Microsoft Quietly Swaps In Its Own AI Models To Cut Costs

The News

Microsoft has quietly started routing some of its most-used software features through AI models it built in-house, rather than leaning entirely on partners such as OpenAI and Anthropic. According to TechCrunch, the company is now serving a share of user prompts inside Excel and Word with its own MAI models, the family it unveiled last month.

The shift is small in scope but large in signal. For years, Microsoft's Copilot features across Office 365 were powered by OpenAI and, more recently, Anthropic systems. Now a "certain percentage" of prompts in two of the world's most widely used productivity apps run on Microsoft's own technology. When approached for comment, Microsoft told TechCrunch it had nothing further to share.

The move follows the debut of seven new MAI models at the Microsoft Build conference in June 2026, a line-up that included an agentic coding model and a text-to-image generator. Together they mark Microsoft's clearest attempt yet to reduce dependence on the very labs it has spent billions backing.

Why It Matters

Running frontier models is expensive, and inference costs, the price of answering each query, scale brutally when your software has hundreds of millions of daily users. Every prompt Microsoft answers with its own model instead of a partner's is one it does not pay a third party for. At Office's scale, even a modest slice of redirected traffic becomes serious money.

Microsoft is not alone. The report frames this as part of wider belt-tightening across the industry, with Amazon, Uber, Meta and Accenture all cited as trimming AI spending. The pattern echoes the cloud playbook of the 2010s, when hyperscalers first rented capacity, then built their own chips and data centres to escape other people's margins. The same logic is now arriving in the model layer.

The strategic tension is obvious. Microsoft remains OpenAI's largest backer, yet it is steadily building the capability to route around it. That is not a divorce, but it is a hedge, and it tells rivals and customers that owning the model is becoming as important as owning the application.

Indian Angle

For India, this previews a decision every large enterprise and IT-services firm will soon face: rent intelligence, or build it. TCS, Infosys, Wipro and HCLTech have layered generative-AI services on top of Azure OpenAI and similar platforms. If the platform owner itself is swapping in cheaper in-house models to protect margins, Indian services firms will feel pressure to do the same on their own delivery costs, especially for high-volume, low-complexity tasks.

There is also a sovereignty thread. India's IndiaAI mission and start-ups such as Sarvam and Ola-backed Krutrim argue the country needs home-grown foundation models rather than perpetual dependence on foreign APIs billed in dollars. Microsoft's pivot validates that thesis commercially: even the firm with the deepest OpenAI ties sees value in controlling its own stack. For Indian developers paying rupee salaries against dollar-denominated inference bills, cheaper first-party models could ease a real cost squeeze.

Finally, for enterprise buyers and regulators, the episode is a reminder that the model behind a familiar product can change silently. As MeitY shapes data-localisation and AI-governance rules, which model processes a sensitive Excel or Word prompt, and where, becomes a compliance question, not just a billing one.

FAQ

What exactly has Microsoft changed?

It has begun using its own MAI models to answer a portion of user prompts inside Excel and Word, instead of routing every request to partner models from OpenAI and Anthropic. Microsoft declined to elaborate beyond confirming the shift.

Why is Microsoft doing this now?

Inference costs rise with usage, and Office serves enormous volumes daily. Handling more prompts on in-house MAI models, seven of which launched at Build in June 2026, lets Microsoft cut payments to third-party labs while keeping its Copilot features running.

What does it mean for Indian tech firms?

Indian IT-services companies built heavily on Azure OpenAI may face pressure to adopt cheaper in-house or home-grown models to protect margins, echoing Microsoft's own move.

Where can I read the original coverage?

TechCrunch first reported the deployment. The full article is linked in the attribution below.

This story was reported by TechCrunch. Read the full original coverage at TechCrunch.