AMD Sees Agent Computers as the Next Step in AI PCs

Key Takeaways

• AMD offers RyzenClaw and RadeonClaw for local AI agents
• RyzenClaw runs Qwen 3.5 35B at 45 tokens/sec
• RadeonClaw achieves 120 tokens/sec with 32 GB GPU
• RyzenClaw system starts around $2,700; GPU $1,299
• Supports up to six concurrent agents on CPU path

Summary

AMD released a guide outlining two local‑AI hardware paths—RyzenClaw and RadeonClaw—designed to run large language models entirely on Windows without cloud services. The RyzenClaw configuration uses a Ryzen AI Max+ system with 128 GB memory, delivering about 45 tokens per second and supporting up to six concurrent agents. RadeonClaw pairs a Radeon AI PRO R9700 GPU, reaching roughly 120 tokens per second but limiting concurrency to two agents. Pricing starts at $2,700 for the CPU‑centric build and $1,299 for the GPU card, positioning the solutions for early adopters.

Pulse Analysis

Artificial‑intelligence agents have traditionally relied on remote data‑centers, but growing concerns about data privacy, latency, and recurring cloud fees are prompting a move toward edge‑compute solutions. AMD’s newly released “Agent Computer” guide embodies this shift by showing developers how to run large language models entirely on a Windows machine using its own silicon. By leveraging WSL2 and the open‑source LM Studio front‑end, the guide eliminates the need for external APIs, giving users full control over model weights and inference costs. This approach aligns with a broader industry trend that values on‑premise AI for both individuals and enterprises.

The guide defines two hardware tracks: RyzenClaw, built around the Ryzen AI Max+ platform with 128 GB of unified memory, and RadeonClaw, which pairs the Radeon AI PRO R9700 GPU with 32 GB of VRAM. Running the 35‑billion‑parameter Qwen 3.5 model, RyzenClaw delivers roughly 45 tokens per second and can handle a 260 K‑token context window while supporting up to six simultaneous agents. RadeonClaw triples the throughput to about 120 tokens per second but reduces the context window to 190 K tokens and limits concurrency to two agents. These trade‑offs let users choose between CPU‑centric flexibility and GPU‑driven speed.

Pricing remains the primary obstacle to mainstream adoption. A RyzenClaw‑style workstation, comparable to a Framework Desktop with a Ryzen AI Max+ 395, starts near $2,700, while the Radeon AI PRO R9700 GPU alone costs about $1,300. For hobbyists and small businesses, these figures place local AI agents firmly in the early‑adopter segment. Nevertheless, the ability to run multiple agents without recurring cloud fees could appeal to regulated industries such as finance, healthcare, and defense, where data sovereignty is non‑negotiable. As silicon costs fall and software ecosystems mature, AMD’s agent‑computer strategy may accelerate the shift toward decentralized, always‑on AI workloads.