Microsoft distances itself from OpenAI and develops its own AI models

Microsoft is changing its artificial intelligence strategy. Mustafa Suleyman, head of AI at the Redmond giant, confirmed to the Financial Times that the company is now developing its own cutting-edge models that will be launched in 2026. This strategic shift marks a sharp break with Microsoft's historical reliance on OpenAI, in which the firm nevertheless invested $13 billion.

A self-sufficiency strategy

Mustafa Suleyman, a former co-founder of Google DeepMind who joined Microsoft in 2024, clearly stated the company's ambition: " We must develop our own foundation models, at the absolute cutting edge of technology, with gigawatt-scale compute power and some of the best AI training teams in the world ". This statement comes after a restructuring of the ties between Microsoft and OpenAI in October 2025, which relaxed the exclusivity terms between the two partners.

Microsoft is investing heavily in assembling and organizing large datasets required to train advanced systems. The stated objective is to conquer the enterprise market with a professional-level AGI, AI tools powerful enough to accomplish the daily tasks of knowledge workers. According to Suleyman, " most tasks of white-collar workers sitting at a computer — lawyer, accountant, project manager or marketer — will be fully automated by an AI within 12 to 18 months ".

The first in-house models already operational

Microsoft took a concrete step in August 2025 with the launch of MAY-1-preview, described as " in-house mixture-of-experts models " which was pre-trained and post-trained on around 15,000 NVIDIA H100 GPUs. This model is designed to provide useful responses to everyday queries and to follow instructions effectively. Microsoft plans to gradually integrate it into some of Copilot's text-based use cases.

At the same time, Microsoft deployed MAY-Image-1, its first internally developed image-generation model, which illustrates the company's desire to strengthen its technological autonomy. These successive launches demonstrate that Microsoft is not content with making announcements: the American giant is building models at significant scale, on the same hardware infrastructure as its competitors.

The Maia 200 chip to challenge Nvidia

In January 2026, Microsoft unveiled the Maia 200, its new chip positioned as an inference accelerator. designed to significantly improve the economics of AI token generation This chip, equipped with more than 100 billion transistors, delivers over 10 petaflops in 4-bit precision and about 5 petaflops of 8-bit performance, a substantial increase over its predecessor.

The new chipset directly targets Nvidia's solutions by pairing custom silicon with a software package intended to loosen CUDA's grip. According to Scott Guthrie, executive vice president for cloud and AI at Microsoft, the chip delivers 30% more performance compared to alternatives at the same price. Microsoft can connect up to 6,144 Maia 200 chips together, which reduces power consumption and total cost of ownership.

A partnership with OpenAI that endures nonetheless

Despite this strategic repositioning, Microsoft officially maintains its relationship with OpenAIThe deal negotiated in October 2025 guarantees Microsoft a 27% stake in OpenAI's for-profit arm and access to its most advanced models until 2032, including post-AGI models. Microsoft also retains rights to intellectual property and the exclusivity of the Azure API extends until 2032.

However, this arrangement gives OpenAI more freedom to seek new investors and infrastructure partners, potentially turning it into a direct competitor. OpenAI can now seek compute elsewhere, while Microsoft can diversify its model suppliers. Each is visibly preparing for the end of their exclusive partnership.

Major diversification

Microsoft has deliberately broadened its AI ecosystem by hosting models from multiple providers in its data centers. The company has invested heavily in Anthropic, a direct competitor to OpenAI, and even agreed to use Anthropic's models in Microsoft 365 Copilot experiments after internal tests found them better for certain Office tasks. This change involved paying AWS for access, a significant strategic choice.

The Windows and Microsoft 365 teams were ordered to adopt Claude Code, to the detriment of GitHub Copilot. This signal could not have been clearer: Microsoft wants to be the place where every winning model runs, and wants at least one winner to carry a Microsoft badge. This strategy aims to ensure that putting "Copilot everywhere" never depends on a single provider.

A long-term medical ambition

Beyond the enterprise world, Microsoft is also targeting the healthcare sector with the goal of build a medical superintelligence capable of helping to solve staffing crises and wait times in overloaded healthcare systems. Last year the company unveiled a diagnostic tool that would outperform doctors on certain tasks.

Mustafa Suleyman believes the ultimate role of AI is to serve as an "extraordinary teacher," an "extraordinary doctor," and a life companion. He explains that Copilot is designed to function like a wise advisor that enables individuals to reach their full potential, offering a personalized program to help each person learn what interests them.

A strained financial context for OpenAI

This strategic repositioning by Microsoft comes at an especially delicate time for OpenAI. The company is spending nearly a billion dollars a month and a prominent analyst at the Council on Foreign Relations predicts its bankruptcy within the next 18 months. Sam Altman is ramping up fundraising and desperately seeking liquidity across Silicon Valley. Even Nvidia has become hesitant about the idea of investing $100 billion in the company.

This Microsoft announcement therefore comes at the worst possible time for OpenAI, which is losing its biggest client and protector just as its financial survival is being questioned. Microsoft's AI pivot remains, however, a resounding commercial failure: despite $88.7 billion in capital expenditures in fiscal year 2025, Copilot is stuck behind ChatGPT and Gemini with only 1.1% market share.

Total control of the value chain

This commercial debacle is precisely pushing Microsoft to take control in-house rather than rely on an external partner. Developing its own models offers total control over costs, performance and the technological roadmap. Inference is the area where bills pile up and where hyperscalers want the most leverage.

Microsoft says that OpenAI remains its "leading model partner", but this claim hides another reality: the company is giving itself maximum latitude to negotiate, pivot and replace its suppliers as needed.

The article "Microsoft distances itself from OpenAI and develops its own AI models" was published on the site Abondance.