At the Build 2026 developer conference in San Francisco, Microsoft CEO Satya Nadella made clear the company intends to structurally reduce OpenAI dependence: seven newly unveiled MAI models were developed entirely internally, without training data or outputs from external AI systems. The constellation is unusual: Microsoft is simultaneously primary investor, infrastructure provider, and direct OpenAI competitor in the model market. In parallel, the company presented Majorana 2, a redesigned quantum chip whose qubits reportedly are 1,000 times more reliable than the previous generation.
From OpenAI distributor to direct developer
Since 2019, Microsoft has invested more than 13 billion dollars in OpenAI and supplies the compute infrastructure for OpenAI's models via Azure cloud. Simultaneously, Microsoft integrates OpenAI models into its products, from Windows to GitHub Copilot. This dependence had strategic weaknesses: pricing, availability, and development direction lay outside Microsoft's control.
The answer came with hiring Mustafa Suleyman, cofounder of Google DeepMind and CEO of Inflection AI until 2023. Suleyman leads Microsoft's superintelligence team and oversaw the new model line. On stage at Build, he made the break explicit: the MAI models emerged with zero distillation from third-party models, without borrowing training data or outputs from external AI systems.
Seven models for seven tasks
The flagship is MAI-Thinking-1, Microsoft's first reasoning model optimized for multi-step complex tasks like code development and document analysis. Per Microsoft's documentation, it has 35 billion active parameters and a context window of 256,000 tokens. Surge, an evaluation firm, tested the model and reported MAI-Thinking-1 was preferred in user surveys versus Claude Sonnet 4.6 and matched Claude Opus 4.6 on programming benchmarks. These results remain unverified by independent auditors.
Alongside it, Microsoft introduced six more models: MAI-Code-1 for programming, embedded directly in GitHub Copilot and Visual Studio Code. MAI-Image-2.5 and its faster Flash variant for generating and editing images, placed by Microsoft at third position on the Arena AI leaderboard for image generation. MAI-Transcribe-1.5 for automatic speech recognition in 43 languages. MAI-Voice-2 with new voice options for 15+ languages and corresponding Flash variant.
Practically significant is less the model count than distribution strategy: all models land first in existing Microsoft products like Office, Teams, Windows, and GitHub. The company builds no open competing ecosystem to OpenAI or Anthropic, but leverages installed base as distribution channel. For millions of Microsoft 365 users, the switch will initially remain invisible.
What follows for users and competitors
Microsoft's own developments shift market dynamics without formally ending the OpenAI partnership. The company remains investor and infrastructure provider for OpenAI; shared Azure contracts run for years more. Nadella spoke of complementarity on stage. Behind that is sober calculation: Microsoft can keep selling OpenAI models for premium applications and integrate own models for standard tasks into its products, improving margins.
For European enterprise, the shift brings little immediate relief: Amazon, Microsoft, and Google already control over 70 percent of European cloud market per the EU Commission. Microsoft's own AI models strengthen this position rather than dissolve it. Add regulatory pressure: the EU's AI Act takes effect August 2, 2026, with transparency requirements. General-purpose models like MAI-Thinking-1 mean documentation obligations, technical descriptions, and data/energy disclosures. Microsoft has roughly two months.
Majorana 2 and the 2029 promise
The conference's second highlight was Majorana 2, Microsoft's quantum chip. The successor to Majorana 1, unveiled January 2025, uses redesigned materials: lead instead of aluminum as superconductor; a semiconductor region now combines indium arsenide with indium arsenide-antimonide. Microsoft claims qubit reliability improved by a factor of 1,000, with average qubit stability of 20 seconds and peak values up to one minute.
The target date is ambitious: 2029 should see the first commercially usable quantum computer based on this technology, with one million qubits on a palm-sized chip. Microsoft halved its original timeline.
Trade publication heise responded skeptically. A commentary titled Oops, they did it again noted Microsoft's quantum research had repeatedly raised high expectations unmet before. No peer-reviewed publications on Majorana 2 exist. Quantum physicists emphasize qubit reliability is only one of several hurdles to practical quantum computing: error correction and scaling to one million qubits are considered equally difficult. The 2029 target date remains a conference promise until independent verification.