Mark Papermaster, AMD’s chief technology officer, delivered a blunt message to enterprise leaders on July 8, 2026: agentic AI is forcing a fundamental rethink of how we build computing infrastructure. Speaking at AMD’s RAISE Summit in a session captured by SiliconANGLE’s theCUBE, Papermaster argued that the shift from simple chatbots to autonomous AI agents that can plan, reason, and act across applications means enterprises can no longer optimize their silicon in isolation. Instead, they must design entire systems—from the smallest sensor to the largest data center cluster—to work as a unified, intelligent platform. For Windows users, that message has immediate, practical implications that stretch from the next Copilot+ PC on your desk to the server racks humming in your IT department.

The Big Pivot: From Chips to Systems

Papermaster painted a picture of an industry approaching a tipping point. For decades, performance gains came from cramming more transistors onto a single die. But agentic AI workloads—where an AI assistant might simultaneously parse a spreadsheet, search the web, generate a report, and schedule a meeting while learning from each step—don’t fit neatly inside a single accelerator. They sprawl across CPUs, GPUs, NPUs, and memory, often in real time. “You can’t just drop a faster GPU in a server and call it a day,” Papermaster said. “The orchestration layer, the memory hierarchy, the interconnects—everything has to be tuned for agents.” That means AMD is pushing its partners and customers to think about balanced system designs, where each component is chosen to eliminate bottlenecks that emerge only when AI agents begin to chain together complex tasks.

What Agentic AI Actually Means for Your Windows Machine

For everyday users, the terminology can seem abstract, but the consequences are concrete. Agentic AI is the technology behind the next generation of Windows Copilot features. Right now, Copilot can answer questions and generate text. Soon, it will be able to autonomously manage your inbox, draft documents based on meeting notes, and even debug code—all without step-by-step human prompting. That requires the AI to run complex models locally for privacy and low latency, while offloading heavier lifting to the cloud when needed. The processor inside your PC must handle a constant ebb and flow of inference tasks, not just a single burst.

Windows users with AMD-powered Copilot+ PCs—machines built around the Ryzen AI 300 series or later—will see the first real taste of this. These chips integrate a dedicated neural processing unit (NPU) capable of over 40 trillion operations per second (TOPS), meeting Microsoft’s threshold for advanced AI features. But Papermaster’s talk suggests that even 40 TOPS is just a starting point. As agents grow more capable, they’ll demand more concurrent compute. That’s where AMD’s system-level thinking pays off: a Ryzen AI processor shares a unified memory architecture with its integrated Radeon graphics, so AI tasks can spill over from the NPU to the GPU without copying data across a slow bus. The result is smoother multitasking when an agent is actively working in the background.

Power users—gamers, content creators, data scientists—face a different inflection. Many already run local AI models for Stable Diffusion, Whisper, or custom fine-tuned LLMs. With agentic AI, those models will need to run persistently, reacting to triggers like new emails or calendar changes. That steady-state load changes the calculus for hardware. A GPU that previously only fired up for rendering or gaming may now be under constant partial load. This makes thermal management and power delivery just as critical as peak TOPS. AMD’s emphasis on system optimization includes its SmartShift technology, which dynamically routes power between CPU and GPU, and could evolve to include the NPU in future Windows PCs.

The Enterprise Angle: Clusters and Software Stacks

For IT professionals, the message hits closer to the bottom line. Many organizations are already building internal AI agents to automate workflows, from HR ticket resolution to supply chain predictions. Papermaster stressed that these agents will not live on a single server; they’ll be distributed across on-premises clusters and cloud instances, often making decisions that cascade across dozens of nodes. AMD’s answer involves its ROCm software stack, which has quietly become more Windows-friendly over the past two years. ROCm 6.1, released in mid-2024, brought official support for Windows on Radeon GPUs, and subsequent updates have improved PyTorch integration. That means a company running Windows Server in its data center can now deploy AMD Instinct accelerators with the same open-source tooling they’d use on Linux, simplifying hybrid environments.

But Papermaster’s vision suggests that simply installing ROCm isn’t enough. He described a future where AI agents require “composable infrastructure”—pools of compute, memory, and storage that can be reconfigured on the fly based on agent demand. For Windows admins, this might translate into new management tools that let them allocate NPU slices across virtual machines, or policies that prioritize agent traffic on the network. Microsoft has been teasing similar capabilities with its Azure Local offering (formerly Azure Stack HCI), and a deeper collaboration with AMD seems likely.

How We Got Here: A Timeline of Acceleration

The RAISE Summit keynote didn’t happen in a vacuum. It’s the culmination of several years of accelerating convergence between Windows and AI hardware.

