1
00:00:00,600 --> 00:00:03,770
A cheap model call can still produce an expensive job.

2
00:00:04,211 --> 00:00:07,020
Retries do that — and almost nobody counts them.

3
00:00:07,833 --> 00:00:10,852
Ask what a job costs and you will get a token price.

4
00:00:11,061 --> 00:00:12,431
That is not the number.

5
00:00:13,127 --> 00:00:17,133
The bill is the whole loop, divided by the jobs that were accepted.

6
00:00:17,400 --> 00:00:22,915
And that is where the accounting breaks. The bill for an agent job is not the token bill.

7
00:00:23,089 --> 00:00:31,402
It is everything the job spent — inference, tools and sandbox time, retries and failed attempts, verification, operating overhead —

8
00:00:31,645 --> 00:00:35,628
divided by the jobs that were actually accepted, with the latency stated.

9
00:00:36,440 --> 00:00:38,809
One number: cost per accepted job.

10
00:00:39,400 --> 00:00:43,731
That number is why enterprises are bringing their baseline back on-premises.

11
00:00:44,543 --> 00:00:46,610
Own the inference you run every day.

12
00:00:46,888 --> 00:00:49,907
Keep the frontier for the work that genuinely needs it.

13
00:00:50,604 --> 00:00:54,168
Treat cloud as a deployment choice — not a destination.

14
00:00:54,800 --> 00:00:58,039
AMD brings execution choices, not a ladder.

15
00:00:58,422 --> 00:01:02,010
A Ryzen AI NPU for the decision next to the user.

16
00:01:02,335 --> 00:01:04,471
Strix Halo for shared team work.

17
00:01:05,168 --> 00:01:12,517
Instinct MI350P in a Dell PowerEdge for shared enterprise inference — with CPU and tool work running beside them.

18
00:01:12,900 --> 00:01:18,438
And the Metrum AI Router is the control point: policy, per-agent keys, and the accounting.

19
00:01:19,000 --> 00:01:29,240
Now make it arithmetic. One modelled workload — 500 seats, 4 always-on agents each, 17 h a day, 3% escalation.

20
00:01:29,937 --> 00:01:39,631
It runs at about $419,451 a month at frontier list prices; about $33,111 on-premises; about $34,644 all-cloud.

21
00:01:40,583 --> 00:01:47,491
Owned silicon wins by a mile against a frontier bill, and only narrowly against well-used elastic GPUs.

22
00:01:48,304 --> 00:01:50,115
Occupancy is the lever.

23
00:01:50,700 --> 00:01:58,815
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?

24
00:01:59,628 --> 00:02:03,099
Right token, right silicon, counted per finished job.
