Open World News

Artificial intelligence infrastructure is evolving on two fronts at once: the software that keeps large models serving smoothly, and the physical systems that make those models possible. Two recent videos from PyTorch and FINOS capture both sides of that story.

In "Elastic Expert Parallelism in vLLM," presented by Itay Alroy of NVIDIA at PyTorchCon North America 2026, viewers get a look at a technique that allows GPUs to be added to or removed from an active Mixture-of-Experts deployment with minimal interruption and downtime. This matters for real-world serving, where traffic fluctuates and operators need flexibility without sacrificing reliability. The talk covers the architecture behind elastic expert parallelism, key implementation details, and the open challenges that remain.

Meanwhile, FINOS offers a broader perspective with "What Actually Powers an AI Factory?" The video breaks down Jensen Huang's five-layer framework, tracing the stack from raw energy and GPU chips up through networking and application layers. Together, these posts connect the dots between low-level infrastructure and high-level AI production, showing how energy, hardware, and software must integrate to scale modern AI systems.

For anyone tracking where AI compute is heading, both videos provide useful context, one focused on the mechanics of efficient model serving,