NVIDIA at GTC Berlin Unveils Vera Rubin, DSX Pod and Open Shell

On October 21, the NVIDIA GTC Berlin keynote took place, where Jensen Huang unveiled Vera Rubin, DSX Pod, Open Shell, the Agile 1S humanoid robot reference platform, and more.

To answer this question, GTC Berlin needs to be placed back into the broader context of global AI competition, Europe's sovereign AI agenda, and the regionalization of supply chains. For industry observers, this set of solutions offers a reference point for judging where the next round of competition will center: competition is shifting from a single computing-power metric to a combined contest of deployment speed, security governance, and supply chain resilience.

The core message NVIDIA conveyed at GTC Berlin can be summed up in one sentence: AI factories are turning from "custom engineering" into "standard products," and Europe is the first complete testing ground for this productization logic.

A $10 billion European investment, a Vera Rubin European supply chain, the DSX Pod containerized AI factory, the Open Shell agent security runtime, and the Agile 1S open-source humanoid robot—viewed individually, these announcements are product news; taken together, they form an infrastructure solution that is "vertically optimized, horizontally open."

This is not a single-point technological lead, but rather breaking the entire lifecycle of infrastructure into modules and then building a reference implementation for each module one by one.

Whoever defines the modules defines the procurement list.

The most impactful real-world example is the DSX Pod pod computing container.

According to NVIDIA, this is a fully assembled, fully tested, fully pre-built, and fully validated containerized AI factory unit. Traditional data center installation may take up to six months to complete rack setup, cabling, piping, validation, and connection to a traditional physical data center; DSX Pod can compress this cycle to 14 days.

The weight of this figure must be viewed within the real cost structure of AI factories: public information shows that AI factories have a power density of 50–100 kW per rack, five times that of standard hyperscale data centers; the construction cost per megawatt is as high as $25 million, compared with only $11.3 million for a standard data center.

AI factories are not a matter of "buying a few more GPUs," but a systems engineering challenge involving power, land, approvals, and construction.

By extracting the most standardized and replicable parts and turning them into container products, NVIDIA has effectively productized in advance the "delivery friction" that both customers and competitors must face.

DSX Pod targets rapid deployment scenarios at the hundred-megawatt level, while traditional gigawatt-scale data centers will still proceed in parallel—but NVIDIA has already used a standard container to pull AI factories from the era of "custom engineering" into the era of "productized delivery."

Also noteworthy are DSX's zero-water-use design and grid interaction capabilities.

The warm-water closed-loop cooling system sets the cooling temperature at 45 degrees Celsius, relying on dry coolers, fans, and radiators to complete heat dissipation, eliminating evaporative water consumption.

The DSX Flex automation API allows data centers to dynamically reduce or defer non-critical operations during grid peak periods, feeding power back to the grid.

This design has clear compliance value in Germany's regulatory environment of PUE 1.2 and a WUE cap of 0.4 L/kWh for new data centers.

The basic trade-off in cooling engineering is that waterless systems typically use more energy than evaporative cooling. NVIDIA's choice is to pay in electricity and capital expenditure under the European regulatory framework in exchange for meeting water-use targets—a move that aligns ahead of Europe's regulatory rhythm.

What is open-sourced is the runtime; what is locked in is the entry point

Agent security was the software-level focus at GTC Berlin.

At the event, NVIDIA introduced Open Shell—a policy-governed runtime based on the open-source Apache 2 license, deployed between agents and various resources within enterprise systems, covering files, credentials, tools, APIs, models, and even networks.

The core mechanism is a kernel-level isolation sandbox: agents run inside the sandbox, and the policy prover function verifies access boundaries before agent execution. Models, applications, and business data remain outside the sandbox, with Open Shell uniformly managing access permissions and tracking all communication interactions.

Open Shell's ecosystem adoption already has public signals: Dell, HPE, and SUSE AI Factory stacks have adopted Open Shell; HPE plans to integrate it into HPE Private Cloud AI in Q4 2026; Red Hat runs Open Shell and DOCA on Red Hat AI Factory; at launch, more than 100 organizations had already connected, including Anthropic, Cisco, CrowdStrike, Microsoft, SAP, Salesforce, and others.

Open Shell's structure deserves accurate description: open-source runtime + NVIDIA-defined reference architecture + BlueField-4 DPU as an independent security monitor.

When the AI factory stacks of Dell, HPE, and Red Hat all use Open Shell as the default security runtime, NVIDIA will have turned the reference implementation of security governance into the de facto entry point for industry procurement through an Apache 2.0-licensed piece of software.

This path has a precise analogy in infrastructure history: InfiniBand was originally an open industry standard, and NVIDIA's manufacturing capability, CUDA software integration, and proprietary extensions such as SHARP enabled it to evolve into a highly integrated technology stack.

