Microsoft bets on its own AI models to cut costs

On July 7, 2026, Microsoft became the latest Silicon Valley giant to announce a push to cut AI spending: the company plans to significantly increase the share of its own language models in its products and services.

Why Microsoft is reducing its reliance on outside models

AI features in Microsoft products have scaled so far that the cost of third-party models has become a significant line item in the operating budget. With hundreds of millions of users across Microsoft 365 Copilot, GitHub Copilot and other AI-integrated services, even a small difference in the cost of a single request adds up to hundreds of millions of dollars in annual savings.

The company is not starting from scratch. The Phi series of small language models, developed by Microsoft Research, has shown that compact, carefully trained models can match larger ones on many practical tasks while costing far less to run. Moving part of the workload from external APIs to its own models lets Microsoft control inference costs, speed up responses and tune behavior more precisely for a given product.

This does not mean a complete break with OpenAI: the partnership with the world's leading AI lab is not going away. Microsoft appears to be building a hybrid approach, in which GPT models handle the most resource-intensive and complex requests, while in-house models take on routine, standard tasks.

Who else is moving this way

Microsoft is joining a trend that has gained momentum among all the major Silicon Valley players.

• Meta is steadily moving its products, from Facebook to WhatsApp, onto its own Llama family of models instead of buying external AI APIs
• Apple has bet on on-device models for iOS and macOS, minimizing its dependence on cloud providers
• Google is balancing its commercial Gemini line with specialized internal models for specific services
• Amazon is developing the Titan models and a number of internal solutions for AWS and its own products

It is telling that Microsoft, OpenAI's largest corporate investor, has become one of the flagships of this movement. This is not a break but pragmatic diversification: use partner models where they are irreplaceable, and switch to in-house solutions where that makes economic sense.

What changes for the AI market

The shift of large corporations away from external APIs is changing the balance of power. AI model providers that counted on long-term growth in enterprise demand now have to rethink their strategy: they will have to compete not only on quality but also on price efficiency and customization options.

For the industry as a whole, the trend encourages further reductions in inference cost and the spread of specialized, lighter models. "Squeezing out the same quality for less money" is becoming the main technical priority of the coming years.

What this means

Microsoft's move into AI cost-cutting means the first era of "buy capacity from the leader" is ending. Large corporations are building their own AI stacks and vertically integrating model expertise. For OpenAI, Anthropic and other providers, this is a signal to look for value where in-house corporate teams have not yet reached parity: advanced reasoning, multimodality and specialized scientific tasks.

*Meta has been designated an extremist organization and is banned in Russia.