Mustafa Suleyman’s Roadmap to AI ‘Self-Sufficiency’

Microsoft is taking significant steps to achieve AI ‘self-sufficiency’ under the leadership of Mustafa Suleyman. The company has announced a 6-step plan to reduce its reliance on OpenAI and develop its own AI capabilities.

Step 1: Creation of a Dedicated ‘Super-intelligence’ Team

A new internal research unit, the MAI Super-intelligence Team, has been created to build ‘frontier-grade’ models that can replace OpenAI’s offerings in Microsoft products. The team consists of ~120 researchers and engineers and is led by Suleyman.

Step 2: Massive Internal Compute & Custom Silicon Programme

Microsoft is investing $2 billion in a custom AI-chip cluster, Azure-AI-Silicon, to run training and inference for the new models. The target is to achieve > 200 petaflops of AI-optimized compute by mid-2025.

Step 3: Development of Microsoft-Owned Models

The company is developing its own models, including MAI-1 and MAI-2, with the goal of replacing OpenAI’s models in Microsoft products. MAI-1 has already been trained at a cost of ≈ $150 million, and MAI-2 is planned for early 2026 at a cost of ≈ $400 million.

Step 4: Data-Ownership Strategy

Microsoft is consolidating all Azure-customer interaction data, GitHub Copilot logs, and LinkedIn activity into a private data lake for model training. The target is to achieve > 10 PB of high-quality, consent-based data by 2026.

Step 5: Talent-Acquisition & Internal Culture Push

The company has hired over 200 former DeepMind/Inflection researchers, including co-founder Karen Simonyan, and is offering $1 million equity grants per senior scientist.

Step 6: Parallel Partnership with OpenAI

Microsoft has extended its OpenAI agreement to 2030, allowing the company to run its own models side-by-side with OpenAI’s models.

Implications for the Broader AI Industry

The move is expected to increase competitive pressure on OpenAI, accelerate the custom-silicon race, and intensify the talent war in the AI industry. It may also lead to a shift towards hybrid partnership models, where cloud providers work with AI companies while building their own in-house capabilities.

Sources