Open World News

Artificial intelligence infrastructure is evolving on two fronts: the systems that train and serve models, and the physical architecture that powers them. Two recent talks from PyTorch and FINOS capture both sides of this story, offering a look at how AI factories are built and how they can adapt in real time.

In "Elastic Expert Parallelism in vLLM," NVIDIA's Itay Alroy presents a technique for adding or removing GPUs from an active Mixture-of-Experts deployment with minimal interruption to serving and minimal downtime. The approach matters because it lets operators scale capacity during live traffic rather than scheduling disruptive restarts. Alroy's session, delivered at PyTorchCon North America 2026, covers the architecture, key implementation details, and the open challenges that remain.

Meanwhile, FINOS offers a broader view with "What Actually Powers an AI Factory?" The video walks through Jensen Huang's five-layer framework, moving from raw energy and GPU chips up through networking and application layers that define NVIDIA infrastructure. This breakdown clarifies how the pieces integrate to scale global AI production.

Together, the two posts connect the micro and macro levels of AI computing: one shows how serving systems can flex under real-world demand, while the other maps the full stack that makes those