Vast Data's answer to enterprise AI's biggest trust problem

Key points

Key takeaways from a Big Technology Podcast interview with Vast Data co-founder and CEO Renen Hallak (September 2026):

The inflection point. Hallak calls OpenAI's reported breakthrough on the Navier-Stokes millennium problem the most exciting AI advance since Vast started, because it marks the first time a model solved something no human could — proof that AI systems can generate genuinely new intellectual property.

The trust standoff. Enterprise adoption is booming — Hallak expects OpenAI and Anthropic combined to approach or exceed $100 billion in ARR this year, up from roughly $10 billion last year — but companies fear proprietary data being absorbed into the next model, while labs risk model weights worth billions of dollars to distillation attacks.

Encrypted weights, secret data. Vast's answer is confidential computing: encrypted memory at hardware level (built on Nvidia's constructs), with Vast managing the keys and control plane, so neither Vast nor the customer can see the model weights while inference still runs — on-premises or within a customer's virtual private cloud.

Models as a managed resource. Echoing Jensen Huang's five-layer cake framing, Hallak argues models are becoming a resource like GPUs, SSDs or network infrastructure, and that the operating system layer — which Vast's platform, born as storage and database, now occupies — must own model management, security and serving.

The breadcrumbs problem. Every interaction with a frontier model may leave breadcrumbs (Satya Nadella's term) that expose metadata about a company's work; Hallak argues protection belongs in the infrastructure layer, not the application, especially as highly regulated firms adopt AI.

Distillation is the weights thief. Query-based distillation can eventually be defended, but once bare weights are exposed there is no recovery — like an MRI revealing the connections of a brain, says Hallak, who expects enterprises to guard fine-tuned weights as their own crown-jewel IP.

Billions of personalised models. A 1,000–10,000-person enterprise could run a million agents within 5–10 years, with fine-tuned models becoming an organisation's most valuable assets; Vast aims to manage that long tail of models securely at scale.

Read more: Big Technology Podcast