Home AI lab with two DGX GB300s serves DeepSeek at 5,774 tok/s, all solar
julianharris · x · 2026-10-13
A solar-powered home AI lab built around two NVIDIA DGX Station GB300s (each with an RTX PRO 6000 sidecar), three DGX Sparks, a Threadripper server and NAS — running everything locally, falling back to frontier APIs only when needed. An open repo (dgx-station-frontier) documents receipts-first benchmarks:
- DeepSeek-V4.1-Flash: 5,774 tok/s at 64 users across both Stations, 430 tok/s single-user, GSM8K 98-99%
- With the 6000 sidecar: 1,856 tok/s at 16 users, 2.7-3.25x faster at 16-32 users
- DeepSeek-V4-Pro (1.6T) on two Stations: 83-95 tok/s single-user, 274-313 at 16 users
- MiMo-V2.6-Pro: 503 tok/s at 16 users
- Repo includes six serving recipes, a microbenchmark atlas, and a validator tying every number to logs
The quip: this isn't "local AI," it's a mini data centre.
More from Infra
- Blog post shrinks from 158KB HTML to 17KB Markdown for AI agents — colinmcnamara · 2026-10-13
- Zuckerberg: an H100 is worth more than last year as AI compute stays scarce — inductionheads · 2026-10-13
- Arm: AI data centers need 3 CPU roles, market to grow from $52B to $300B by 2030 — firstadopter · 2026-10-13
- Every AI compute fix just moves the bottleneck: stacking, optics, big dies each spawn new constraints — BenBajarin · 2026-10-13
- Meta's Hyperion to be world's first 5GW datacenter, a $400B bet on 4,000 acres — bookwormengr · 2026-10-13
- AMD unveils 6th-gen EPYC "Verano" CPU with differentiated memory architecture — ryanshrout · 2026-10-13