Nadella Says Microsoft Overcorrected on OpenAI Reliance, Unveils 7 In-House MAI Models

Nadella's Candid Admission

Speaking with Stratechery's Ben Thompson on 3 June 2026, alongside Microsoft Build 2026, CEO Satya Nadella offered an unusually frank assessment of the company's AI journey. Asked whether Microsoft had been "lulled to sleep" by offloading too much of its AI strategy to OpenAI, he did not deny it, instead framing the situation as one that has produced a stronger competitive posture. "Here we are competing with a bunch of new people, a bunch of old people, and we have our own game," he said. Nadella described the broader shift as moving from "consuming a frontier model to fully participating at the frontier."

Seven New MAI Models

At Build, Microsoft unveiled seven in-house MAI models, anchored by MAI-Thinking-1, its first dedicated reasoning model, and MAI-Code-1-Flash, its first purpose-built coding model. MAI-Thinking-1 is a mid-sized model trained from scratch on clean, commercially licensed data, with 35 billion active parameters. Independent raters preferred it over Anthropic's Claude Sonnet 4.6 in blind testing, and it matches Claude Opus 4.6 on SWE-bench Pro coding tasks. MAI-Code-1-Flash, a compact 5-billion-parameter model, outperforms Claude Haiku 4.5 by 16 percentage points on SWE-Bench Pro while using 60% fewer tokens, and is rolling out to GitHub Copilot users in Visual Studio Code.

Nadella described a "dual frontier" approach: Microsoft maintains its own MAI lineage with clean data lineage whilst continuing to use OpenAI's intellectual property through to 2032, comparing outputs across both to improve quality. Mustafa Suleyman, who leads Microsoft's rebranded Superintelligence team, told Fortune the restructured OpenAI deal now lets Microsoft train larger frontier models, though he estimated it would take two to three years to match the top labs.

GPU Discipline and Enterprise Priorities

Nadella was direct about infrastructure trade-offs, revealing Microsoft is deliberately withholding GPU capacity from what he called "Neolabs" — smaller AI companies seeking raw compute on Azure — to prioritise enterprise customers and internal model development. "We're not selling raw GPUs to a bunch of Neolabs," he said, characterising such sales as "easy money" Microsoft is forgoing in favour of long-term positioning across its hyperscale cloud, its own applications, and research compute for MAI models.

The End of Pure SaaS Pricing

Nadella also predicted that traditional per-seat software pricing will give way to a hybrid model combining subscriptions with usage-based billing, answering "100%" when asked if hybrid pricing would define the future. He pointed to autonomous agents that consume tokens continuously as the driving force, identifying coding, security, and knowledge work as the three domains where Microsoft will concentrate its agent businesses. The company says nearly 90% of Fortune 500 firms now have active agents built with Microsoft tools, and its AI business has surpassed a $37 billion annual revenue run rate.