AMD Announces Agent Computers: RyzenClaw (128GB, 6 Concurrent Agents) and RadeonClaw (120 tok/s) Redefine Local AI Agent Hardware

AMD has published its vision for the Agent Computer — a local system designed to run AI agents directly on client hardware without cloud dependency. The company released official guidance for running OpenClaw locally on Windows through two hardware paths: RyzenClaw and RadeonClaw. Guru3d covered the announcement in detail on March 16, 2026.

AMD AGENT COMPUTER CONCEPT:

AMDs core argument is that not every AI workload belongs in a remote data center, especially when users want better privacy, fixed costs, always-on access, and direct control over models and data. The full stack runs through WSL2, with LM Studio and llama.cpp handling local LLM inference. Memory.md is supported through local embeddings, keeping the environment self-contained.

RYZENCLAW — MEMORY-OPTIMIZED PATH:

• Platform: Ryzen AI Max+ with 128 GB unified memory
• Recommended: Reserve 96 GB as variable graphics memory for AI workloads
• Model: Qwen 3.5 35B A3B
• Performance: ~45 tokens per second
• Prompt processing: 10,000 input tokens in ~19.5 seconds
• Context window: 260K tokens
• Concurrent agents: Up to 6 simultaneous agents
• Best for: Multi-agent workflows where memory footprint and concurrency matter more than raw speed
• Starting price: ~$2,700 for system

RADEONCLAW — SPEED-OPTIMIZED PATH:

• Platform: Radeon AI PRO R9700 workstation GPU with 32 GB VRAM
• Model: Qwen 3.5 35B A3B (same model)
• Performance: ~120 tokens per second (2.7x faster than RyzenClaw)
• Prompt processing: 10,000 input tokens in ~4.4 seconds (4.4x faster)
• Context window: 190K tokens
• Concurrent agents: Up to 2 simultaneous agents
• Best for: Single-agent or dual-agent workflows requiring fast response times
• GPU price: Starting at $1,299 (plus rest of system)

STRATEGIC SIGNIFICANCE:

1. HARDWARE BRANDING: AMD is the first chip manufacturer to create product categories specifically branded for AI agent workloads (RyzenClaw, RadeonClaw). This signals that AI agents are becoming a distinct computing workload category alongside gaming, content creation, and scientific computing.

2. LOCAL-FIRST AGENTS: The Agent Computer concept directly challenges the cloud-centric model for AI agent deployment. Privacy, fixed cost, and always-on access are the key value propositions.

3. MULTI-AGENT SUPPORT: RyzenClaws support for 6 concurrent agents suggests AMD is thinking ahead to orchestrated multi-agent workflows where specialized agents collaborate on complex tasks.

4. NVIDIA COMPETITION: This announcement lands on the same day as NVIDIAs GTC 2026 keynote, where NVIDIA is expected to unveil its own Vera Rubin and potentially Feynman architectures. AMD is positioning itself as the local/edge AI agent platform while NVIDIA dominates datacenter.

5. SETUP SIMPLICITY: AMDs claim that the full stack can be configured in under an hour signals the target is developers and enthusiasts, not just enterprise IT teams.

LIMITATIONS:

• Current pricing ($1,299-$2,700+) limits mainstream adoption
• Target audience remains developers, workstation users, and enthusiasts
• WSL2 dependency means Windows-only for now
• Model quality constrained to what runs locally (35B parameters vs 100B+ cloud models)