Nvidia's Berlin Blitz: 14-Day AI Factories, a Safe-Agent Runtime and a $279 Billion Supply Bet
NVIDIA used its GTC Berlin keynote to launch DSX POD, a containerised Vera Rubin AI factory design aimed at roughly 14-day site deployments, alongside OpenShell, a security reference architecture for AI agents. The same week's quarterly filing revealed supply and capacity commitments had more than doubled to $279 billion, driven mainly by memory procurement.
A keynote built for Europe's build-out
NVIDIA's GTC Berlin ran October 20–22 at STATION-Berlin, and the main event was Jensen Huang's keynote at the Tempodrom on October 21, a two-hour session starting at 11 a.m. CEST. The framing was explicitly regional: Europe building its own intelligence, with local control over the infrastructure that produces it. The company's newsroom moved fast — eight separate press releases landed inside a 59-minute window, between 05:34 and 06:33 ET, covering data centres, suppliers, agent safety and robotics.
DSX POD: AI factories that arrive pre-assembled
The headline hardware announcement was DSX POD, a reference design for complete Vera Rubin AI factories that partners build inside modular containers. The units are integrated and validated before they ship, so they arrive ready to connect out of the box to a prepared site's power, cooling, networking and other infrastructure. NVIDIA positions this as a way to collapse construction schedules — the target is roughly 14 days from delivery to a running system, with first deliveries planned for early 2027. It is a supply-chain play as much as an engineering one: rather than shipping components to a muddy building site, the factory moves to the container.
OpenShell: putting guardrails outside the model
On the software side, NVIDIA launched a reference architecture for OpenShell, its open runtime for AI agents, combining Vera Rubin, Vera CPUs and BlueField-4 with in-silicon security enforcement. The design pairs OpenShell, the runtime, with Sentry, an out-of-band control layer. The architectural point is deliberate: controls sit outside the model, so an agent cannot simply talk its way past its own restrictions while reaching for external tools and systems. "AI capability and safety must advance together," Huang said. OpenShell and related platform software are available through NVIDIA's developer resources and on GitHub.
The $279 billion number behind the slides
The most consequential figure of the week was not printed on a product slide. In its quarterly filing for the second quarter of fiscal 2027, which ended July 26, 2026, NVIDIA disclosed that supply and capacity commitments had risen from $119 billion to $279 billion — more than doubling in a single quarter, driven primarily by memory procurement. The company also pointed to a 300% increase in supplier capacity, four times its starting level. Vera Rubin is already in full production at Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure, on the back of a quarter that delivered $96 billion in revenue with a 117% data-centre surge.
What it means for buyers
Read together, the Berlin announcements describe a company optimising for speed to power rather than peak benchmark numbers. Two-week deployments, containerised factories and a runtime that assumes billions of agents will eventually operate semi-autonomously all point the same direction: the binding constraint on AI in 2027 is not model quality but whether the physical and governance plumbing can keep up. For European enterprises and the governments courting them, that is the offer being made — and the $279 billion is the bet that it gets bought.