Running AI features for hundreds of millions of everyday users gets ridiculously expensive fast. For a while, Microsoft didn't seem to mind paying OpenAI every time someone generated an image in PowerPoint or transcribed a call. They just wanted shiny features out the door as fast as possible to win the headline race. That honeymoon period is officially over.

Microsoft has quietly started kicking third-party models out of its software stack and replacing them with its own tech, branded under the name MAI. 

They recently announced two new variants called MAI-Image-2.5-Pro and MAI-Voice-2-Flash. The technical specs are standard industry updates, but the real story is how aggressively Microsoft is shoving these models into its actual products.

Bing Image Creator is now running entirely on Microsoft’s homebuilt models. Key image editing features in OneDrive made the switch too, which Microsoft claims improved latency and saved system resources. 

But the biggest hit is inside PowerPoint. Ditching GPT-Image for their own native MAI engine allegedly chopped GPU processing costs by 84 percent. When you operate at Microsoft scale, an 84 percent drop in data center hardware costs represents millions of dollars saved.

The same cost-cutting push is hitting their voice tools. The new Flash model is getting jammed directly into Dynamics 365 Contact Center. That software handles customer support lines for giant companies like T-Mobile and EasyJet. Microsoft says swapping in their own voice engine lowers GPU expenses by up to 89 percent while running twice as fast. Even Dragon Copilot, the dictation tool used by 170,000 healthcare providers, is ditching third-party models across 58 languages to use Microsoft’s internal transcription engine instead.

This whole move comes down to basic math. Paying external vendors for model access across millions of enterprise applications was always going to burn a hole in their margins. OpenAI might still build impressive frontier research, but routine corporate work does not require paying a markup to a partner. Generating a graphic for a presentation slide or routing a call center ticket needs to be cheap.

Microsoft wants full ownership of its infrastructure from top to bottom. Swapping out partner models saves them a fortune on server bills and keeps their own software locked into their own cloud. You can expect them to keep quietly pulling the plug on outside tech everywhere else they can.