Meter
What the lab’s owned hardware has served. Every number below is either read from instrument files on disk or labeled as an estimate with its heuristic stated. Physics, not bills.
local tokens served
19,152,281
1.7M measured17.5M estimated
as of 2026-07-03 19:51 UTC · regenerated at each deploy
16.2
effective tok/s
completion tokens ÷ full benchmark wall time — includes test execution, so it understates raw generation speed
181
gated benchmark cases
across 16 instrumented runs — every token counted by the harness, not guessed
18,618
squad transcript turns
the working record of the local crew — token volume estimated at chars/4, labeled accordingly
the metal
the Mac
MacBook Pro M5 Max · 128GB
squad host · 5 resident brains · benchmark rig
18.8 W
powermetrics package avg under live 27B benchmark load, 2026-06-10 (band 17.8–19.6W)
—
paired watts×tok/s sampling scheduled
the Jetson
Jetson Orin Nano · 8GB
always-on floor · federation seat
20 W
board envelope under sustained generation, 2026-06-10
0.8 J/tok
measured: gemma4 e2b-it-qat, 25 tok/s sustained at ~20W, 2026-06-10
the 3090
RTX 3090 · 24GB (drone)
image-gen drone · LoRA bakery
—
instrumentation pending
—
instrumentation pending
methodology
Measured numbers come from instrument files on disk; estimates are labeled and computed with stated heuristics. No cloud-pricing comparisons — the meter reports physics, not bills. Generated 2026-07-03T19:51:47+00:00 by jarvis/tools/local_meter.py; regenerated with each site deploy. 1 Wh = 3,600 J — the Jetson’s measured 0.8 J/tok is ~0.00022 Wh/token.