Microsoft took the wraps off Project Solara at Build 2026, an ambitious chip-to-cloud platform engineered specifically for a new generation of agent-first Windows devices. The initiative, developed in close collaboration with silicon giants Qualcomm and MediaTek, marks a fundamental shift in how the company envisions the interplay between hardware, operating system, and artificial intelligence. Instead of grafting AI onto existing paradigms, Solara bakes autonomous agent capabilities into the very silicon and stretches their reach seamlessly into the cloud.

Satya Nadella introduced the concept during the opening keynote, describing it as “the culmination of over three years of rethinking the PC from the transistor up.” The demonstration showed an agent negotiating a complex travel itinerary, pulling live pricing from airline APIs, coordinating with a calendar, and proactively adjusting a schedule—all without a browser open. That experience, Nadella said, wasn’t just a cloud demo; it was running locally on a Solara reference device powered by a next-generation Qualcomm Snapdragon NPU.

From Copilot to autonomous agents

Windows has been on an AI journey for years, with features like Cortana, the AI-powered Bing integration, and the Copilot sidebar gradually weaving machine intelligence into the user experience. But Microsoft’s internal telemetry revealed a persistent friction: users were still spending too much time navigating between applications, copying data, and manually orchestrating tasks. Copilot could answer questions or summarize documents, but it couldn’t take action across multiple silos without extensive user prompting.

Project Solara aims to close that gap by defining a new class of “agent-native” hardware. Instead of running a traditional OS with an AI assistant bolted on, Solara devices treat the agent as the primary interaction model. A dedicated neural processing unit (NPU) manages always-on listening, visual context sensing, and intent understanding, while a tightly integrated cloud fabric handles heavier reasoning, large-scale data retrieval, and multi-step planning. The goal is a device that anticipates needs and executes tasks proactively, such as ordering supplies when inventory runs low, rescheduling meetings around travel delays, or even composing reports from raw data pulled from multiple cloud services.

What chip-to-cloud really means

The “chip-to-cloud” moniker isn’t marketing fluff. It describes a full-stack architecture where security boundaries, AI model sharding, and data privacy controls are coordinated from the silicon up to Azure servers. At the hardware level, Solara relies on a dedicated security processor—Microsoft calls it the Pluton AI Enclave—that isolates agent-specific operations from the rest of the system. This ensures that sensitive data, like biometric authentication or confidential documents, never leave the device without explicit user consent and strong encryption.

From there, the NPU pipeline handles real-time inferencing for lightweight models that handle natural language understanding, screen context analysis, and basic task routing. When a request exceeds the local compute budget—say, analyzing a 50-page PDF or cross-referencing live market data—the request is broken into encrypted micro-tasks that are scattered across Azure’s edge nodes. Each micro-task only processes anonymized data vectors, making it impossible for any single cloud node to reconstruct the user’s original information. Microsoft claims this approach meets the strictest data sovereignty requirements, which will be critical for enterprise adoption.

The silicon partnerships: Qualcomm and MediaTek

Building agent-first devices required a ground-up reimagining of the system-on-chip (SoC). Both Qualcomm and MediaTek have built custom NPUs that exceed 40 TOPS (trillions of operations per second), specifically tuned for the Mixture of Agents (MoA) framework Microsoft developed for Solara. Unlike a single monolithic AI model, MoA decomposes tasks into specialized sub-agents—one for web navigation, one for spreadsheet formulas, one for image generation—and schedules them intelligently across local and cloud resources.

Qualcomm’s Snapdragon X Elite Gen 2 will be the launch platform, with engineering samples already being tested by select OEMs. The chip features a dedicated Solara IP block that handles agent state caching, context window compression, and seamless cloud handoff. MediaTek, meanwhile, is targeting thin-and-light commercial laptops and education devices with its Kompanio 1000T series, which takes a slightly different approach by offloading more agent orchestration to a low-power microcontroller that runs even when the main CPU is asleep. This allows a Solara device to remain “agent-alive” for up to two weeks on standby, continuously monitoring calendars, emails, and system notifications.

Both partner chips integrate a hardware root of trust that validates every agent action against a user-defined policy engine. For instance, an IT administrator could restrict agents from accessing HR files or making purchases above a certain threshold. Those policies are enforced in silicon, making them resistant to tampering even if the OS is compromised.

Windows 2026 and the agent-first UX

Solara isn’t just a hardware spec; it demands a rearchitected OS. The version of Windows shipping alongside these devices—tentatively called Windows 2026—introduces the Agent Shell, a complete replacement for the traditional desktop metaphor. Instead of a Start menu and icons, users interact with a persistent agent canvas that surfaces relevant tools, files, and suggestions based on current context. Microsoft demonstrated switching from an email about a budget review to a pre-populated spreadsheet with real-time Power BI data, all prompted by a simple “help me prepare for this.”

