1
00:00:03,500 --> 00:00:13,897
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.

2
00:00:16,704 --> 00:00:27,466
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.

3
00:00:30,294 --> 00:00:39,985
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.

4
00:00:42,547 --> 00:00:54,380
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.

5
00:00:55,771 --> 00:01:09,590
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.

6
00:01:15,192 --> 00:01:27,391
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.

7
00:01:28,192 --> 00:01:40,757
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.

8
00:01:41,192 --> 00:01:52,738
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.

9
00:01:54,782 --> 00:02:04,055
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.

10
00:02:05,782 --> 00:02:19,575
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.

11
00:02:19,782 --> 00:02:32,582
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.

12
00:02:33,782 --> 00:02:43,343
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.

13
00:02:45,782 --> 00:02:57,433
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.

14
00:02:57,782 --> 00:03:09,145
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.

15
00:03:09,782 --> 00:03:18,638
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.

16
00:03:21,830 --> 00:03:29,797
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.

17
00:03:34,670 --> 00:03:43,891
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.

18
00:03:47,659 --> 00:03:59,910
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.

19
00:04:02,139 --> 00:04:18,857
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.

20
00:04:20,799 --> 00:04:39,320
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.

21
00:04:40,889 --> 00:04:53,036
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.

22
00:04:54,899 --> 00:05:04,826
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.

23
00:05:08,899 --> 00:05:20,367
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.

24
00:05:21,899 --> 00:05:33,419
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.
