Microsoft Build 2026 Highlights? 7 Self-Developed AI Accelerate Autonomy, Quantum Chips See 1,000-Fold Leap

TradingKey - On Tuesday, June 2, Eastern Time, Microsoft ( MSFT )'s annual developer conference, Build, kicked off at the Fort Mason Center in San Francisco. This is the first time the conference has moved away from its home base in Seattle since 2016. While the scale was streamlined to about 2,500 people, the signal it sent was disruptive—Microsoft is attempting to redefine Windows from an operating system for human users to a native runtime environment for AI agents, pushing AI from the stage of "assisting human work" toward a new phase of "executing tasks on behalf of humans."

At this high-profile conference, Microsoft released seven new AI models at once, covering core areas such as reasoning, coding, vision, and multimodality, signaling that its AI "autonomy" strategy has entered a critical implementation phase.

Mustafa Suleyman, head of Microsoft AI, stated clearly that Microsoft is forging a development path distinct from that of Google ( GOOGL ), Meta ( META ), and OpenAI.

He emphasized: "We are more focused on an Anthropic-style direction—the enterprise, developer, and coding markets." While continuing to deepen its partnership with OpenAI, Microsoft is accelerating the build-out of its own AI technology ecosystem, especially building competitive advantages in enterprise applications, developer tools, and programming scenarios.

Microsoft Launches 7 Self-Developed AI Models

Microsoft launched seven new AI models simultaneously, all integrated into the MAI (Microsoft AI) family. These models cover a full stack of capabilities, including reasoning, coding, vision, speech, and multimodality, marking a critical implementation phase of Microsoft's AI "autonomy" strategy.

Microsoft describes this series of models as core components of a "hill-climbing machine"—achieving iterative self-improvement through continuous investment in computing resources, optimized training data, and refined evaluation systems, ensuring users remain at the forefront of technology.

The seven models released this time do not merely pursue parameter scale; instead, they aim to build a complete capability stack of "thinking, reasoning, execution, and coding" to support next-generation AI agent systems.

Core products include the MAI Thinking series of reasoning models, ultra-efficient coding models, vision and multimodal models, lightweight models for agent systems, and specialized models optimized for enterprise and developer scenarios.

The two most watched products are MAI-Thinking-1, Microsoft's first flagship reasoning model, and MAI-Code-1-Flash, an ultra-efficient coding model specifically built for GitHub scenarios.

As a core weapon in Microsoft's strategy for the enterprise AI market, MAI-Code-1-Flash was trained using an end-to-end "clean and licensed" dataset and is currently being rolled out in batches to individual GitHub Copilot users in VS Code.

Users can either manually switch to the model via the model selection menu or have it intelligently assigned by the system's automatic selector based on task complexity.

In addition to MAI-Code-1-Flash, Microsoft's MAI-Thinking-1 reasoning model is equally noteworthy. The most significant announcement is Microsoft's first introduction of a reasoning model family—MAI Thinking.

Reasoning models are becoming the new battleground for AI competition in 2026. Unlike traditional chat models that focus on natural language interaction, reasoning models emphasize logical thinking—breaking down complex problems into executable steps, completing long-chain planning tasks, handling mathematical and code reasoning, and supporting the autonomous execution of agent systems. This capability perfectly aligns with the core requirements of enterprise applications, making it a strategic high ground for tech giants.

Microsoft's MAI-Thinking-1 is aimed directly at this market. According to official data, this mid-sized model's performance on key software engineering benchmarks is comparable to industry-leading models, specifically approaching the level of Claude Sonnet 4.6 in coding capabilities.

Mustafa Suleyman, head of Microsoft AI, admitted in a media interview that Anthropic still maintains a lead of several months, but emphasized that Microsoft is narrowing the gap at an astonishing pace, having achieved leapfrog progress over the past six months.

From a technical architecture perspective, the design philosophy of MAI-Thinking-1 is highly similar to Anthropic's Claude series—avoiding a blind pursuit of parameter scale in favor of focusing on practical reasoning capability and efficiency. This approach allows the model to handle complex tasks more effectively while lowering deployment costs, making it better suited for large-scale enterprise applications.

Next-generation quantum chip Majorana 2 unveiled

Simultaneously, Microsoft officially unveiled its next-generation quantum chip, Majorana 2. The chip is a successor to the Majorana project that sparked industry controversy last year and serves as the latest milestone in Microsoft’s 20-year commitment to the "topological qubit" roadmap. Departing from the superconducting quantum path pursued by giants like Google and IBM, Microsoft has opted for a more challenging technical route: utilizing Majorana quasiparticles to construct naturally noise-resistant qubits.

