Shared Flash as AI Data Plane
Diving deeper into
Renen Hallak, CEO of VAST Data, on AI agents creating infinite storage demand
pushing AI labs (xAI), neoclouds (CoreWeave, Lambda, Crusoe), and hyperscalers (Microsoft, Google) toward VAST Data’s separated compute/shared flash architecture
Analyzed 4 sources
Reviewing context
This shift shows that AI storage is turning from a box attached to one cluster into a shared data layer that thousands of GPUs can hit at once across clouds and sites.
-
VAST is built for many machines reading and writing to one pool of flash at the same time, which fits AI training, checkpoints, embeddings, and agent memory better than older setups built either for one supercomputer job at a time or for ordinary enterprise backups and databases.
-
This matters most for neoclouds and labs because their product is rented GPU time. If storage stalls the GPUs, the customer pays for expensive idle chips. CoreWeave has become a multi billion dollar GPU cloud, and VAST has become its standard storage layer, which makes storage architecture part of the cloud product, not back office infrastructure.
-
The competitive split is concrete. Weka and DDN optimize for peak speed on specialized high performance clusters. Pure, Dell, and NetApp optimize for reliability in normal IT estates. VAST is trying to combine scale, speed, and resilience in one system, then add database and data orchestration on top so customers can build AI pipelines on the same substrate.
The next step is that shared flash becomes the default data plane for agentic infrastructure. As AI clouds consolidate and hyperscalers retrofit their stacks for AI, the winning vendor is the one that can keep data portable across on-prem, neocloud, and public cloud while feeding larger GPU estates without forcing customers to rebuild their applications around storage limits.