcomed the focus on task-specific efficiency over blanket frontier deployments. One user summarized the financial appeal, noting that "cost and performance both improve when you stop overusing the biggest model."
However, some industry watchers remain skeptical about Microsoft's ability to execute this vision seamlessly. Critics point to historical struggles with user feedback, suggesting the company must prove it can maintain quality while aggressively optimizing for cost.
