Microsoft has scrapped the Copilot+ PC exclusivity that once defined its on-device AI ambitions, announcing at Build 2026 in San Francisco that GPUs—not just NPUs—will drive local intelligence on Windows going forward. The shift upends a two-year narrative that tethered advanced AI features to premium Arm hardware and later to newer x86 chips with dedicated neural processing units, instead opening the door for hundreds of millions of existing PCs with capable discrete or integrated graphics.

Speaking to a packed Moscone Center, Windows and Surface chief Pavan Davuluri confirmed that the next iteration of the Windows AI platform will target any DirectX 12-capable GPU with at least 4GB of VRAM, alongside a refreshed set of agentic developer tools. “We built an incredible foundation with Copilot+ PCs, but the ecosystem vote was loud,” Davuluri said during the keynote. “Developers told us they need GPU compute for the models they're building, and enterprises told us they can't rip and replace every desktop. So we're meeting them where they are.”

The announcement effectively retires the bright-line Copilot+ PC branding as a prerequisite for on-device AI experiences, replacing it with a tiered capability model that scales from lightweight tasks on NPU to heavier inference and fine-tuning on GPU.

The Copilot+ Detour

When Microsoft launched Copilot+ PCs in mid-2024, the message was binary: a Windows PC needed a Qualcomm Snapdragon X Elite or equivalent Arm chip with a 40-trillion-operations-per-second (TOPS) NPU to run features like Recall, Cocreator, and Live Captions locally. The company argued that only a dedicated AI engine could deliver the privacy, latency, and battery life users expected. Retailers stamped “Copilot+ PC” on laptop boxes and Microsoft predicted 50 million units sold by the end of 2025.

Reality fell short. Adoption lagged as businesses balked at refreshing fleets for features many considered novelties. The x86 camp—Intel with Core Ultra and AMD with Ryzen AI 300—eventually met the 40 TOPS bar, but by then the conversation had shifted. Developers complained that NPUs were fragmented, each with its own toolchain and quantized model zoo, while the GPU ecosystem was mature, unified, and dramatically more powerful. Even Microsoft's own AI teams, insiders admit, did most of their heavy lifting on NVIDIA and AMD GPUs in Azure before porting down to an NPU.

GPU Ascendant

At Build 2026, Microsoft laid out a new vision it calls “Windows AI Everywhere,” which leans hard into the GPUs already sitting in enterprise workstations, gaming rigs, and even high-end thin-and-lights. The centerpiece is a revamped DirectML runtime that can schedule AI workloads across any DX12-capable GPU, integrated or discrete, without requiring explicit code changes. Partner demos showed Stable Diffusion XL, Llama 4, and Phi-4-mini running locally on everything from a Surface Laptop with Intel Arc graphics to a desktop powered by an NVIDIA GeForce RTX 5060.

The technical underbelly: a new GPU Work Scheduler in the Windows display driver model that can preempt rendering jobs for AI inference slices as short as 500 microseconds, plus a memory-sharing API that lets GPU buffers and system RAM be treated as a unified pool for models up to 8 billion parameters. These are low-level changes seeded through an Insider build scheduled for June 2026, with general availability alongside the Windows 11 24H2 Moment 8 update in the fall.

Critically, the toolchain aligns with what developers already use. Visual Studio 2026 gets a built-in AI Profiler that visualizes GPU occupancy, memory pressure, and token throughput. The ONNX Runtime is updated to natively target DirectML, while a new “Agent Runtime” built on top of Semantic Kernel lets ISVs define multi-step reasoning agents that run locally without round-tripping to the cloud.

Agentic Tools and Enterprise Play

Microsoft’s pivot isn't just about hardware—it's about agency. The keynote hammered the theme of “agentic AI on the edge,” where small language models perform iterative tasks, chain tools, and access local data with full privacy. Davuluri demonstrated a procurement agent that could read email attachments, cross-reference a local SQLite database, update a Teams tab, and generate a draft contract—all on a Dell laptop with an NVIDIA RTX 2000 Ada GPU, offline.

