3 PCs, 120+ threads, 36 coding agents: he says local setups need CPUs, not GPUs
kevinnbass · x · 2026-10-30
Kevin Bass shows off a 3-PC mini-cluster in his apartment (main + 2 workers, 120+ threads total) and argues a contrarian take: local workloads need more CPUs, not GPUs — "stop trying to run local inference." He plans to buy 6 more machines.
The attached dashboard shows 38-70% load across the machines running 36 interactive sessions, 24 Codex lanes and multiple Claude tasks, with per-lane CPU waterfalls, memory gating and disk水位 monitoring. The core idea: instead of stacking GPUs for inference on one box, spread massive parallel coding-agent workloads across cheap CPU machines.
More from coding & agent
- Tactics for rolling out 'software factories': the human and org changes that matter — blaizedsouza · 2026-10-30
- AI Demos Flop in Production Because You Picked the Examples: Build a 50-100 Case Eval Set First — salespire · 2026-10-30
- Engineering is now 20% prompting and 80% tuning the agents that execute your prompts — blaizedsouza · 2026-10-30
- Xcode 27.2 ships an Instruments skill letting coding agents diagnose app hangs — JordanMorgan10 · 2026-10-30
- AI speedup PRs don't need full merges — spotting optimization opportunities is the real goal — jmschreiber91 · 2026-10-30
- Banks Detect Unauthorized AI Agents Operating Inside Customer Accounts — ChuckDBrooks · 2026-10-30