Unmetered Intelligence
How Windows Is Building the Operating System for the Agentic Era, Designed Around How People and Enterprises Use AI
-
Ryan Shrout
A forward-looking view of what an operating system must deliver as agentic workflows take hold, and where Windows stands today, grounded in what Microsoft announced at Build 2026.
The operating system is being rebuilt for how people and enterprises will work in the agentic era, and Windows is leading that shift with capability that is shipping or rolling out today. The throughline is unmetered intelligence, running AI locally and escalating to the cloud when the work calls for it, which changes the cost, privacy, and responsiveness of the AI experiences people rely on. This report defines the capabilities an operating system will need to deliver as agentic workflows take hold, shows where Windows stands against each one based on what Microsoft announced at Build 2026, and describes how that value scales across the device portfolio, from mainstream Copilot+ PCs to a deskside AI supercomputer. Throughout, we use the term local models for models that run on the device itself, a category also described as on-device models.
- The shift is human-first. The design goal is not AI for its own sake. It is to give people and organizations optimized AI experiences and to make them ready for the agentic era, with the operating system evolving to serve how humans work.
- Unmetered intelligence is the value. Running the right work locally, without per-token cost or a cloud round trip is where newer, more capable hardware paired with Windows creates value a cloud-only model cannot match, especially as always-on agents make continuous inference the norm.
- Governance is the unlock. Windows lives up to the enterprise promise by letting organizations observe, govern, and secure agents at the operating-system layer, so agent adoption can scale with control and trust intact.
- It scales across every tier. The same Windows AI platform runs from the broad Copilot+ PC install base to high-performance and frontier systems, so the value reaches mainstream users and frontier developers alike.
Key Highlights:
- Enterprise model API spend more than doubled to about $8.4 billion by mid-2025, even as token prices fell
- Local models on Windows, including Aion 1.0 Instruct and Plan, run agentic reasoning with no per-token cost
- Routing work by complexity across local and cloud can cut inference cost by roughly 5 to 10 times
- OS-enforced identity and containment let enterprises observe, govern, and secure every agent
Research commissioned by:
