The Superintelligence Podcast

Why AI GPUs wait for data

Sven Breuner

Sven Breuner

VAST Data

VAST Data

“There’s no point in having your GPUs fully utilized if you don’t think that it’s really making a difference for your research or for your business.”

In this conversation

More GPUs do not solve every AI infrastructure problem, because expensive chips waste money when storage cannot deliver data fast enough. Sven Breuner from VAST Data joins Kim Isenberg on The Superintelligence Podcast to explain why the data layer decides whether GPUs are put to good use. Breuner, who started the development of the parallel file system BeeGFS at Fraunhofer, contrasts classic HPC file systems, where each server owns its own slice of data, with VAST’s design. He describes what VAST means by an “AI Operating System”: a platform that stores data, adds database and vector capabilities, and runs functions as new files arrive, so a new video can be queried right away. They discuss why separate GPU, storage and networking teams slow AI projects down, why systems that grow organically become fragile patchwork, and why some enterprises should rent GPUs from neoclouds instead of buying them.

Key takeaways

  • AI workloads read data far more randomly than classic HPC workloads, which is why spinning disks no longer work and VAST put all data on flash.

  • Archives that used to be rarely touched are now the most valuable training material, so AI needs fast access to all data, not just a small working set.

  • VAST’s DASE (disaggregated shared everything) architecture gives every server equal access to every drive. This enables lossless data reduction and lets compute scale independently of storage capacity.

  • With retrieval augmented generation, the system fetches documents on behalf of the user who asks, so access rights are enforced and new data can be queried right away instead of after retraining.

  • When AI factories scale to tens of thousands of GPUs, storage access is the first thing to break, Breuner says.

The daily AI briefing people actually read.

What shipped, who’s winning, and what it means. Six days a week, in under ten minutes.

The newsletter

The people building the AI future. Read by 160,000+ subscribers.

Subscribe free