In the same week NVIDIA set out when to split multimodal inference across separate resources, a run of Vision-Language-Action papers pushed models further toward acting, not just seeing and reading.
NVIDIA's technical blog described when to place the Encode, Prefill and Decode stages of multimodal serving on separate resources — "EPD disaggregation" — to cut latency and lift throughput. In the same week, arXiv carried a run of Vision-Language-Action work, including TANGO, which drives a humanoid's whole body through cluttered environments, asking a single model to see, read and act.
EPD disaggregation is a serving-layer design: how to deliver models that already exist more cheaply and quickly. The VLA work sits on the other layer, extending what a model is asked to do at all. The two do different jobs and neither substitutes for the other, so watching only one of them misjudges how close deployment actually is. Efficiency in serving and growth in capability advance at different speeds. Note: the arXiv papers are pre-review, and their performance claims await independent replication.
Deployed cases where EPD disaggregation is adopted and its effect reported in numbers. On the research side, how long it takes for VLA results to reach physical robots or shipping products.