METRUM AI · launch_overview.mp4 — narration transcript
Runtime 94 s · external film · 3840x2160 authoring grid, 1920x1080 delivery
Voice: ElevenLabs Roger (eleven_multilingual_v2, voice id CwhRBWXzGAhq8TQ4Fs17)
Music: ElevenLabs Music (upbeat build, peak at ~2/3, clean resolve)
Captions: launch_overview.srt, timed to the delivered voice track via the TTS alignment
Every on-screen figure carries its claim ID and label from shared/CLAIMS.md.

[00:00:00,200] A failed attempt cost almost what a successful one cost: 48 core-seconds against 47.
[00:00:07,178] So why do we price AI by the token?
[00:00:09,767] An agent gets a job, plans, calls a model, runs tools, retries, and checks its own work.
[00:00:15,400] In a measured pilot, coding agents on pinned AMD EPYC cores solved the same SQLite build-and-test task.
[00:00:22,888] Failures were not cheap exits: a median 48 core-seconds, against 47 for the ones that passed.
[00:00:30,470] Every failure multiplies the cost of the jobs that get accepted.
[00:00:35,000] The work decides where it runs.
[00:00:37,450] A decision model on the Ryzen AI NPU guards the request next to the user.
[00:00:43,022] A Strix Halo team box carries shared team work.
[00:00:46,842] Instinct MI350P in a Dell PowerEdge is the shared enterprise inference target.
[00:00:52,809] Execution choices — not a ladder every request must climb.
[00:00:56,583] The Router applies policy per agent key.
[00:01:00,000] The unit has to be the job, not the token: inference, tools, retries, verification,
[00:01:06,362] divided by the jobs that were accepted, with the latency stated.
[00:01:10,623] Here is that receipt for one accepted job: about $0.0359 at frontier list prices;
[00:01:16,521] on owned silicon, about $0.0036 at the margin, and $0.0057 all-in.
[00:01:21,675] Same job.
[00:01:23,500] Next: cost per accepted job, measured end to end, with the latency stated.
[00:01:28,829] Right token, right silicon — counted per finished job.
