Microsoft Goes All-In on Model Independence

Microsoft Goes All-In on Model Independence

Microsoft just announced seven new MAI models at Build 2026, and the subtext is clear: they're done being OpenAI's biggest customer [5][6]. The lineup includes MAI-Thinking-1 for reasoning, MAI-Code-1-Flash (137B params, 5B active) for GitHub Copilot, and variants for image, voice, and transcription tasks.

The key detail: these models use "clean licensed data" with no distillation from third-party models [6]. Microsoft is building from scratch to own their AI stack completely. MAI-Code-1-Flash will power the next generation of GitHub Copilot and VS Code integration, while MAI-Thinking-1 is already rolling out to select enterprise partners.

This is Microsoft's declaration of AI independence. They've learned from being dependent on OpenAI's roadmap and pricing. Now they control the entire pipeline from model training to enterprise deployment, with cost and performance optimized for their specific use cases.

Generalist AI Raises $400M, Physical AI Gets Real Money

A San Francisco robotics startup called Generalist AI just closed $400 million at a $2 billion valuation, with NVIDIA's NVentures leading alongside Bezos Expeditions and Fei-Fei Li [7][8]. They're building GEN-1, a foundation model for embodied robot learning that can handle complex physical tasks.

The funding round perfectly times with NVIDIA's Cosmos 3 release. We're seeing the physical AI stack come together: foundation models for world simulation, massive funding for robotics applications, and the compute infrastructure to train it all. This isn't lab research anymore — it's a coordinated industry push toward robots that can actually work in unstructured environments.

The investor lineup tells the story. NVIDIA wants customers for their robotics compute platform. Bezos sees the Amazon warehouse automation potential. Fei-Fei Li understands the computer vision foundations. Everyone's betting that 2026 is when physical AI moves from demos to deployment.