# technical_story.mp4 — transcript

Metrum AI · “The inside story” · the making of the AMD + Dell + Metrum AI Router launch media.
Narration: synthetic voice (ElevenLabs, voice “Roger”, eleven_multilingual_v2). Music: ElevenLabs Music.
Runtime 5:40.1 · 1920×1080 · captions: technical_story.srt.

Every figure that appears on screen carries a claim ID and a label (MEASURED / MODEL / VENDOR CLAIM / ILLUSTRATION / PLAN). Claim text: shared/CLAIMS.md.

**0:03.5** — Two agents get the same job. One finishes it. One fails and retries. Which cost more? Price AI by the token and the answer is: almost the same. That is the problem.

**0:16.7** — This is measured. Across one pilot, a failed attempt took a median of 48 core-seconds of CPU. A passing attempt took 47. The failure cost slightly more than the success.

**0:30.3** — And you pay for it twice: once for the attempt, once for the retry that finally passes. At a 50% failure rate, one accepted job costs twice as much.

**0:42.5** — These numbers come from one pilot: SQLite coding agents on pinned AMD EPYC cores, with Qwen3 Coder Next on Amazon Bedrock. Not NPU, Halo or MI350P validation.

**0:55.8** — First surprise. Across roughly 4,000 model calls, zero were throttled. None. And agents still spent a large share of every run waiting on the model. Zero throttling does not mean zero waiting — the bottleneck was never the rate limit.

**1:15.2** — Two specialist agents went deep before anyone wrote a slide. The first spent days with one laptop and its NPU, and proved the chip was doing the work: busy time grows with prompt length, and every other path read zero.

**1:28.2** — Then it put real models on it. Thirty-five billion parameters, on the NPU: 58 out of 60 on a fixed suite — the best score of anything tested, above both cloud references. The iGPU sat idle.

**1:41.2** — And then the surprise. A decision took 0.40 s on the NPU. With a chat model on the same NPU: 39 s. Not a slow chip — a chip serving one request at a time.

**1:54.8** — Three decision models, between a fifth of a second and nine tenths of a second each. 23/24, 23/24 and 24/24. All three passed the fidelity check.

**2:05.8** — Then the whole edge tier, on the laptop. Easy prompts stayed local; hard ones went to a core of two Instinct GPUs. 48 prompts, 48 routed as labelled. And the full set ran in 2.9 minutes, down from 22.1.

**2:19.8** — The second agent took the team-sized box. One machine, 128 GB of unified memory: 16 interactive chat streams at 15.8 tokens a second each. Sixteen people, one desk-side box.

**2:33.8** — One host software upgrade made the mixture-of-experts model up to 64% faster per stream, and made the 122B model slower. Same silicon, different stack.

**2:45.8** — Then a speed trick that looked like a win. Hard code went from 29 passes out of 40 to 6, with 25 new syntax errors. A benchmark that only measured speed would have scored it as an improvement.

**2:57.8** — Modelled, with assumptions on screen: $0.77 to $1.05 per million output tokens on one box. Hosted, the same model: $0.95. Frontier APIs: $10 to $20.

**3:09.8** — Meanwhile the sizing work asked two questions. First: how many agents keep the CPU above 80% busy? Five and eight on one core. Twelve on two.

**3:21.8** — And a fine-tuned model that looked like a seven-fold win on tool calls — until it lost most of its general knowledge. The gate refused promotion.

**3:34.7** — Six sustained studies, all invalid — not because the jobs failed, but because the instrumentation did. A prediction sealed before the run missed. That is the finding.

**3:47.7** — So here is the mental model. Stop counting tokens; count jobs. Add up everything one accepted job took — inference, tool work, retries, verification, overhead — then divide by the jobs that were accepted.

**4:02.1** — Now the twist. Frontier list: $0.0359. Owned silicon: $0.0036 marginal, about $0.0057 all-in. Just under ten times cheaper — or 6.3 times, counting the hardware. The tool seconds, retries and verification did not change.

**4:20.8** — Scale it up. 500 seats, 4 always-on agents each, 17 hours a day, 3% escalating to a frontier API. All-frontier: about $419,451 a month. Owned plus that escalation: about $33,111. Elastic cloud: about $34,644.

**4:40.9** — Owned wins above 75% occupancy; below that, rent. Escalation share moves the bill most, cached input dominates the frontier price, and retries decide how many attempts one accepted job costs.

**4:54.9** — Then the swarm assembled the proposal. More than a hundred individual agents, distributed, coordinating with each other: one story, one claim register, one cost model, and the films.

**5:08.9** — And the next step is the interesting one. The plan: the solution will be built by a swarm that routes its own tokens through the Metrum AI Router — built on the system it describes, every hop priced in the ledger.

**5:21.9** — Which brings it back to the question we started with. The token is a unit of purchase; the job is the unit of work. Price the work. Label every number, and when a gate says no, that is the result.
