Together, these notes suggest builders must plan where and how inference runs, not only which model to call.
From the Field Notes:
Inference is becoming a placement and capacity-planning problem across hardware and workloads
In one minute
NVIDIA PAIR connects local devices behind one endpoint for private inference, while Gimlet describes coordinating heterogeneous silicon through an execution planner, compiler, and runtime. An inference-sizing guide frames memory, batching, latency, throughput, and service targets as linked capacity decisions.