AI工厂的隐藏瓶颈：数据基础设施

The Hidden Bottleneck in AI Factories: Data Infrastructure with Sven Breuner, VAST Data

Everyone talks about GPUs. But in real-world AI infrastructure, the bottleneck is often somewhere else: the data layer.

In this episode, I speak with Sven Breuner from VAST Data about the infrastructure behind modern AI systems - from GPU utilization and storage architecture to HPC, inference, data movement, and the emerging concept of AI factories.

As companies invest billions into GPUs, many are discovering that compute alone is not enough. If data cannot move fast enough, GPUs sit idle, training becomes inefficient, and inference systems struggle to scale.

In this conversation, we discuss:

• Why GPUs often wait for data • How AI workloads differ from traditional HPC workloads • What happens when storage, metadata, and networking become bottlenecks • Why the data layer matters for both training and inference • What VAST means by an “AI Operating System” • How enterprises should think about infrastructure before buying more GPUs • Why the next generation of AI factories will require more than faster chips This is a deep dive into one of the most important but least understood parts of the AI stack: the infrastructure that makes large-scale intelligence possible.

YouTube: https://youtu.be/V7QsxbS2UXs