[ glyph ]

AI Hub

What the fleet actually did, in real numbers.

Snapshot as of August 16, 2026 — updated periodically, not live.

7.87 billion tokens processed · 53 active days · 39 local models · 1 RTX 3090

⚠️ What the dollar figure actually means. This fleet runs on a flat-rate Claude subscription, not metered API billing — so there is no bill this number avoided. It’s an API-equivalent value: what the same work would have cost at standard cloud-API pricing, priced consistently and shown so the scale of the work is honest, not so it reads as money saved. $5,006 in API-equivalent value, accumulated over 53 active days since June 1, 2026.

Every one of those tokens did real work — drafting newspaper issues, transcribing podcast episodes, indexing documents, running the personas in the Superdawg agency, answering questions against the private librarian. None of it touched a cloud model that wasn’t supposed to see it.

Local models available The working roster on the 3090 — MoE models for agent work, dedicated picks for translation, recall, vision, and embeddings, each chosen by a real bake-off rather than a guess. 39 models · one GPU · one tenant at a time
Cloud routing Work is routed to the cheapest lane that can actually do the job: local first, subscription for agentic coding and audits, metered cloud only for the rare second opinion. Nothing defaults to the expensive lane. local · subscription · metered, in that order
GPU One RTX 3090, 24GB VRAM, running Whisper transcription, embeddings, image ML, and every local LLM in the lab — on a strict one-big-tenant-at-a-time rule so nothing silently spills into slow CPU fallback. 24GB VRAM · ~17× realtime Whisper · one tenant at a time

Numbers come from the fleet’s own usage logs, summed directly — not estimated. This page is a snapshot, not a live feed; it gets refreshed by hand for now.