NVIDIA Packs an AI Factory Into a Container: Onsite Build Drops From Six Months to 14 Days, and Agents Get a Kernel-Level Sandbox
Synopsis
At GTC Berlin, NVIDIA introduced DSX POD, a reference design for complete Vera Rubin AI factories that partners integrate and validate inside modular containers before delivery, cutting onsite deployment from six months to 14 days once site utilities are ready. A companion OpenShell reference architecture places a policy-controlled runtime with kernel-level sandboxing between an agent and the systems it can affect, since each launched agent carries different privileges.
Interpretation
The delivery form of an AI factory shifts from onsite construction to modular containers pre-integrated and validated by partners in a factory. Data centers previously had to be installed and built onsite, a practice Huang said this change will alter. NVIDIA's stated comparison is a drop from six months to 14 days for onsite deployment once site utilities are ready.
An agent is broken into five parts: a model, a harness that orchestrates its work and memory, a sandbox that isolates the harness, knowledge and tools, so it is more than a model answering questions. Under this framing an agent can draw on knowledge and act through tools rather than only producing answers. Huang stressed that agents can reach files, resources, networks and tools, so privileges should be minimal and monitored, and differ on every launch.
The OpenShell reference architecture puts policy enforcement, privileges and monitoring between the agent and external systems, adapting on each launch. Security is treated as an engineering problem rather than a property of model output. The architecture builds on the open source, policy-controlled OpenShell runtime with kernel-level sandboxing.
Perspective
The report covers what was announced in the Berlin GTC keynote: the DSX POD reference design, the OpenShell reference architecture, the Agent Knowledge Foundation, CUDA-X tooling and the Agile ONE SN humanoid reference platform. It is aimed at engineers working on AI infrastructure, agent deployment and robotics research platforms. The keynote states DSX POD can shorten deployment to 14 days once site utilities are ready, a precondition to read alongside specific site conditions.
Figures such as up to 100% less cooling water and up to 30% less power for the same AI output are stated as upper bounds without measurement conditions. Agile ONE SN, a 31-degree-of-freedom body expected next year, and the reduction of researcher setup from weeks to days are expectations rather than delivered results. How OpenShell's dynamic privilege adjustment maps to specific enterprise policies is described only in principle.