Seven Models Trained From Scratch

The centerpiece came from Mustafa Suleyman, who runs Microsoft AI. He presented seven models spanning reasoning, coding, image generation, voice and transcription, all trained from scratch on licensed data with no distillation from rival labs. The flagship reasoning model, MAI-Thinking-1, uses a sparse mixture of experts design with roughly 35 billion active parameters and a context window of 256,000 tokens, and it is in private preview through Microsoft Foundry rather than general release. Microsoft reported that blind human raters preferred it to Anthropic's Claude Sonnet 4.6 and that it matched Claude Opus 4.6 on the SWE-bench Pro coding benchmark, results that come from the company's own evaluations and await outside testing. A smaller and more efficient coding model, MAI-Code-1-Flash, is reaching GitHub Copilot users inside the editor.

Microsoft describes the program behind these models as a hill-climbing machine, a training pipeline meant to improve cycle after cycle as global compute keeps scaling. The framing connects to silicon. In its keynote, Microsoft said it co-designed the MAI models with its Maia 200 inference accelerator and reported efficiency gains from pairing the two. It also introduced Frontier Tuning, which applies reinforcement learning inside a customer's own compliance boundary so a model adapts to how a specific business actually operates. Microsoft pointed to one internal example where task completion rose from 13% to 87% after tuning, and said a version adapted for Excel work matched a frontier OpenAI model at up to 10 times lower cost. All of these figures come from Microsoft and have not been independently checked.

A Full Stack, Not Just Models

The model news sat on top of fresh infrastructure, most of it early. Microsoft said its Azure Cobalt 200 Arm-based virtual machines, now in preview, can deliver up to a 50% improvement in processor performance depending on the workload, and the company is aiming them at Linux-based agentic AI. It also added Azure HorizonDB, a Postgres-compatible service for AI applications with features such as vector search and connections into Foundry and Fabric. A version of its Fabric data warehouse with GPU acceleration ran up to seven times faster than three rival cloud warehouses in internal Microsoft testing during May, a result the company has not benchmarked publicly against named competitors.

On the agent side, Microsoft moved its Agent 365 software development kit to general availability and reorganized its knowledge layer around Foundry IQ, which is generally available and unifies Work IQ, Fabric IQ, Azure SQL, file search and external sources, with Web IQ added for live web grounding. A new GitHub Copilot desktop app pushes Copilot past chat into managing tasks and pull requests, and Visual Studio is moving onto the GitHub Copilot software foundation underneath. Microsoft also showed MDASH, a multi-model scanning system in expanded private preview that pairs Defender with GitHub to find and remediate vulnerabilities, alongside Windows containers that isolate agents under policy. On hardware, the Surface RTX Spark Dev Box, built with Nvidia, delivers roughly 1 petaflop of local AI compute, and a concept device called Project Solara imagines machines that run agents in place of applications.

The most distant bet was Majorana 2, Microsoft's next quantum chip. The company claims an average qubit lifetime of 20 seconds, reliability 1,000 times higher than its previous generation and a path toward 1 million qubits on a chip that fits in a palm, with a scalable quantum machine targeted for 2029.

Why The Independence Push Lands Now

Microsoft's reliance on OpenAI has defined its AI strategy since 2023, when it built Copilot on GPT models and committed billions to the partnership. Owning models, and co-designing them with its own Maia and Cobalt chips, gives Microsoft room to negotiate on cost and to set its own roadmap rather than wait on a partner. That positioning moves it closer to Google, which pairs Gemini with its custom tensor processing units, and Amazon, which pairs Nova models with Trainium silicon. Both rivals have argued for years that owning the full stack lowers cost and tightens integration, and Microsoft is now making that case with its own parts.

The partnerships on stage were a reminder that independence has limits. Satya Nadella appeared with Nvidia's Jensen Huang and Qualcomm's Cristiano Amon, because Microsoft still depends on Nvidia for training compute and on chip partners for the devices that run its agents. Self-sufficiency in models does not extend to the silicon that trains them at frontier scale.

What Microsoft Still Has To Prove

The benchmarks are the obvious gap. Every performance figure Microsoft shared came from its own evaluations, and much of what it showed is not yet generally available. MAI-Thinking-1 is in private preview, the Cobalt 200 virtual machines are in preview, MDASH is in expanded private preview and Project Solara is a concept device, so buyers cannot test most of these claims against their own workloads today. The Frontier Tuning result rests on a single internal example, which is encouraging but far from a pattern across industries.

The quantum timeline is years out, and quantum roadmaps across the industry have a long history of slipping. MAI also does not replace OpenAI or Anthropic inside Copilot today, where those models still handle most production traffic, so the practical dependency remains even as the strategic one loosens. Agent governance is early as well. The Windows isolation containers and a proposed agent control specification are sensible, though they are version-one answers to security questions that enterprises are only starting to ask.

What This Means For Buyers

None of this calls for ripping anything out. The right read is optionality. Microsoft is giving customers a path to run first-party models where cost or data residency matters, while keeping OpenAI and Anthropic available for the work those models do best. Watch Frontier Tuning closely if your organization runs high volumes of repetitive, well-defined tasks, because a model tuned inside your compliance boundary at lower cost is a real budget lever once the benchmarks hold up outside Microsoft's labs.

The harder work is governance. Agents that act across Windows, Azure and GitHub need identity, policy and audit controls in place before they scale, not after. Decision makers who treated Build as a model launch missed the larger move. Microsoft is assembling the pieces to make its own intelligence the default inside its own platform, and the leverage that creates over pricing and roadmap will outlast any single benchmark.