NVIDIA AI Updates: September 12, 2026

1. Personal AI Router Pools Every GPU on a Local Network Into One Inference Target

NVIDIA put Personal AI Router (PAIR) into beta, which combines the inference capacity of multiple machines on a local network and distributes incoming AI requests across them automatically. The target workload is local multi-agent setups, where a handful of independent model calls fired in parallel will saturate a single GPU while other boxes on the same LAN sit idle. Source

2. BioNeMo Inference Runtime Targets Proteome-Scale Structure Prediction Without Leaving PyTorch

NVIDIA published a walkthrough of BioNeMo Inference Runtime (BioIR), which accelerates supported biomolecular structure-prediction models through optimized kernels and, where applicable, CUDA Graphs. The design constraint is that it keeps the familiar PyTorch workflow intact, so the speedup applies to moving a whole worklist through a pipeline rather than requiring teams to port models into a separate serving stack. Source