4x AMD R9700 local LLM rig for 10-person startup costs $18K

A user shared a local inference build for a 10-person startup costing $18K, with a Threadripper 9970X, 128GB DDR5 ECC, and 4 AMD Radeon R9700 AI Pro GPUs (128GB total VRAM). Serving Qwen 3.8 and other models via a Radiance fork, it reached about 6.3-6.8K tok/s aggregate prefill and 900 tok/s decode at 16 concurrent sessions.

• LemonSeed: Qwen3.8-27B Hits 159 tok/s on iPad via eGPU R97002026-10-09
• Trump unveils $2.4B in tech giant AI pledges2026-10-09
• AMD Unveils Full-Stack Agentic AI Platform2026-10-08
• ROCm 10.1 Targets Data-Movement Bottleneck2026-10-08
• Saluki 27B claims 96% of Qwen 3.8 performance at 1/7 size2026-10-08