Analyst(s): Nick Patience
Publication Date: October 22, 2026
At NVIDIA GTC Berlin, Jensen Huang introduced DSX POD, a containerized reference design for Vera Rubin AI factories that partners build and test before delivery, alongside OpenShell for agent security and a slate of physical AI partners. His message to Europe was that it can produce AI rather than simply consume it. Futurum believes the stack is strong, although the constraint that matters most in Europe, grid power, remains a constraint.
What Is Covered in This Article:
- NVIDIA introduced DSX POD, a reference design for containerized Vera Rubin AI factories that 17 partners will deliver from early 2027, cutting onsite deployment from six months to 14 days once site utilities are ready.
- DSX POD and the wider DSX platform improve deployment speed and power efficiency but do not address access to grid power, the binding constraint on AI factories in Europe.
- OpenShell and Sentry extend NVIDIA’s stack from silicon to the agent runtime, positioning the company as a security layer for enterprise agents.
- NVIDIA’s open-model case for European sovereignty rests on Nemotron, which 6.7% of respondents in Futurum’s latest AI Product Series survey say their company currently uses.
- Physical AI announcements spanned new Hyperion vehicle partners, the Agile ONE reference humanoid, AI-RAN, and the RTX Spark PC line.
The News: At NVIDIA GTC Berlin this week, NVIDIA founder and CEO Jensen Huang used his keynote to set out how Europe can build what he called an AI economy, with NVIDIA’s full stack as its foundation. The headline announcement was NVIDIA DSX POD, a reference design for complete Vera Rubin AI factories that partners build inside modular containers, integrate, and validate before delivery. Each pod holds seven Vera Rubin NVL72 racks, or 504 GPUs, connected by Spectrum-X networking, with liquid cooling, dry coolers, and 800-volt DC power delivery. NVIDIA says onsite deployment can shrink from six months to 14 days once site utilities are ready. Seventeen partners, including Bull, Dell Technologies, Delta, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology (QCT), Schneider Electric, Vertiv, and Wistron, will deliver DSX PODs worldwide from early 2027. NVIDIA also extended its DSX AI factory platform with power controls that it says produce the same output with up to 30% less power, and cooling designs that it says can cut cooling water use by up to 100%.
Beyond infrastructure, NVIDIA presented OpenShell, an open-source (Apache 2.0) secure runtime that sandboxes AI agents and enforces policies on file, network, and system-call access, plus Sentry, an optional monitor running on BlueField-4 that can quarantine a misbehaving agent in milliseconds. Huang announced that Anthropic and NVIDIA are partnering on Claude agents that run in OpenShell. NVIDIA also added a Volkswagen van, a new JLR model, Volvo, and next-generation Mercedes-Benz vehicles to its Hyperion autonomous driving platform, previewed the Agile ONE reference humanoid robot built with Agile Robots for early 2027, and showed RTX Spark, a new line of Windows PCs built on its N1X chip with Microsoft. Huang said NVIDIA invested nearly $10 billion in Europe over the past year and now has more than 12,000 partnerships across the region.
At GTC Berlin, NVIDIA Puts the AI Factory in an Actual Container
Analyst Take: NVIDIA GTC Berlin confirmed that NVIDIA now sells something closer to a finished AI factory than a set of components, and its pitch to Europe is that buying that factory is the fastest route to an AI economy of its own. The engineering is impressive, and DSX POD in particular addresses a real cost, the months that expensive GPUs sit idle while racks are cabled and commissioned on site. However, the binding constraint on AI factories in Europe is access to grid power, and DSX POD doesn’t address that.
DSX POD Turns the AI Factory Into a Supply Chain Product
Huang’s economics were blunt. A one-gigawatt AI factory costs about $50 billion and, by his estimate, should generate around $40 billion a year in revenue, so a delay of a single quarter costs roughly $10 billion. DSX POD extends the logic of NVL72, which moved rack assembly from the data center into the supply chain, to the scale of a multi-rack pod. Containers are cabled, burnt in, and tested at the factory and then shipped to site, which reduces both onsite labor and the idle capital that waits for commissioning.
The partner list shows where that work will go. Alongside Dell, HPE, Lenovo, and Bull, the builders include Taiwan’s Foxconn, Pegatron, QCT, Wistron, and GIGABYTE, with Delta supplying power systems. More of the value of each AI factory shifts to these manufacturers, but so does more of the integration risk, since a pod that fails acceptance testing becomes the supplier’s problem rather than the customer’s. For buyers, the appeal is predictability, because a standard unit with a known delivery schedule is easier to plan and finance than a bespoke build.
Efficiency Is Not the Same as Power at NVIDIA GTC Berlin
The weakness in the DSX POD story lies in NVIDIA’s own qualification that the 14-day deployment applies once site utilities are ready. In Europe’s main data center markets, including Frankfurt, London, Dublin, and Amsterdam, operators have faced grid-connection restrictions and multi-year queues, and that wait dwarfs the six months DSX POD removes. Huang places energy at the base of his five-layer cake of energy, chips, AI factories, models and applications, a layer to be addressed by partners such as Schneider Electric and Siemens Energy.
