The Data Foundation for HPC, AI, and Quantum

FEATURED SESSION AT ISC 2026

Rethinking Data Infrastructure Conventions for Modern AI and HPC

This was one of the hot topics at ISC 2026: as AI workflows reshape HPC, the old conventions built around running one large application as fast as possible no longer hold.

Glenn Lockwood broke down how training, fine-tuning, and inference each demand distinct data access patterns and what that means for integrating AI with traditional simulation workloads.

FEATURED DEMOS

Experience the end-to-end AI pipeline showcase

VAST AI OS for Genomics: An End-to-End AI Pipeline

An end-to-end genomics AI pipeline on one VAST AI OS, raw DNA to GPU-accelerated variant calling (NVIDIA Parabricks) to AI clinical synthesis, closing with a preview of NVIDIA BioNeMo drug discovery. One platform connecting HPC analysis to AI discovery, with no data staging.

Document RAG on the VAST AI OS: Grounded Answers from Your Own Corpus

An end-to-end Document RAG pipeline on one VAST AI OS: files vectorize automatically the moment they land, and a research assistant returns grounded, cited answers. No separate vector database, fully observable, and built on a no-code pipeline builder.

Video Search and Summarization on the VAST AI OS: Ask Your Video Anything

Turn unstructured video into a searchable, cited source of truth on one VAST AI OS. A DataEngine pipeline segments, reasons over (vision-language model), embeds, and stores video so you can ask it questions in natural language and verify each answer against the source frame.

VAST AI OS for Genomics: An End-to-End AI Pipeline

An end-to-end genomics AI pipeline on one VAST AI OS, raw DNA to GPU-accelerated variant calling (NVIDIA Parabricks) to AI clinical synthesis, closing with a preview of NVIDIA BioNeMo drug discovery. One platform connecting HPC analysis to AI discovery, with no data staging.

Document RAG on the VAST AI OS: Grounded Answers from Your Own Corpus

An end-to-end Document RAG pipeline on one VAST AI OS: files vectorize automatically the moment they land, and a research assistant returns grounded, cited answers. No separate vector database, fully observable, and built on a no-code pipeline builder.

Video Search and Summarization on the VAST AI OS: Ask Your Video Anything

Turn unstructured video into a searchable, cited source of truth on one VAST AI OS. A DataEngine pipeline segments, reasons over (vision-language model), embeds, and stores video so you can ask it questions in natural language and verify each answer against the source frame.

Six reasons HPC and AI leaders are moving to VAST.

From TACC to CINECA, the world's leading compute centers are choosing VAST. Here's what's driving it.

No more parallel file system complexity

VAST replaces Lustre and Spectrum Scale with a completely parallel data and metadata path. No MDS tuning, no mandatory tiering. TACC said it best: VAST "just makes the bad I/O smoother than Lustre."

GPU utilization goes from 60% to where it belongs

When storage can't deliver mixed HPC and AI I/O patterns at GPU speed, every idle GPU cycle is wasted capital. VAST's DASE architecture removes the bottleneck entirely: certified GPUDirect Storage, TB/s sustained throughput, no I/O stalls eating into your compute budget.

One platform for simulation and AI. No staging, no silos

HPC simulation jobs and AI training pipelines access the same governed namespace via NFS-over-RDMA, S3, and GPUDirect Storage simultaneously. DataEngine closes the loop with event-driven triggers: simulation output flows directly into training pipelines.

No more parallel file system complexity

VAST replaces Lustre and Spectrum Scale with a completely parallel data and metadata path. No MDS tuning, no mandatory tiering. TACC said it best: VAST "just makes the bad I/O smoother than Lustre."

GPU utilization goes from 60% to where it belongs

When storage can't deliver mixed HPC and AI I/O patterns at GPU speed, every idle GPU cycle is wasted capital. VAST's DASE architecture removes the bottleneck entirely: certified GPUDirect Storage, TB/s sustained throughput, no I/O stalls eating into your compute budget.

One platform for simulation and AI. No staging, no silos

HPC simulation jobs and AI training pipelines access the same governed namespace via NFS-over-RDMA, S3, and GPUDirect Storage simultaneously. DataEngine closes the loop with event-driven triggers: simulation output flows directly into training pipelines.

Multi-tenant research computing without the overhead

Per-tenant encryption keys, independent QoS, resource quotas, and audit trails, enforced at the storage layer. Physics, genomics, and the AI team share infrastructure without seeing each other's data.

All-flash economics that actually pencil out

Up to 3:1 average data reduction makes all-flash competitive with, or better than, tiered spinning-disk architectures on a cost-per-usable-PB basis. Lower power, higher density, and no hot/warm/cold management overhead.

The full AI lifecycle: from training to RAG to agents

InsightEngine automates the complete RAG pipeline on the same platform where data lives. AgentEngine orchestrates agentic AI workflows with persistent memory and policy enforcement. One namespace, one governed data layer, from first training run to production.