NVIDIA has detailed how its Dynamo inference framework and Brev platform tackle planetary-scale AI inference, cutting per-token costs 35x on GB200 hardware.