Cloud AI seats are rent
Eight seats at public list price, month after month. The appliance is a one-time purchase, quoted in writing, and the marginal question after that is electricity.
is 36 months of cloud AI seats for a team of eight, before overages.
Source: modelled — assumptions listed. 8 seats × $30 per seat per month × 36 months, public list prices of common cloud AI assistants; months on the x axis
Where the last thirty answers were produced
Every answer the appliance logs carries the tier that produced it. The graph answers from facts, the small local model reasons on the box, and the frontier is a cloud model reached only by consent.
of the last thirty logged answers were produced on the appliance; the frontier was not called.
Source: measured on the reference deployment, as of 2026-09-10. the appliance's turn log read on 2026-09-10 (last 30 turns, 2026-09-09 18:22–21:12 UTC); those turns are the ask-harness runs r4 and r5, a question mix balanced across tiers on purpose, so the split is the harness's, not a visitor's
From question to a spoken answer
The same two-sentence answer rendered to speech on the appliance, before and after the speech model moved to the appliance's GPU.
from question to a spoken answer, rendered on the appliance, down from 12.2 s.
Source: measured on the reference deployment, as of 2026-09-09. say path measured through the tunnel on 2026-09-09, three fresh 46-word renders each — CPU image 12534 / 9577 / 12177 ms, GPU image 1385 / 1415 / 1407 ms, medians shown
Your numbers, on your operation
A 30-minute walkthrough scopes your exact stack. The quote comes in writing.
Talk to us