Nvidia's AI factory pitch is also a tenancy problem. The Vera Rubin release includes confidential computing, attestation, BlueField-4 isolation and zero-trust policy enforcement, not only faster racks. [1]

That detail belongs beside the paper's account of Vera Rubin entering full production and its earlier note that photonics moved AI infrastructure toward uptime. Scale makes the security question sharper. Shared AI factories are not just machines waiting for customers. They are places where proprietary prompts, model weights, outputs and telemetry may share infrastructure.

The company names security features in the same official release that names adopters and fall shipment expectations. [1] That does not prove how each customer will deploy them. It does show that Nvidia understands the buyer's fear.

X celebrates rack scale because scale photographs well. Enterprises buy only if isolation can be inspected. The next receipt is not another benchmark. It is whether customers can verify attestation details and policy enforcement before they put real work inside the factory.

-- DAVID CHEN, Beijing