Open Shell is taking the same path today: the more widespread the open-source runtime becomes, the deeper NVIDIA's involvement in the "default form of security infrastructure" grows.

This is not a negative judgment, but the key to understanding the business logic of open-source infrastructure: open standards and ecosystem entry points can coexist. Security governance is sinking from the model layer down to the infrastructure layer, and NVIDIA is using Open Shell to help define the technical implementation path of this lower layer.

While regulators are still discussing model alignment, NVIDIA has already turned "executable governance" into deployable components.

Europe's procurement list has more than one name; NVIDIA uses its supply chain to strengthen stickiness

Over the past week, NVIDIA announced the launch of a shared AI supercomputer (Sferical One) in Sweden, pledged $250 million in computing resources for UK AI safety research, and brought its cumulative investment and commitments in Europe's tech ecosystem to nearly $10 billion.

At the supply chain level, France's Bull is expanding production at its 20,000-square-meter facility in Angers, while Foxconn's Czech plant handles component production and testing before shipping to Angers for final assembly; European capacity from OEMs such as Dell supports GB200 rack supply, and Siemens is advancing an 800-volt DC power distribution solution for AI factories.

The forward-looking intent of this layout is clear: NVIDIA is upgrading Europe from a "sales market" to a "manufacturing and delivery hub," trading a regionalized supply chain for policy compliance and delivery certainty.

In September 2026, EuroHPC JU signed a €387.8 million LUMI-AI contract with Bull. This is Europe's sixth official AI factory, and its core GPUs are AMD Instinct MI430X.

The report explicitly points out that this is a landmark competitive outcome in European sovereign computing procurement, and the signal is that European buyers are willing to diversify risk away from the dominant GPU supplier.

NVIDIA's response precisely reflects its sophistication: the $10 billion European investment and supply chain layout serve two purposes at once—satisfying the political requirement of "made in Europe" and ensuring that AI factories procured with European public funds pass, at the physical level, through as many NVIDIA-defined components and reference architectures as possible.

The EU has signed letters of intent with AMD, NVIDIA, and Qualcomm to ensure hardware supply. This diversified landscape is itself part of Europe's sovereign AI agenda, and NVIDIA has chosen to reinforce its position in this diversified landscape through manufacturing localization.

The final piece of the puzzle at GTC Berlin is embodied AI.

NVIDIA offers a figure: physical AI covers 70% to 80% of global GDP related to the physical world, corresponding to a global market opportunity of more than $50 trillion.

In the first half of 2026, NVIDIA's physical AI business revenue was close to $11 billion, 30% higher than the full previous year.

GTC Berlin announced four items related to embodied AI: Siemens Healthineers using edge AI to reshape personalized radiotherapy; European automakers such as Volvo and Volkswagen showcasing L2-to-L4 autonomous driving solutions based on NVIDIA's full-stack platform; the Halo Safe complete safety system open test suite; and, in partnership with Munich-based Agile, the launch of the open-source humanoid robot reference platform Agile 1S, with the full software stack open to all researchers.

Now, with end-to-end models capable of reasoning and simulation as well as agentic AI technology, the relevant technology has "undoubtedly reached a tipping point." Siemens has fully introduced digital twin technology into its 55-year-old factory and, based on agentic AI, deployed autonomous inspection AI agents that formulate and execute plans daily, alongside general-purpose robots.

What is open-sourced is the reference platform; what is accumulated is scenario data and the developer ecosystem—this is NVIDIA's opening move to replicate the CUDA path in the physical AI field.

Three checkpoints, one industry and its signals

After October 21, three indicators are worth tracking: whether DSX Pod forms verifiable repeat orders at the hundred-megawatt level in Europe; whether IREN's flagship Sweetwater site moves from "announced" to "operational" on schedule; whether Open Shell enters actual production environments in at least three European regulated industries (finance, healthcare, energy); and, in the enforcement of Germany's EnEfG PUE 1.2 requirement, whether DSX warm-water cooling can pass third-party audits under real loads.

These three checkpoints test, respectively, the quality of "productized delivery," "security governance as infrastructure," and "European compliance adaptation."

Viewed against the broader backdrop of global AI competition, NVIDIA's intent at GTC Berlin can be summarized in three layers: regionalized hedging of the supply chain, moving from "selling chips" to "participating in defining standards," and turning "security" and "governance" into deployable infrastructure components before regulation becomes more detailed.

For the industry, the real signal of this launch event is not the specifications of any single product, but the shift in the dimension of competition: the contest in AI infrastructure is moving from "scale of computing power" to a multidimensional contest of "deployment speed, security governance, and supply chain resilience."

This article is from the WeChat public account "Tech Cloud Report" (ID: ITCloud-BD), author: Tech Cloud Report, published by 36Kr with authorization.