NVIDIA GTC Berlin 2026: what builders can use, test and wait for

What NVIDIA announced in Berlin

NVIDIA used GTC Berlin to connect three parts of its platform: systems that produce model inference, controls for agents that call tools, and smaller computers for local development and use. That connection matters because an agent deployment depends on more than the model. Teams also need compute, a harness, tool permissions, monitoring and a way to stop unsafe actions.

NVIDIA's keynote report [1] lists DSX POD, OpenShell, Sentry and local systems among the announcements. It is NVIDIA's account of its own event, so product claims below remain vendor claims unless another source or a buyer's test confirms them.

DSX POD moves AI-factory assembly away from the customer site

DSX POD is a reference design for Vera Rubin systems assembled inside modular containers. Partners complete cabling, power distribution, liquid cooling and validation before shipment. A customer still needs a prepared site with enough power, cooling, networking and other supporting infrastructure.

NVIDIA says [2] a prepared site can move from a six-month installation to as little as 14 days. That number describes onsite deployment after site utilities are ready. It does not include permitting, grid connection, procurement, model integration or production acceptance.

The first question for a buyer is therefore not whether a container arrives quickly. Ask when usable capacity begins serving an accepted workload, what dependencies sit outside the quoted schedule, and who owns failures across the system and facility partners.

OpenShell and Sentry put agent controls below the model

The OpenShell announcement [3] describes a full-stack reference architecture. OpenShell provides the secure runtime, Vera CPUs run harness and tool work, and BlueField-4 DPUs run Sentry outside the host.

NVIDIA's Sentry technical article [4] says Sentry can inspect access requests, correlate network activity with runtime state, preserve tamper-evident records and enforce policy from a hardware-isolated layer. This design can make it harder for a compromised host or agent to disable its own monitor.

Hardware isolation does not decide whether a business action is appropriate. Teams still need identities, narrow permissions, approval rules, data classification, incident response and tests for prompt injection, confused-deputy behavior and valid but harmful tool calls.

Local AI now spans PCs, developer boxes and workstations

NVIDIA's local systems report [5] positions RTX Spark for Windows personal agents, DGX Spark for Linux development, and DGX Station for larger enterprise work. NVIDIA also announced new DGX Spark configurations, multi-system Sync support and a Windows version of DGX Station planned for December.

Local execution can help with latency, data control and disconnected work, but local does not automatically mean private or secure. Check telemetry, updates, downloaded models, connector traffic, disk encryption and whether an agent can reach cloud services before making a privacy claim.

Choose hardware from the measured model, context length, quantization, concurrency and latency target. Peak compute and maximum memory alone do not predict whether an interactive agent will meet its service target.

How builders should evaluate the announcements

• Write one production workload with an acceptance rate, latency target, concurrency level and cost ceiling.
• List every date and dependency, including preview status, partner availability, site readiness and software support.
• Threat-model the agent host, DPU, model endpoint, tools, identity provider and human approval path separately.
• Test whether policy still applies after a host compromise, malformed tool request, prompt injection and network failure.
• Measure model load time, tokens per second, end-to-end latency, power draw and recovery time on the exact model configuration.
• Require logs that connect a user request, agent decision, tool call, policy result and final side effect.
• Keep a smaller baseline so you can quantify what the NVIDIA stack adds in cost, security and operations work.

What remains uncertain

The published material does not provide independent deployment-time measurements, production security evaluations or total ownership costs. Partner participation shows ecosystem interest, but it does not prove that each integration has reached the same release stage.

DSX POD systems are planned through partners in early 2027. OpenShell describes an architecture that spans several hardware and software layers, so availability and feature coverage may differ by operating system, partner and deployment size.

Sentry can enforce technical policy from an isolated layer, but organizations must still define the correct policy and handle actions that are permitted technically but wrong in context.

Frequently asked questions

• Did NVIDIA release one product called the GTC Berlin stack? No. The event covered separate systems, software and reference architectures with different schedules.
• Is DSX POD available now? NVIDIA said partners plan to deliver modular AI factories worldwide in early 2027.
• Does the 14-day claim include site construction? No. NVIDIA applies it to deployment at a prepared site with supporting utilities ready.
• Can Sentry replace application authorization? No. It can enforce policy outside the host, but teams must still define identities, permissions and valid business actions.
• Does local AI keep all data offline? Only if the complete application, model, connectors, updates and telemetry are configured that way.