Why this Local AI writer picked an AMD 395 / 128GB box for local LLMs
julianharris · x · 2026-09-20
Julian Harris, author of a Local AI newsletter, reveals he bought an AMD 395 / 128GB box, calling it the current price/performance sweet spot for local LLMs. He credits work by antirez and others on SSD streaming of model weights (loading weights from NVMe) for making high-RAM + NVMe setups viable.
- He'd like an NVMe Gen 5 bus for 14.5GB/s throughput for SSD streaming, but it's still niche.
- 128GB is the RAM ceiling on the 395, and more memory is simply too expensive.
- He'll explain in an upcoming newsletter why AMD was absent from his earlier "Local AI is finally usable" piece.
More from Infra
- Open-source Dynamo deployment guide with benchmarks on a 16×H100 cluster — TheZachMueller · 2026-09-20
- Kubernetes learning series: logging, monitoring and Helm explained in videos — _jaydeepkarale · 2026-09-20
- Dev shares video series that finally made Kubernetes click: why it exists — _jaydeepkarale · 2026-09-20
- Offer a DGX Station / GB300 as a sign-on bonus and any AI hire signs on the spot — Jasonio · 2026-09-20
- Is Now a Bad Time to Buy a Strix Halo for Local LLMs? RAM Crisis Doubts — -mattmason- · 2026-09-20
- Hacked account ran up an $80K AI bill: no major provider offers a hard spend cap — MaverikSh · 2026-09-20