1
00:00:00,000 --> 00:00:04,911
Before a single slide existed, two agents
spent days arguing with silicon.

2
00:00:06,300 --> 00:00:10,521
A failed attempt costs exactly as much as a
successful one.

3
00:00:10,521 --> 00:00:14,455
Forty-eight core-seconds against forty-
seven, measured.

4
00:00:14,455 --> 00:00:20,178
So why does the whole industry price AI by
the token? Because tokens are easy to

5
00:00:20,178 --> 00:00:22,467
count, and accepted work is not.

6
00:00:22,467 --> 00:00:26,545
We spent a week working out what an agent
actually costs.

7
00:00:27,467 --> 00:00:31,901
A hundred individual agents, spread across
machines, working in parallel.

8
00:00:31,901 --> 00:00:35,972
Eleven leads, each owning one workstream,
with their own subagents.

9
00:00:35,972 --> 00:00:37,916
They coordinate with each other.

10
00:00:39,567 --> 00:00:42,967
The work ran as a constellation, not a
pipeline.

11
00:00:42,967 --> 00:00:47,005
Phosphoros went deep on the NPU inside a
Dell Pro laptop.

12
00:00:47,005 --> 00:00:49,768
Hestia went deep on the Strix Halo box.

13
00:00:49,768 --> 00:00:53,948
Enodia owned the router hierarchy and the
containment demo.

14
00:00:53,948 --> 00:00:56,003
Hephaestus the Instinct core.

15
00:00:56,003 --> 00:00:58,482
Astraea the evals and the final QC.

16
00:00:58,482 --> 00:01:00,182
Plutus priced every hop.

17
00:01:00,182 --> 00:01:02,662
Lynceus recorded what actually ran.

18
00:01:02,662 --> 00:01:05,071
Thaumas built the show experience.

19
00:01:05,071 --> 00:01:08,967
Daedalus the brand pack and the three.js
wireframe kit.

20
00:01:08,967 --> 00:01:10,171
Apollo the films.

21
00:01:10,171 --> 00:01:14,989
And Calliope held the narrative, the scripts
and the claim register.

22
00:01:15,400 --> 00:01:20,454
More than a hundred agents at peak: eleven
leads and eighty-nine named subagents,

23
00:01:20,454 --> 00:01:21,390
plus their own.

24
00:01:21,390 --> 00:01:24,073
An estimate, not a claim about concurrency.

25
00:01:25,400 --> 00:01:27,779
Then the swarm assembled the proposal.

26
00:01:27,779 --> 00:01:30,658
One story, one claim register, one cost
model,

27
00:01:30,658 --> 00:01:35,666
a deck and the films — a hundred agents
working in parallel on the same problem.

28
00:01:39,267 --> 00:01:41,571
The NPU specialist went first,

29
00:01:41,571 --> 00:01:47,794
and its first job was not to make the NPU
look good — it was to prove the NPU did

30
00:01:47,794 --> 00:01:49,024
anything at all.

31
00:01:49,024 --> 00:01:55,400
Busy time grew in a straight line with
prompt length: 0.377 milliseconds per prompt

32
00:01:55,400 --> 00:01:58,704
token, R² 0.991, over twenty-five requests.

33
00:01:58,704 --> 00:02:01,393
Every other path read exactly zero.

34
00:02:02,233 --> 00:02:03,372
Then the models.

35
00:02:03,372 --> 00:02:09,205
A thirty-five-billion-parameter mixture-of-
experts ran on the NPU at 13.5 tokens a

36
00:02:09,205 --> 00:02:14,469
second and scored 58 out of 60 on a fixed
quality suite — above both cloud

37
00:02:14,469 --> 00:02:15,251
references.

38
00:02:15,251 --> 00:02:19,448
Whisper ran a clip in 2.74 seconds against
4.33 on the CPU.

39
00:02:20,600 --> 00:02:26,691
And then the Router itself ran on the
laptop: all forty-eight test prompts routed

40
00:02:26,691 --> 00:02:31,578
exactly as labelled, the full set in 2.9
minutes, down from 22.1.

41
00:02:31,578 --> 00:02:34,210
The edge of the hierarchy, working.

42
00:02:35,100 --> 00:02:37,491
That is the finding, not a defect.

43
00:02:37,491 --> 00:02:42,623
One device, one request at a time — and it
changes the recommended split.

44
00:02:45,867 --> 00:02:50,068
Wait, really? The same decision, three
configurations,

45
00:02:50,068 --> 00:02:54,736
one machine: 0.40 seconds alone, 39 seconds
sharing the NPU,

46
00:02:54,736 --> 00:02:57,771
0.45 seconds with the chat on the iGPU.

47
00:02:57,771 --> 00:03:04,150
Fifty out of fifty decisions still correct,
and the chat only nine percent slower.

48
00:03:04,150 --> 00:03:07,496
Preliminary, one device, labelled that way.

49
00:03:12,100 --> 00:03:15,283
It changed the recommended split, it changed
a slide,

50
00:03:15,283 --> 00:03:17,625
and it became a scene in all the films.

51
00:03:17,625 --> 00:03:22,610
It also went upstream as an issue — a
limitation you can measure is worth more
than

52
00:03:22,610 --> 00:03:23,751
a claim you cannot.