From a technical specification standpoint, the improvements in Majorana 2 are revolutionary. The number of qubits on the chip has increased from eight in the previous generation to 12, but the genuine breakthrough lies in qubit stability. According to data disclosed by Microsoft, the average qubit lifespan of the new chip exceeds 20 seconds, with some reaching over a minute, whereas the first-generation product released last year lasted less than 12 milliseconds. This represents a more than 1,000-fold increase in reliability, an advancement Microsoft compares to "replacing a phone battery that lasts one day with one that lasts nearly three years."

Majorana 2 moved away from the aluminum-based superconducting materials used in its predecessor, opting instead for lead-based superconductors and updating the semiconductor active region to a combination of indium arsenide and indium antimonide. This new material stack creates a more stable topological phase, significantly bolstering the qubits' resilience against environmental noise.

Chetan Nayak, a researcher and executive at Microsoft's quantum hardware division, stated that this progress gives the company the confidence to halve the R&D timeline for a practical quantum computer, moving the goal forward from 2035 to 2029.

It is noteworthy that the entire R&D process for Majorana 2 utilized AI-assisted design. Through the Microsoft Discovery agent, the research team accelerated material screening and architecture optimization, drastically shortening a development cycle that would typically take years.

Web IQ: Dedicated Search for AI Agents

Microsoft not only demonstrated breakthroughs in quantum computing but also launched a game-changing product in the field of AI agents—Web IQ.

Designed specifically for AI agents, this search API suite reconstructs its underlying architecture based on Bing's two decades of technical expertise. It aims to address the pain points of high search-related costs and slow response times in current AI applications, serving as the "information foundation" for the era of agents.

Unlike traditional search engines, Web IQ serves AI agents rather than human users. Jordi Ribas, President of Microsoft Search and AI, explained in an interview that while human searches require engines to rank and display results, AI agents need highly condensed, structured information snippets for rapid parsing and use without excessive token consumption. Consequently, Web IQ rebuilt its entire architecture from the ground up, leveraging Bing's twenty years of technical accumulation to provide "tailor-made" search services for AI agents.

According to official data disclosed by Microsoft, 95% of requests can be responded to within 165 milliseconds, averaging approximately 2.5 times the speed of industry competitors; through grounding technology, the returned information is more compact, reducing token consumption by 60% compared to traditional search APIs.

This performance metric is particularly crucial in the current AI application environment. A McKinsey report for the first quarter of 2026 indicates that search-related token consumption accounts for 35% of total costs in AI applications, and instances where response latency exceeds 300 milliseconds account for 40%, becoming a major bottleneck affecting the agent experience.

The core capability of Web IQ lies in its powerful grounding functionality. It helps AI agents obtain real-time, reliable internet information—including the latest news, real-time prices, dynamic inventory, web content, API documentation, and corporate information—thereby effectively reducing AI hallucination issues.

More importantly, Web IQ does not simply return web content; it provides executable information structures, enabling AI agents to directly call website services, automatically complete transactions, understand page semantics, operate online tools, and even collaborate with external agents. This design is highly consistent with Microsoft's previously promoted Model Context Protocol (MCP) strategy, signaling an evolution of the internet from "browsers reading webpages" to "AI agents reading services."

Moving Beyond Single Dependency: Microsoft Accelerates In-House AI Development

Over the past four years, Microsoft has staked nearly all its AI bets on its partnership with OpenAI. From Copilot to Azure AI, and from enterprise services to consumer applications, OpenAI's model technology has formed the core backbone of Microsoft's AI capabilities. However, as the partnership undergoes adjustments, Microsoft is moving toward a "truly self-sufficient" path of AI development.

The starting point of this transformation can be traced back to the renegotiation of the partnership agreement between the two parties last year. Although Microsoft still holds an approximately 27% stake in OpenAI and retains long-term access to advanced models, the company has internally begun explicitly building a multi-model strategy to "eliminate single-source dependency."

Mustafa Suleyman, Microsoft's head of AI, admitted in an interview that over-reliance on a single partner carries structural risks, and that possessing in-house R&D capabilities is the only way to ensure long-term strategic autonomy.

With the rapid rise of competitors such as Google, Meta, and Anthropic in the AI sector, Microsoft urgently needs to strengthen its technological moat. Specifically, Google's Gemini model series has surpassed OpenAI's GPT-4o in certain performance benchmarks, leading Microsoft to realize that excessive reliance on external technology could leave it at a disadvantage in future AI competition.