WHERE THE WORK RUNS
ON-PREMISES · LOCAL · OPTIONAL CLOUD
Local decision · NPU
Triage and guard the request next to the user.
Team workload · Strix Halo
Shared team work on one local box.
Shared inference · MI350P
Instinct MI350P in a Dell PowerEdge, at a stated service level.
I2 · MEASURED CLAIM · METRUM TESTS
~250 concurrent coding agents per MI350P
TERMINAL-BENCH CODING WORKLOADS + RAW MI350P INFERENCE
8-card node counts are extrapolated · MODEL
MEASURED · ONE TEST SET
48 of 48 prompts routed as labelled
LAPTOP NPU → INSTINCT GPUS · METRUM AI ROUTER ON THE EDGE TIER
Execution choices — not a ladder every request must climb
CPU AND TOOL WORK, ORCHESTRATION AND INFERENCE SIT IN ONE STORY
ROUTER = GATEWAY, NOT THE SERVING RUNTIME
No guaranteed quality, atomic budgets or automatic privacy.