Just a few months ago, Microsoft looked more like a distribution giant than an artificial intelligence creator: the Redmond firm was building its empire on OpenAI's models, injecting tens of billions of dollars into a partnership that allowed it to place Copilot at the heart of Windows, Office, and Azure. This Tuesday, June 2, during its annual Build conference, Microsoft made a surprisingly discreet but brutally significant farewell: it now has its own models.

MAI-Thinking-1, spearheading a new family

The MAI (Microsoft AI) family is presented by the Microsoft AI Superintelligence Team, an entity founded in November 2024 and led by Mustafa Suleyman, co-founder of DeepMind — one of the most respected minds in the industry. The spearhead of this new offering is MAI-Thinking-1: a reasoning model with a MoE (Mixture of Experts) architecture and 35 billion active parameters. Its promise? To be more efficient than competing models with equivalent parameters, by tailoring specifically to the needs of businesses and developers using Azure.

Efficiency over raw power

Suleyman's strategy is clearly articulated: Microsoft does not claim to compete head-on with the most powerful models on the market — GPT-5, Gemini Ultra, or Claude Opus — but intends to occupy the efficiency flank, that of compact and specialized models that are cheaper to run and deliver better performance in specific areas. An approach strangely reminiscent of Google's strategy with its Gemini Flash models, or Anthropic's Haiku versions. The AI market is no longer solely a race for the most powerful model: it is an industry that is segmenting, specializing, and seeking to monetize colossal investments.

An $80 billion titans' war

The most direct consequence of this announcement is economic and relational: Microsoft is reducing its dependence on OpenAI and Anthropic, its two major model providers until now. This strategic pivot comes at a time when both startups are preparing for their IPO — a market entry that will be scrutinized by the entire tech industry as a valuation test for the generative AI sector as a whole. Google, for its part, is reportedly looking to raise up to $80 billion to support an AI infrastructure that demand continues to outstrip. A titans' war whose front lines are being drawn for the next eighteen months.

Editorial opinion

Microsoft's move is less technical than political. By developing its own models, Redmond ceases to be a commercial hostage of OpenAI and regains control of the only asset that truly matters in generative AI: gross margin. MAI-Thinking-1 may not shake GPT-5 on public benchmarks, but it will shake Sam Altman's contractual negotiations. For the sector, this marks the beginning of a maturity phase where value will no longer be measured solely in parameters, but in marginal cost per token. Suleyman understood this before others.

Key takeaways

  • MAI-Thinking-1: MoE architecture, 35 billion active parameters.
  • MAI family led by Mustafa Suleyman (ex-DeepMind, ex-Inflection AI).
  • Objective: Reduce Microsoft's dependence on OpenAI and Anthropic on Azure.
  • Google seeks to raise up to $80 billion for its AI infrastructure.
  • OpenAI and Anthropic are both preparing for an IPO.