  • 2023: AMD launches the Ryzen 7040 series, its first mobile chips with a dedicated AI engine (the “Ryzen AI” block). Windows Studio Effects begin using the NPU for background blur and eye contact correction, but the capability is modest.
  • 2024: Microsoft introduces Copilot+ PCs, requiring an NPU with at least 40 TOPS. AMD answers with the Ryzen AI 300 series (Strix Point), packing a 50-TOPS NPU alongside Zen 5 CPU cores and RDNA 3.5 graphics. At Computex, AMD also releases ROCm 6.1 for Windows, giving developers a path to run machine learning frameworks directly on Radeon GPUs.
  • 2025: Microsoft rolls out Windows 11 24H2 updates that unlock new AI features like Recall (later reworked for security) and Click to Do, both of which lean heavily on the NPU. AMD ships the Ryzen AI Max series, pushing NPU performance to 55 TOPS.
  • 2026 (current): Agentic AI frameworks such as Microsoft’s Copilot Studio and open-source LangChain agents mature. At RAISE Summit, AMD makes the case that the next leap isn’t about a single piece of silicon but about co-engineering the entire platform.

What You Should Do Right Now

If you’re evaluating a new Windows PC or planning infrastructure refreshes, Papermaster’s insights offer a clear checklist.

For home users and prosumers:
- Prioritize a Copilot+ PC with an NPU of 40 TOPS or higher. Current AMD Ryzen AI 300 and Intel Core Ultra 200V systems meet the mark, but look for AMD’s just-announced Ryzen AI 400 series or later if you can wait until fall 2026. These next-gen chips are expected to double NPU throughput and add hardware acceleration for transformer attention mechanisms.
- Demand unified memory. Agentic AI thrives on low latency, and systems that split memory pools between CPU and GPU will struggle. AMD’s APU designs naturally unify memory, but check that the specific laptop or mini-PC you’re considering actually leverages that architecture fully.
- Keep Windows and drivers updated. Many NPU optimizations come through Microsoft’s Neural Processing Unit driver stack, which is delivered via Windows Update. The latest version as of July 2026 is 31.0.82.0; ensure you’re on at least that build.

For developers and data scientists:
- Install ROCm for Windows if you haven’t. The current HIP SDK (version 24.Q2 as of this writing) lets you compile and run CUDA-like code on AMD GPUs without dual-booting Linux. This will be essential as more agentic AI workloads remain local to a developer’s workstation.
- Experiment with ONNX Runtime and DirectML. Microsoft’s cross-platform inference engine is now deeply optimized for AMD NPUs and GPUs, and it includes built-in agent chaining templates that can split a workflow across local and cloud resources automatically.
- Start prototyping agent frameworks now. Tools like Microsoft’s Semantic Kernel and AutoGen have native Windows support and can be configured to prefer the NPU for certain tasks. Getting familiar with these will pay off as enterprise rollouts accelerate.

For IT administrators:
- Audit your server fleet for AMD Instinct MI300X or MI400 accelerators if you’re planning large-scale agent deployments on Windows Server. These GPUs support ROCm natively and integrate with Microsoft’s latest AI toolchains, including the Azure AI Foundry deployment pipelines.
- Investigate dynamic resource allocation via Windows Admin Center. Early previews of the “AI Resource Manager” extension (expected in Windows Server 2026) allow you to assign NPU/GPU time slices to specific VMs or containers—a capability that will become critical once agents start competing for compute.
- Budget for increased cooling and power. Always-on agents generate a continuous thermal load that differs from bursty rendering or database queries. Papermaster specifically mentioned that data center design must evolve to handle “sustained heterogeneous utilization,” a phrase that should prompt facility reviews.

Looking Ahead: The Platform Play

AMD’s RAISE Summit message wasn’t just a technology forecast—it was a competitive positioning statement. Intel is pushing its own AI PC platform with Lunar Lake and the forthcoming Nova Lake, while Qualcomm’s Snapdragon X Elite chips brought ARM-based Copilot+ PCs to market. Nvidia, for its part, dominates the cloud AI training space with CUDA. But AMD’s bet is that the physical proximity of CPU, GPU, and NPU on a single die, combined with an open software stack that now spans Windows and Linux, will prove decisive as agents move from the data center to the edge and onto the desktop.

For Windows users, the timeline is aggressive: Microsoft has indicated that many agentic features will require a Copilot+ PC by the end of 2026, and 40 TOPS may soon look like the bare minimum. Papermaster hinted that future NPUs will exceed 100 TOPS by 2027, which suggests a generational leap is coming. The prudent move, whether you’re buying a single laptop or provisioning a thousand servers, is to treat AI compute not as a discrete accelerator but as a foundational component of the system—just like memory or storage. AMD’s RAISE call to action is clear: start optimizing your entire stack today, or risk being overtaken by the agents of tomorrow.