Mustafa Suleyman's Bid to Make Microsoft a Top-4 AI Lab

Mustafa Suleyman told The Verge at Build 2026 that his goal is to make Microsoft one of the top four AI labs globally, placing it alongside OpenAI, Anthropic, and Google DeepMind, roughly nine months after a renegotiated contract with OpenAI formally freed his division to build frontier models from scratch, VentureBeat reported in June 2026. The same week, Bloomberg reported that Suleyman called Anthropic's models "extremely expensive" and said Microsoft's near-term objective is to eliminate what it pays for external AI inference entirely.

Breaking Free From the OpenAI Contract

For five years, Microsoft's partnership with OpenAI gave it commercial access to the most capable frontier models available, at a cumulative investment of $13 billion across multiple tranches beginning in 2019. The arrangement also carried a constraint: Microsoft's AI products were effectively structured as a distribution channel for OpenAI's technology rather than an independent lab. That changed in November 2025 when the two companies amended their agreement to grant Suleyman's division formal authority to train large-scale frontier models independently, Fortune reported at the time.

"We were only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence," Suleyman told VentureBeat in June. The revised terms launched what he named the MAI Superintelligence Team, built around a framework he calls "humanist superintelligence", AI that, in his framing, is engineered to serve people rather than pursue open-ended objectives. Suleyman rejected the distillation approach used by DeepSeek and others, where capabilities are compressed from larger models at reduced cost. The MAI team was instructed to build from first principles using Microsoft's proprietary data pipelines and its Maia custom silicon.

The financial logic is straightforward. The $13 billion invested in OpenAI was a licensing fee paid to a competitor; converting that relationship to an in-house R&D function redirects the margin to Microsoft's own product revenue. Suleyman's bet is that Microsoft's infrastructure scale can substitute for the years of model development that its competitors accumulated before Microsoft entered the field.

Seven Models in Six Months

At Build 2026 in late May, Microsoft shipped seven MAI models spanning five capability areas: reasoning, code, vision and image generation, transcription, and voice. The flagship reasoning model, MAI-Thinking 1, delivers performance comparable to an earlier generation of Anthropic's Claude Opus 4.6 on programming tasks at a lower per-token cost, Microsoft said at the conference. "We got here in six months, which is itself a remarkable achievement," Suleyman told Semafor's Gina Chua in a June 2 interview. "We're now neck and neck with essentially what was state of the art just a few months ago."

The deployment strategy is tightly integrated with Microsoft's existing developer products. Suleyman told Semafor that effective model performance depends on matching the model to the harness: "It's very important that the model is tuned to the harness." All seven MAI models are being deployed within GitHub Copilot and the VS Code AI tooling, where Microsoft controls the full stack from model training through to the developer interface. The company also restricted employee access to Anthropic's Claude Code coding assistant, The Decoder reported, directing staff to use its own Copilot tools instead. The decision functions both as a cost measure and as a quality signal: if internal developers can rely on MAI models for their own work, the product is ready for external customers.

The MAI-1 flagship model was reported to carry approximately 500 billion parameters, making it competitive in scale with the largest publicly known frontier models. Suleyman has been explicit that size alone is not the objective; the team is optimizing for specific task performance within Microsoft's product ecosystem rather than benchmark leaderboard positioning.

Pricing Anthropic and OpenAI Out of the Stack

The competitive framing Suleyman has chosen is primarily economic. Bloomberg reported on June 4 that he described Anthropic as "extremely expensive" and said Microsoft's goal is to eliminate what it pays for external AI APIs. The Next Web reported the direct language: Microsoft wants to "eliminate" the cost of Anthropic, not merely reduce it. The Decoder subsequently confirmed that Copilot is phasing out OpenAI and Anthropic models in favour of MAI-family replacements across its product line.

The cost argument is well-timed. StartupHub.ai tracks 177 companies building frontier AI models and large language models globally, most of which are pre-Series B, suggesting enterprise customers are still in early stages of vendor selection. A Microsoft product stack that prices below Anthropic and OpenAI while offering comparable benchmark performance has structural advantages in that environment, particularly among enterprises already running Microsoft 365 and Azure. Suleyman told Semafor: "We are one of the largest tech companies in the world, and we have the resources to make sure that we do catch up."

He also told Semafor that Microsoft has less than 1 percent penetration of global coding and reasoning model usage, meaning the market for its MAI models is almost entirely untapped. The pricing strategy reflects that: capturing a large volume of low-margin inference revenue now is preferable to holding a premium position in a market that is still forming.

Suleyman told Fortune that he believes all white-collar work could be automated by AI within 18 months, a position that clarifies why he views the competitive window as narrow. For a company whose core revenue comes from productivity software used by white-collar workers, owning the AI layer is not an optional strategy.

What It Means

Suleyman's description of Microsoft's position as "the greatest game of catchup ever played" is accurate about the competitive gap and optimistic about the timeline to close it. Google DeepMind has 16 years of frontier model research behind it; OpenAI has 11. Suleyman's team achieved first-generation parity in under one, which is a meaningful result, but parity at a given capability level is not the same as the institutional depth that drives sustained model improvement. The MAI Superintelligence Team is still building the research infrastructure, data pipelines, and evaluation frameworks that its competitors built over years. Suleyman's argument is that Microsoft's financial resources and Azure compute can substitute for that accumulated advantage. His competitors' argument is that they cannot. The answer will be visible in the model releases that follow Build 2026.

What is not in dispute is the structural shift. Microsoft spent over a decade as OpenAI's distributor; it is now OpenAI's direct competitor in the market for developer AI tooling and enterprise inference. That change alone reshapes the economics of the industry. Related coverage: Satya Nadella's $37B AI Run Rate: The July 29 Earnings Breakdown.

Sources

• Semafor: Microsoft's AI chief on the greatest game of catchup ever played
• VentureBeat: Microsoft AI chief says company was set free from OpenAI to pursue superintelligence
• Bloomberg: Microsoft Says Anthropic Models Are Too Expensive
• Fortune: Microsoft launches new AI humanist superintelligence team
• The Next Web: Microsoft's AI chief says the company wants to eliminate what it pays Anthropic
• The Decoder: Copilot goes cheap as Microsoft phases out OpenAI and Anthropic models
• Fortune: Microsoft AI chief gives it 18 months for all white-collar work to be automated by AI

Editorial standards: every claim is sourced. Tips: [email protected]