Microsoft has officially begun pursuing "superintelligence" using its own in-house research, data, and custom silicon after a contractual renegotiation with OpenAI roughly six months ago. Mustafa Suleyman, CEO of Microsoft AI, stated at Microsoft Build 2026 that the company is no longer solely dependent on OpenAI for frontier AI development. While Microsoft maintains its partnership with OpenAI, it has established the MAI Superintelligence Team to build its own sovereign AI capabilities.

The company introduced the MAI model family, a collection of seven in-house models spanning reasoning, code generation, and multimodal tasks. The flagship MAI-Thinking-1, a 35-billion-active-parameter reasoning model, was trained from scratch on commercially licensed data without distilling outputs from third-party models. Other models include MAI-Code-1-Flash for GitHub Copilot, MAI-Image-2.5 for image generation, and speech-generation system MAI-Voice-2. These models are available through Microsoft Foundry, the company's deployment infrastructure, which now allows developers to tune model weights through platforms like OpenRouter and Fireworks.

A key component of this strategy is Frontier Tuning, an enterprise capability allowing organizations to customize MAI models using proprietary data within secure compliance boundaries. According to Suleyman, this process shifts AI from simple conversation to autonomous action. Early deployments include healthcare-focused models at the Mayo Clinic and tax-advisory agents at EY. Microsoft reported that when tuned for specific organizational standards, its models achieved higher win rates at one-tenth the cost of competitors. The firm also launched Microsoft Scout, an always-on "Autopilot" agent built on open-source OpenClaw technology, designed to execute complex tasks across enterprise applications like Excel and Jira.

Microsoft’s compute strategy involves a massive scale of hardware investment, with Suleyman confirming the company is the world's largest buyer of Nvidia GB200 and GB300 accelerators. Simultaneously, Microsoft is deploying its second-generation AI accelerator, Maia 200, in data centers across Iowa, Arizona, Italy, Australia, and South Korea. Microsoft claims Maia 200 is 30 percent more cost-efficient than the GB200, and when co-optimized with MAI models, it provides an additional 1.4x improvement in performance per watt. Suleyman argues that model quality depends on high-quality, curated training data rather than mere brute-force scale, ensuring that Microsoft's vertically integrated stack provides a distinct competitive advantage in the enterprise market.

Microsoft has officially begun pursuing "superintelligence" using its own in-house research, data, and custom silicon after a contractual renegotiation with OpenAI roughly six months ago. Mustafa Suleyman, CEO of Microsoft AI, stated at Microsoft Build 2026 that the company is no longer solely dependent on OpenAI for frontier AI development. While Microsoft maintains its partnership with OpenAI, it has established the MAI Superintelligence Team to build its own sovereign AI capabilities.

The company introduced the MAI model family, a collection of seven in-house models spanning reasoning, code generation, and multimodal tasks. The flagship MAI-Thinking-1, a 35-billion-active-parameter reasoning model, was trained from scratch on commercially licensed data without distilling outputs from third-party models. Other models include MAI-Code-1-Flash for GitHub Copilot, MAI-Image-2.5 for image generation, and speech-generation system MAI-Voice-2. These models are available through Microsoft Foundry, the company's deployment infrastructure, which now allows developers to tune model weights through platforms like OpenRouter and Fireworks.

A key component of this strategy is Frontier Tuning, an enterprise capability allowing organizations to customize MAI models using proprietary data within secure compliance boundaries. According to Suleyman, this process shifts AI from simple conversation to autonomous action. Early deployments include healthcare-focused models at the Mayo Clinic and tax-advisory agents at EY. Microsoft reported that when tuned for specific organizational standards, its models achieved higher win rates at one-tenth the cost of competitors. The firm also launched Microsoft Scout, an always-on "Autopilot" agent built on open-source OpenClaw technology, designed to execute complex tasks across enterprise applications like Excel and Jira.

Microsoft’s compute strategy involves a massive scale of hardware investment, with Suleyman confirming the company is the world's largest buyer of Nvidia GB200 and GB300 accelerators. Simultaneously, Microsoft is deploying its second-generation AI accelerator, Maia 200, in data centers across Iowa, Arizona, Italy, Australia, and South Korea. Microsoft claims Maia 200 is 30 percent more cost-efficient than the GB200, and when co-optimized with MAI models, it provides an additional 1.4x improvement in performance per watt. Suleyman argues that model quality depends on high-quality, curated training data rather than mere brute-force scale, ensuring that Microsoft's vertically integrated stack provides a distinct competitive advantage in the enterprise market.