METRUM AI · launch_about.mp4 — narration transcript Runtime 124.5 s · external film · 1920x1080 delivery Voice: ElevenLabs Roger Music: ElevenLabs Music (upbeat build, peak at ~2/3, clean resolve) Captions: launch_about.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. [00:00:00,600] A cheap model call can still produce an expensive job. [00:00:04,211] Retries do that — and almost nobody counts them. [00:00:07,833] Ask what a job costs and you will get a token price. [00:00:11,061] That is not the number. [00:00:13,127] The bill is the whole loop, divided by the jobs that were accepted. [00:00:17,400] And that is where the accounting breaks. The bill for an agent job is not the token bill. [00:00:23,089] It is everything the job spent — inference, tools and sandbox time, retries and failed attempts, verification, operating overhead — [00:00:31,645] divided by the jobs that were actually accepted, with the latency stated. [00:00:36,440] One number: cost per accepted job. [00:00:39,400] That number is why enterprises are bringing their baseline back on-premises. [00:00:44,543] Own the inference you run every day. [00:00:46,888] Keep the frontier for the work that genuinely needs it. [00:00:50,604] Treat cloud as a deployment choice — not a destination. [00:00:54,800] AMD brings execution choices, not a ladder. [00:00:58,422] A Ryzen AI NPU for the decision next to the user. [00:01:02,335] Strix Halo for shared team work. [00:01:05,168] Instinct MI350P in a Dell PowerEdge for shared enterprise inference — with CPU and tool work running beside them. [00:01:12,900] And the Metrum AI Router is the control point: policy, per-agent keys, and the accounting. [00:01:19,000] Now make it arithmetic. One modelled workload — 500 seats, 4 always-on agents each, 17 h a day, 3% escalation. [00:01:29,937] It runs at about $419,451 a month at frontier list prices; about $33,111 on-premises; about $34,644 all-cloud. [00:01:40,583] Owned silicon wins by a mile against a frontier bill, and only narrowly against well-used elastic GPUs. [00:01:48,304] Occupancy is the lever. [00:01:50,700] So here is the question AMD, Dell and Metrum are answering together: what does an accepted agent job actually cost — and where should it run? [00:01:59,628] Right token, right silicon, counted per finished job.