For IT departments, the value proposition sharpens. Instead of purchasing Copilot+ certified machines at a premium, organizations can now deploy the same AI features to their existing Windows 11 fleet, provided the GPU meets the baseline. Microsoft released a Windows AI Readiness Analyzer that scans through Intune and delivers a dashboard of eligible devices, estimated inference latency, and annual GPU-cycle cost—modeled on Azure spot pricing but for on-prem hardware. Early enterprise adopters at the show, including Accenture and Siemens, reported that 60–70% of their deployed Windows PCs already meet the 4GB VRAM floor, largely thanks to Intel Iris Xe and mainstream AMD Radeon integrated graphics.

Security also gets a boost. The local agent runtime runs inside a hardened container with cryptographically signed manifest files, preventing model tampering. All local data access is governed by Windows Information Protection policies, meaning corporate documents never leave the device during agent workflows.

Not Abandoning NPUs—Yet

Jimmy Chen, CVP of Windows Silicon and Systems Integration, walked a careful line: NPUs aren't disappearing. They still excel at sustained, low-power workloads like real-time transcription, gaze tracking, and background noise suppression. Microsoft will continue to expose NPUs through the DirectML abstraction for ISVs that want to optimize for battery life. But Chen admitted that “the NPU development ecosystem isn't where we hoped it would be after two years” and that Microsoft would stop marketing an NPU TOPS score as a purchasing bright line.

This is already evident in Microsoft’s own hardware. The Surface Pro 11 refreshed for 2026—shown briefly on stage—pairs a Qualcomm Snapdragon X2 Elite with an AMD RDNA 4–based integrated GPU, and Microsoft’s demos deliberately ran the meaty AI workloads on the GPU tile rather than the Hexagon NPU. When asked if future Surface devices might omit an NPU entirely, Chen said it would depend on the form factor and use case but that “GPUs bring the flexibility we need for the next wave of agentic experiences.”

Developer Reception

In the expo hall, reaction was broadly positive but not without skepticism. “Finally, I can target one hardware path and know it’ll run on any decent laptop, gaming PC, or workstation,” said a developer from a major CAD vendor. But others worried about fragmentation creeping back in through GPU capability tiers. A session on “Designing for VRAM Budgets” outlined four compliance tiers: Light (<2GB), Standard (4GB), Pro (8GB), and Ultimate (12GB+), with each unlocking progressively larger models and longer context windows. Microsoft promised that the Windows Store and winget will surface AI-capability badges, but independent devs fear user confusion.

Still, the unified stack won over the crowd. “I’d rather manage five VRAM tiers than five different NPU toolchains,” said an AI engineer from a healthcare ISV. “This lets us deploy the same model to a nurse’s tablet and a radiologist’s workstation.”

Competitive Context

Microsoft’s GPU-first strategy brings Windows closer to the Apple model, where the M-series Neural Engine is present but often bypassed in favor of the GPU for heavy ML tasks via Core ML and Metal Performance Shaders. Google has taken a similar approach with ChromeOS, using GPU inference for Gemini Nano features on Intel and AMD Chromebooks.

The difference is scale: by tapping DirectX, Microsoft instantly addresses a hardware base of over 1.5 billion active Windows devices, roughly 40% of which contain a DX12-capable GPU with at least 4GB RAM, according to Canalys estimates cited during the keynote. No other OS platform can claim a local AI runtime that spans such a heterogeneous fleet.

Roadmap and the Next Build

Looking ahead, Microsoft teased a Windows AI Roadmap that extends through 2027. Near-term deliverables include the general availability of the Agent Runtime SDK in Q4 2026, a public model catalog hosted on GitHub with pre-optimized GGUF and ONNX formats for DirectML, and a “bring your own model” fine-tuning toolkit that runs locally on NVIDIA and AMD GPUs using QLoRA.

Longer term, the company is experimenting with disaggregated compute: splitting inference across GPU and NPU simultaneously, with a scheduler in the Windows kernel that allocates tokens based on available thermal headroom and battery state. A proof-of-concept shown in a breakout session ran a 7B-parameter model with the attention layers on an NPU and feedforward layers on an iGPU, achieving a 35% energy reduction over GPU-only execution for bursty chat workloads.

Build 2026 will be remembered as the moment Microsoft reset its AI hardware story. By releasing Copilot+ from its narrow silicon mandate, the company has traded marketing simplicity for developer pragmatism and enterprise reach. Whether that trade-off leads to a Cambrian explosion of local AI apps on Windows depends on execution, but for now, the GPU door is wide open.