1
00:00:00,200 --> 00:00:06,365
A failed attempt cost almost what a successful one cost: 48 core-seconds against 47.

2
00:00:07,178 --> 00:00:09,442
So why do we price AI by the token?

3
00:00:09,767 --> 00:00:15,339
An agent gets a job, plans, calls a model, runs tools, retries, and checks its own work.

4
00:00:15,400 --> 00:00:22,076
In a measured pilot, coding agents on pinned AMD EPYC cores solved the same SQLite build-and-test task.

5
00:00:22,888 --> 00:00:29,657
Failures were not cheap exits: a median 48 core-seconds, against 47 for the ones that passed.

6
00:00:30,470 --> 00:00:34,301
Every failure multiplies the cost of the jobs that get accepted.

7
00:00:35,000 --> 00:00:36,637
The work decides where it runs.

8
00:00:37,450 --> 00:00:42,210
A decision model on the Ryzen AI NPU guards the request next to the user.

9
00:00:43,022 --> 00:00:46,029
A Strix Halo team box carries shared team work.

10
00:00:46,842 --> 00:00:51,997
Instinct MI350P in a Dell PowerEdge is the shared enterprise inference target.

11
00:00:52,809 --> 00:00:56,141
Execution choices — not a ladder every request must climb.

12
00:00:56,583 --> 00:00:59,381
The Router applies policy per agent key.

13
00:01:00,000 --> 00:01:06,037
The unit has to be the job, not the token: inference, tools, retries, verification,

14
00:01:06,362 --> 00:01:09,810
divided by the jobs that were accepted, with the latency stated.

15
00:01:10,623 --> 00:01:16,137
Here is that receipt for one accepted job: about $0.0359 at frontier list prices;

16
00:01:16,521 --> 00:01:21,350
on owned silicon, about $0.0036 at the margin, and $0.0057 all-in.

17
00:01:21,675 --> 00:01:22,662
Same job.

18
00:01:23,500 --> 00:01:28,388
Next: cost per accepted job, measured end to end, with the latency stated.

19
00:01:28,829 --> 00:01:32,370
Right token, right silicon — counted per finished job.
