Lead

VAST Data has set a Q1 2027 general-availability date for DataEnclave, a confidential-computing layer that runs frontier AI models against sensitive enterprise data without exposing either side's proprietary assets. The software enters preview today and will ship through VAST and OEMs including Cisco and Supermicro.

What deadlock does it try to break?

The product targets a structural split in the AI market. Enterprises will not hand sensitive records to large language models running in public clouds or neocloud providers. Model builders will not let proprietary weights execute inside customer data centres they cannot control.

DataEnclave aims to remove both objections with one runtime. The two sides of the market do not yet agree on a common trust model.

How does the protection work?

Execution happens inside confidential virtual machines built on CPU-level trusted execution environments, with Nvidia GPUs operating in Confidential Compute mode. VAST layers three protections on top:

• Hardware-isolated encryption of guest memory, GPU memory and NVLink traffic
• Cryptographic attestation that releases model weights only after the runtime environment passes verification
• Independent key control, with each party holding its own keys through Bring Your Own KMS integrations

The attestation chain runs on the open-source CNCF Trustee project, with Fortanix supplying on-premises key brokering for air-gapped deployments. VAST puts the runtime overhead at single-digit to low-double-digit percent of GPU time, depending on model size. The performance tax is real, but smaller than the cost of turning sensitive workloads away.

Who has signed up?

Renen Hallak, VAST co-founder and CEO, framed the architecture as an operating-system problem. "Models are becoming a resource the operating system has to manage, the same way it manages data," he said. "That means knowing which model fits which task, what it can see, who can use it and under what rules."

The committed model roster covers Cohere, CrowdStrike, Deepgram, Factory, Fundamental, Nvidia and TwelveLabs. The big frontier labs are not yet on the list.

Server partners at launch are Cisco and Supermicro. AI cloud partners are limited to Buzz, Nscale and Sharon.AI. The launch roster is broad on servers, narrow on model builders, and thin on AI clouds.

Justin Boitano, Nvidia's vice president of enterprise AI, tied the work to the broader agent market. "Enterprise data is essential to accurate, usable AI – and keeping business data confidential is critical to protecting IP in the age of agents," he said. "VAST Data's integration of Nvidia Confidential Computing delivers protection for both enterprises and model builders."

Where does the standard sit?

VAST is aligning the runtime with three open anchors: the Kata container interface for confidential containers, the CNCF Trustee attestation project, and Nvidia's confidential-computing GPU mode.

John Mao, VAST's vice president of business development and alliances, said the company is taking a "multi-pronged approach" to standards support and will follow the market as it converges. Convergence, not fragmentation, is the stated goal.

What else rides on the same runtime?

The same DataEngine secure runtime underpins VAST AgentEngine, a sandboxed execution environment for autonomous AI agents. Agents inherit the same hardware isolation, key separation and audit trail that DataEnclave applies to model weights and enterprise data. An agent receives narrow, attested access to a model and a dataset without exposing either to the agent's operator.

Model isolation and agent isolation share the same mechanism.

Rack-scale systems from Dell, HPE, Lenovo, Supermicro and a long list of Taiwanese ODMs already ship with Vera Rubin NVL72 configurations that combine Nvidia GPUs, Vera CPUs, NVLink encryption and attestation at the rack level. Those platforms give DataEnclave a ready hardware base, though pricing for the software has not been disclosed.

What happens next?

VAST's commercial thesis, articulated by John Mao, is that bringing leading models to sensitive enterprise data opens roughly half of the AI market that public-cloud deployments cannot reach. The expectation is that on-premise and air-gapped buyers will pay a premium for confidentiality once general availability arrives in Q1 2027.