- Microsoft’s Build 2026 partnership with NVIDIA puts RTX Spark chips into Windows PCs and Azure infrastructure, enabling local execution of 120-billion-parameter models on a single device with 128GB of unified memory.
- The Majorana 2 quantum chip claims a 1,000-fold improvement in qubit reliability over its predecessor, with mean qubit lifetimes of 20 seconds, up from milliseconds on Majorana 1, pushing Microsoft’s commercial quantum target from around 2035 to 2029.
- Microsoft’s 2025 Work Trend Index found that 81% of business leaders expect AI agents to be extensively integrated into their strategies within a year, a number Microsoft cited at Build to justify how aggressively it is packaging agentic tooling across Azure and Windows.
Microsoft’s Build 2026 brought two announcements that don’t often share a stage: a laptop-sized device capable of running a 120-billion-parameter model locally, and a quantum chip that has pushed the company’s commercial viability target from 2035 to 2029. The RTX Spark chip, built with NVIDIA, drives the first. Majorana 2 is the second, and it is drawing as much scepticism from external researchers as excitement from Microsoft.
RTX Spark and the local AI push
The RTX Spark is a system-on-chip pairing a Blackwell RTX GPU with a 20-core Grace CPU. NVIDIA says the combination delivers up to one petaflop of AI compute with 128GB of unified memory, enough to run 120-billion-parameter models locally with a million-token context window, no cloud required. For enterprise deployments, local execution means lower latency and data that never leaves the device.
The reference hardware is the Surface RTX Spark Dev Box, a compact machine built for AI development. Alongside it, Microsoft introduced Project Solara, a reference design platform for what the company calls “agent-first” hardware, intended as a standard architecture that device makers can build against across form factors rather than a single product line.
On the cloud side, the partnership adds new Azure Cobalt 200 VMs targeting enterprise workloads that run too large for local execution. The two-tier approach gives developers a real choice about where inference runs, based on sensitivity and latency requirements, rather than defaulting everything to cloud. That kind of architecture is increasingly relevant as Windows AI agents gain deeper access to local systems.
MAI models and the Scout agent
Microsoft debuted seven proprietary “MAI” models at Build, developed entirely in-house without OpenAI involvement, spanning reasoning (MAI-Thinking-1, a 35-billion-parameter model benchmarked against Claude Opus 4.6 and Sonnet 4.6), coding (MAI-Code-1), image generation (MAI-Image-2.5), transcription (MAI-Transcribe-1.5) and voice (MAI-Voice-2).
Scout, an autonomous agent built into Microsoft 365, was more concrete. CEO Satya Nadella described it as an “Autopilot” for enterprise tasks in Microsoft’s Build presentation, capable of scanning communications and surfacing critical items without being prompted. Whether Scout’s autonomous operation stays within boundaries enterprise IT teams will accept is something deployments will answer. Microsoft cited its 2025 Work Trend Index at the conference, which found that 81% of business leaders expect AI agents to be extensively integrated into their strategies within the coming year.
Majorana 2’s qubit leap
Majorana 2 was the most surprising Build announcement. Microsoft claims the topological quantum chip achieves a 1,000-fold improvement in qubit reliability over Majorana 1. Where Majorana 1 qubits lasted milliseconds, Majorana 2 qubits have a reported mean lifetime of 20 seconds, with some sustaining for up to a minute. The Microsoft Quantum team credits a materials stack that incorporates lead as the specific engineering change that produced that stability.
Microsoft says Majorana 2 was developed with assistance from Microsoft Discovery, an AI platform for research and development that reached general availability at Build 2026. The company credits agentic AI within Discovery for accelerating the materials research behind the chip, and points to that acceleration as what moved its commercial quantum target from around 2035 to 2029. External researchers have pushed back, asking for more reproducible data before accepting Microsoft’s more assertive claims. The company says progress is measurable; how the research community receives that evidence will matter more than the conference announcement. For context on how custom silicon is reshaping the broader AI hardware race, the OpenAI and Broadcom Jalapeño chip announcement covers similar ground from a different angle.