Under the hood, the OS uses a new kernel scheduler that treats NPU time as a first-class resource alongside CPU and GPU time. Real-time agent tasks get priority, ensuring that latency-sensitive operations like voice dictation or camera-based gesture recognition never stutter. The compositing system has also been overhauled to support agent-driven window management; windows can be resized, arranged, or even created on the fly by an agent without breaking the user’s mental model of the workspace.

Legacy Win32 applications won’t be left behind. Microsoft has developed a compatibility layer called Agent Bridge that lets traditional programs expose their core functions to agents through a declarative manifest. If an application lacks a manifest, the agent can still interact with it using computer vision and accessibility APIs, but Microsoft is actively encouraging developers to adopt the new Agent API set for a smoother experience. Early partners include Adobe, SAP, and Salesforce, who all demonstrated agent-enhanced workflows during the Build keynote.

Enterprise security and manageability

For IT departments, the prospect of autonomous agents roaming across corporate data might raise alarms. Microsoft’s answer is the Solara Trust Framework, a comprehensive set of controls rooted in the Pluton AI Enclave and extended through Microsoft Intune. Agents operate within a least-privilege model, where each action—reading a file, sending an email, accessing a web API—must be pre-authorized by policy. All agent activity is logged to a tamper-proof audit ledger that feeds into Sentinel for anomaly detection.

During the demo, a simulated agent attempted to access a confidential M&A document. The system immediately blocked the action, notified the user with a prominent visual alert, and recorded the attempt for security operations. Administrators can define fine-grained policies based on user role, data classification, and even time of day. Solara also introduces “agent firewalls” that inspect the micro-tasks flowing between local NPUs and the cloud, ensuring no sensitive data leaks in an unencrypted form.

Microsoft is banking on these controls to convince regulated industries—healthcare, finance, government—that agent-first devices can meet compliance standards like HIPAA, GDPR, and FedRAMP. The company has already completed SOC 2 Type II audits for the Solara cloud components and is working with several European data protection authorities to certify the chip-to-cloud architecture under GDPR’s data minimization principles.

Developer opportunities and tools

Solara isn’t just a consumer play; it’s a new platform for developers. At Build, Microsoft released the Solara Agent SDK, which includes Visual Studio templates, a local simulator, and integration with GitHub Copilot. Developers can define agent skills using a declarative YAML syntax, then compile them into containerized modules that can run on-device or in the cloud. The SDK also supports multi-agent coordination, allowing developers to compose complex workflows from smaller, reusable agent components.

One of the more compelling demos came from a startup called Allium Health, which built a clinical agent that monitors patient vitals from wearable devices, cross-references them with electronic health records, and drafts pre-visit summaries for doctors. The entire stack was built in under two weeks using the Solara SDK, and it runs latency-sensitive anomaly detection locally on the patient’s device, preserving privacy while still alerting care teams within seconds of a critical event.

Microsoft is also launching the Solara Ready certification program, which tests devices for agent performance, battery life, and security compliance. OEMs like Dell, HP, and Lenovo have already pledged to deliver Solara Ready models in time for the Windows 2026 launch, with pre-orders expected to open in early Q4 2026. Pricing hasn’t been finalized, but executives hinted that enterprise configurations will start around $1,200, targeting the premium segment where most pilot programs are expected.

The road ahead

Project Solara is undeniably ambitious, and its success hinges on execution across a complex supply chain. The chip shortages of 2025 taught Microsoft the perils of relying on a single silicon partner, which is why the dual-source strategy with Qualcomm and MediaTek is so critical. But delivering a consistent agent experience across two distinct NPU architectures, each with their own cloud partitioning schemes, will be a herculean engineering effort.

Furthermore, user acceptance remains an open question. While early focus groups responded positively to the idea of a proactive agent that handles mundane chores, many expressed unease about ceding control to an AI that can act without explicit commands. Microsoft’s Trust Framework is designed to address those fears, but trust is earned slowly and lost quickly. A single high-profile agent error—accidentally sending a sensitive email or purchasing an item without confirmation—could set the project back by years.

Nevertheless, the vision is compelling. By building the agent into the silicon and vaulting it to the cloud under zero-trust principles, Microsoft is attempting to leapfrog the current AI assistant market and define a new computing category. If Solara ships on time and delivers on its promises, the Windows landscape five years from now may look nothing like it does today. The model of staring at a grid of icons and clicking through menus may finally give way to an ambient computing fabric where the device understands intent and acts on it, securely and seamlessly. That transition started on a Build stage in 2026, and the industry is now watching to see if Microsoft can pull it off.