53
00:03:28,367 --> 00:03:30,685
The Halo specialist went second.

54
00:03:30,685 --> 00:03:31,264
One box,

55
00:03:31,264 --> 00:03:35,973
128 gigabytes of unified memory — and memory
decides what fits: a

56
00:03:35,973 --> 00:03:40,827
122-billion-parameter model fits only
because the memory is shared.

57
00:03:40,827 --> 00:03:45,318
Sixteen interactive chat streams at 15.8
tokens a second each.

58
00:03:45,318 --> 00:03:51,041
Against an NVIDIA GB10 it leads by 1.3 times
for one user, and ties at sixteen.

59
00:03:51,867 --> 00:03:56,804
Then the surprise: software moved the
numbers more than silicon did.

60
00:03:56,804 --> 00:04:01,960
One host upgrade sped the mixture-of-experts
model up 45 to 64 percent.

61
00:04:01,960 --> 00:04:07,987
And one n-gram trick looked like a big win —
until the quality gate caught it: hard

62
00:04:07,987 --> 00:04:12,488
code fell from 29 of 40 to 6 of 40, with 25
new syntax errors.

63
00:04:12,488 --> 00:04:16,918
A speed-only benchmark would have called
that an improvement.

64
00:04:17,700 --> 00:04:20,814
None of it counts until the telemetry is
honest.

65
00:04:20,814 --> 00:04:23,538
GPU power here was reporting socket power.

66
00:04:23,538 --> 00:04:25,484
Two wrong metrics became none.

67
00:04:26,100 --> 00:04:29,036
Then the swarm assembled the media itself.

68
00:04:29,036 --> 00:04:34,699
A deterministic wireframe kit — laptops,
racks, particle streams, the agent loop,

69
00:04:34,699 --> 00:04:37,495
a router gate — rendered frame by frame,

70
00:04:37,495 --> 00:04:40,990
and every frame audited for colour and
legibility.

71
00:04:45,400 --> 00:04:50,265
A hundred agents, in parallel, on one
problem: the analysis, the charts,

72
00:04:50,265 --> 00:04:54,588
the diagrams, the films, the deck —
assembled and cross-checked.

73
00:04:54,588 --> 00:04:56,345
Nothing was written twice.

74
00:04:58,967 --> 00:05:01,976
Which is why provenance was the hard part.

75
00:05:01,976 --> 00:05:06,991
A prior result, a current capture and an
illustration must never blur.

76
00:05:06,991 --> 00:05:11,218
Every frame carries its label, its source
and its timecode.

77
00:05:15,100 --> 00:05:19,188
And the discipline: every number on screen
carries a claim ID,

78
00:05:19,188 --> 00:05:23,407
and every claim that could not be sourced
was cut, not softened.

79
00:05:32,800 --> 00:05:35,690
The cost question took the longest to
settle.

80
00:05:35,690 --> 00:05:40,827
The unit was settled by the pilot's own
evidence: forty-eight core-seconds for a

81
00:05:40,827 --> 00:05:43,909
failed attempt, forty-seven for one that
passed.

82
00:05:43,909 --> 00:05:46,542
If a failure costs the same as a success,

83
00:05:46,542 --> 00:05:49,753
then the only honest denominator is accepted
jobs.

84
00:05:55,000 --> 00:05:56,600
So they built one model,

85
00:05:56,600 --> 00:06:00,866
and everything else reads from it: the
receipt for a single job,

86
00:06:00,866 --> 00:06:05,466
the three-scenario comparison, the break-
even occupancy, the payback,

87
00:06:05,466 --> 00:06:10,333
and the four levers — retries, escalation
share, cached input, occupancy.

88
00:06:10,333 --> 00:06:15,199
Every cost number in the deck and the films
comes out of that one script,

89
00:06:15,199 --> 00:06:19,132
and it refuses to print if it stops matching
the plan page.

90
00:06:19,132 --> 00:06:23,865
And one thing never appears in any of it:
what it cost us to make this.

91
00:06:27,333 --> 00:06:28,835
Lessons, internal only.

92
00:06:28,835 --> 00:06:34,252
Test the harness before the model — three of
our own failures came from test setup,

93
00:06:34,252 --> 00:06:35,231
not the models.

94
00:06:35,231 --> 00:06:39,734
Label the scope, or the number will travel
further than its evidence.

95
00:06:39,734 --> 00:06:43,259
Instrumentation gaps are findings, not
embarrassments;

96
00:06:43,259 --> 00:06:48,676
the pilot's six invalid sustained studies
are more useful than a smooth curve would

97
00:06:48,676 --> 00:06:49,328
have been.

98
00:06:50,833 --> 00:06:53,513
And here is the part that is still a plan.

99
00:06:53,513 --> 00:06:58,680
The next campaign will be built by a swarm
that routes its own tokens through the

100
00:06:58,680 --> 00:07:02,124
Metrum AI Router, with every hop priced in
the ledger.

101
00:07:02,124 --> 00:07:05,122
The solution, built on the system it
describes.

102
00:07:05,800 --> 00:07:06,741
Internal only.

103
00:07:06,741 --> 00:07:10,237
What ships is the deck, the films and their
sources.

104
00:07:10,237 --> 00:07:14,002
What stays here is how they were made — and
who said no.
