AMD Shows How to Run AI Agents on PCs with 128 GB of Memory

AMD has published a practical guide for running OpenClaw, an open-source AI agent, in a Windows environment on consumer PCs, aimed at configurations with 128 GB of RAM. The guide describes two ready-made AMD hardware builds — RyzenClaw and RadeonClaw — on which users can deploy the agent locally without relying on cloud infrastructure.

What are the RyzenClaw and RadeonClaw configurations

Based on their names, both configurations are built around current AMD processor and graphics card lines — Ryzen and Radeon respectively — and offer two alternative paths to local agent deployment: one emphasizing processor compute and platform-unified memory, the other focusing on a discrete graphics card as the main inference accelerator. This choice of two builds allows AMD to demonstrate that running modern AI agents does not necessarily require dedicated server infrastructure with Nvidia accelerators — the task can be solved with top-tier consumer hardware assembled from mass-market components.

The fact that the manufacturer recommends such an amount of RAM for full-fledged agent scenarios reflects a broader trend: local AI agents are becoming a new level of hardware requirement, comparable in workload to professional workstations rather than ordinary home PCs. For buyers, this is another argument when choosing a computer configuration — alongside gaming performance or video memory size, the question of how much memory is needed to locally run a full-fledged AI agent rather than a stripped-down demo version increasingly comes up.

• Tool — OpenClaw, an open-source AI agent.
• Operating system — Windows.
• AMD configurations — RyzenClaw and RadeonClaw.
• Memory guideline — approximately 128 GB of RAM.
• Publication format — official guide from AMD.

Why AI agents need so much memory

Running a local AI agent differs from a simple local chatbot in that the agent, in addition to the language model itself, keeps in memory the history of reasoning, results of tool calls, intermediate files, and often — multiple parallel task chains simultaneously. The larger the local model and the longer the context window it has to process, the more RAM is required to fit everything without constant disk access, which severely slows down performance. The 128 GB guideline indicates that AMD is targeting users who want to run reasonably large models locally, not lightweight versions stripped down for minimal configurations.

What this means for local AI agents on PCs

Publication of such a guide by a major processor and graphics card manufacturer is part of a broader movement toward local, private AI agents that operate on the user's own hardware without sending data to the cloud. For enthusiasts and small teams, this reduces both subscription and API call costs as well as risks associated with sending sensitive data to third-party servers. At the same time, this benefits AMD itself — it positions its consumer Ryzen and Radeon platforms as a real alternative to the Nvidia-plus-cloud combination for those who want to keep AI agents at home or in the office on their own hardware.

Beyond saving on subscriptions, local agent deployment also eliminates the question of latency — calling a cloud API always adds network latency and dependence on internet connection stability, whereas a locally deployed agent responds as fast as the user's own hardware permits. For developers and enthusiasts, this also means the ability to experiment with model settings and agent parameters without the limitations typically imposed by cloud providers.

Also notable is that such a guide is released by the component manufacturer rather than the agent developer: for AMD, this is a way to demonstrate that its processors and graphics cards can handle not just games but full-fledged agent AI workloads, which until recently were associated almost exclusively with professional accelerators and cloud services. This positioning is especially important against the backdrop of growing user interest in running AI tools offline, without a subscription, and without the risk that a cloud service might one day become paid, change its terms of use, or restrict access to some features.