Vast Data bets stitched-together AI pipelines can’t keep sensitive data in sync

Vast Data is arguing that the real bottleneck in enterprise AI is no longer model training. It’s inference, and more specifically the mess that shows up when companies try to bolt together separate tools for storage, embeddings, permissions, and queries. Vaughn Stewart, Vast’s vice president of systems engineering, says that setup creates too much risk that the data, the vector embeddings, and the access rules drift apart.

That matters because the most valuable data is often the least movable. Stewart said about half of the world’s data still lives in corporate data centers or colocation facilities, largely because it’s sensitive and sits in regulated industries. Much of it can’t legally be sent to a public cloud frontier model. So the question is not just speed anymore. It’s whether the infrastructure can respect who is allowed to see what.

Vast’s pitch is that a tightly integrated, disaggregated shared-everything architecture can do more than old storage categories suggest. Stewart compared it to systems like Hadoop data lakes and Kafka event buses, which were never called storage platforms but still held huge amounts of data. His point is blunt: if the pipeline has to move data from system to system, the access controls can get lost on the way, and the promise of near-real-time handling starts to look shaky at mass scale.

The company has also been getting paid for that argument. Vast raised about $1 billion at a $30 billion valuation in April 2026. And in September 2026 it announced DataEnclave, a confidential AI runtime built with partners including Nvidia and Cohere. It runs models in secured virtual machines, demands remote cryptographic attestation before keys are released, and logs access in a tamper-proof database. Vast is betting that this is the missing bridge between frontier models and regulated data that companies actually want to use.