Microsoft used its Build 2026 keynote on Monday to wheel out MAI-Thinking-1, the company’s first in-house reasoning model, and the framing was unsubtle. The model was trained from scratch on commercially licensed enterprise data. No third-party distillation. No “we used OpenAI outputs to bootstrap our training pipeline.” Mustafa Suleyman has been telegraphing this move since he took over Microsoft AI in 2024, and the Build 2026 stage was where the receipts finally landed.
The architecture: 35 billion active parameters out of roughly one trillion total in a sparse Mixture of Experts setup, with a 256,000-token context window. Microsoft is claiming the model beats Sonnet 4.6 in blind preference tests and matches Opus 4.6 on SWE Bench Pro, which if the third-party benchmarking holds up puts MAI-Thinking-1 in the same coding-quality tier as the model every enterprise procurement team has been signing annual contracts for. Available now in private preview through Microsoft Foundry.
The number worth circling is the hardware story. MAI-Thinking-1 was tuned for Microsoft’s own Maia 200 chip and benchmarked against Nvidia’s GB300. Microsoft is reporting a 1.4x performance-per-watt advantage running the MAI model end-to-end on its own silicon. That figure is a Microsoft claim on Microsoft hardware running a Microsoft model, so apply the appropriate squint, but it is also the first quantitative data point on whether the Maia program is going to be cost-competitive with Nvidia for Microsoft’s own workloads. If the 1.4x holds, Microsoft has a real reason to keep moving internal inference off the Hopper and Blackwell rentals it has been writing checks for.
What this announcement actually says to the market is that the OpenAI-Microsoft partnership is now structurally a hedged bet on Microsoft’s side. Azure still hosts OpenAI’s most lucrative workloads. Microsoft Foundry still ships GPT-5.5. But the in-house model is no longer a research project, the in-house chip is no longer a roadmap rumor, and Microsoft now has the ability to fall back to its own stack if the OpenAI relationship goes sideways during the OpenAI IPO. That optionality is the whole point.