NVIDIA’s counterargument is that when a site’s power allocation is fixed, efficiency becomes capacity, and NVIDIA says DSX MaxLPS delivers the same output with up to 30% less power, while DSX Flex adjusts load to grid conditions, a capability that could make AI factories easier for grid operators to accept. Closed-loop liquid cooling paired with dry coolers also removes one of the most common local objections to new facilities.
OpenShell Makes Agent Security an Infrastructure Problem
The most strategically important announcement at NVIDIA GTC Berlin may prove to be OpenShell, which was previewed at GTC San Jose. Huang described an agent as five components: a model, a harness, a secure sandbox, knowledge, and tools, and NVIDIA now offers its own version of each. OpenShell enforces policy outside the agent, at the level of system calls and network connections, rather than relying on guardrails inside the model, and Sentry adds a hardware-isolated monitor on BlueField-4. We think that is the right design principle, because security that depends on an agent obeying its instructions offers no real protection. Huang said frontier labs, operating system vendors, and cybersecurity companies have adopted OpenShell since its launch, and the Anthropic partnership gives it a prominent early user.
There is a competitive point here as well. By open-sourcing the agent runtime while placing the strongest enforcement on BlueField hardware, NVIDIA follows a familiar pattern of an open software layer that works best on its own silicon. Enterprises should welcome the openness, although they should expect hyperscalers and security vendors to offer their own runtimes, and the question for 2027 is whether OpenShell becomes a standard or one option among several.
Sovereign Intelligence Rests on Models Few Enterprises Use
Huang’s message to Europe at NVIDIA GTC Berlin was that it should produce AI rather than consume it, and NVIDIA’s offer is credible in parts. Its open Nemotron models ship with weights, training data, and recipes, which enables companies – European or otherwise – to verify provenance and post-train models on their own data. Futurum data shows how far that is from mainstream practice. In Futurum’s AI Product Series survey of September 2026, 6.7% of respondents said their company currently uses Nemotron for AI application development, compared with 74.2% for Anthropic Claude and 67.5% for OpenAI GPT (n=511). Nemotron’s pipeline is healthier, with 18.0% planning to evaluate or adopt it, up from 13.8% in March 2026, and current use has edged up from 5.6%.
Futurum’s view is that sovereignty is ultimately a question of jurisdictional reach across every layer of the AI supply chain. NVIDIA’s pitch gives Europe control over the upper layers, its data, models, and applications, while the compute layer remains designed in the US, manufactured largely in Taiwan, and subject to US export controls. Combined with the power constraint at the bottom of the stack, that leaves Europe with more control over its AI than it has today but well short of independence, not that NVIDIA is promising that. European policymakers should welcome NVIDIA’s investment while recognizing that it does not, by itself, amount to a sovereignty strategy.
Physical AI Extends the Same Playbook
The physical AI announcements apply NVIDIA’s standard-platform approach to new markets. Hyperion now spans Volkswagen, JLR, Volvo, and next-generation Mercedes-Benz vehicles, with NVIDIA betting that robotaxis from many brands will feed shared ride-hailing networks. The Agile ONE reference humanoid, built with Munich-based Agile Robots, aims to do for robotics what NVL72 did for data centers by giving developers a pre-integrated starting point, although humanoid demand remains largely prospective. NVIDIA’s first commercial call over AI-RAN on the T-Mobile network with Nokia moves its telecom ambitions from roadmap to deployment. RTX Spark, built on the N1X chip after four years of work with Microsoft, brings one petaflop and 128 GB of memory to a laptop, the same compute and memory as the first DGX-1 NVIDIA delivered to OpenAI in 2016. The consistent theme is that NVIDIA wants to define the reference configuration in every market it enters, so that partners build on its platform rather than around it.
What to Watch:
- Named DSX POD customers and first deliveries in early 2027, and which of the 17 partners, particularly the Taiwanese manufacturers, will win the largest share of integration work.
- Whether European governments and grid operators speed up connections for flexible-load AI factories, the factor most likely to determine how quickly DSX PODs are switched on in Europe.
- OpenShell adoption beyond launch partners, and whether hyperscalers back it or ship competing agent runtimes.
- Whether Nemotron’s evaluation interest converts into production use in Futurum’s next AI Product Series survey.
- Shipments of RTX Spark PCs and DGX Station for Windows, and developer uptake of local agent workloads.
See the complete press release on DSX POD on the NVIDIA Newsroom.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Other Insights From Futurum:
NVIDIA Cosmos 3 and Open Agent Tools: Is Physical AI About to Leave the Lab?
Nokia’s AI-RAN Platform Puts a Number on the Software-Defined RAN Bet
At GTC 2026, NVIDIA Stakes Its Claim on Autonomous Agent Infrastructure
Author Information
Nick Patience is VP and Practice Lead for AI Platforms at The Futurum Group. Nick is a thought leader on AI development, deployment, and adoption - an area he has researched for 25 years. Before Futurum, Nick was a Managing Analyst with S&P Global Market Intelligence, responsible for 451 Research’s coverage of Data, AI, Analytics, Information Security, and Risk. Nick became part of S&P Global through its 2019 acquisition of 451 Research, a pioneering analyst firm that Nick co-founded in 1999. He is a sought-after speaker and advisor, known for his expertise in the drivers of AI adoption, industry use cases, and the infrastructure behind its development and deployment. Nick also spent three years as a product marketing lead at Recommind (now part of OpenText), a machine learning-driven eDiscovery software company. Nick is based in London.
