OCTOPAI ENGINEERING ARTS · INTERNAL · THE QUESTION
A FAILED ATTEMPT COSTS WHAT A SUCCESS COSTS
TOKENS ARE EASY TO COUNT · ACCEPTED WORK IS NOT
THE NPU SPECIALIST
Days on the Ryzen AI NPU
Proof the NPU did the work, real models on it, and the one-request-at-a-time surprise.
THE HALO SPECIALIST
Days on the Strix Halo box
What fits in 128 GB, how many people it serves, and the trick the quality gate caught.
THE INSTINCT SPECIALIST
Instinct MI350P on Dell
Shared enterprise inference. Short-turn throughput is not agent-workload throughput.
Why do we price AI by the token?
A failed attempt costs exactly as much as a successful one. Tokens are easy to count; accepted work is not. Two specialist agents went deep on the hardware before anyone wrote a slide.
A4 · FAILED 48 vs PASSED 47 CORE-SECONDS
N1 · NPU 58/60 · 13.5 tok/s
H1 · 16 STREAMS AT 15.8 tok/s
SOURCES · NPU REPORT (PRELIMINARY) · LOCAL BOX BENCH · SHORT-TURN INFERENCE · PILOT #cpuonly
OCTOPAI ENGINEERING ARTS · INTERNAL · THE TEAM
A hundred agents, coordinating
More than a hundred individual agents, spread across machines, working in parallel. Eleven leads, each owning one workstream, with their own subagents.
P3 · INTERNAL · 100+ AGENTS · 11 LEADS
NOT A CONCURRENCY CLAIM
SOURCE · OCTOPAI WORK BREAKDOWN · 100+ AGENTS (CONCEPT/ESTIMATE)