Key Insights
- DSX POD modularizes the AI factory. Each pod packs roughly 1.2 Megawatts (MW) across seven liquid-cooled racks, integrated and tested before shipping. NVIDIA says on-site deployment falls from about 6 months to 14 days, with first solutions expected in production in early 2027.
- Speed to revenue outranks silicon in messaging. Vera Rubin was positioned as the engine inside a bigger product. The vendor-stated figures of about US$50 billion per 1 Gigawatts (GW) and close to US$40 billion in annual revenue make the point: turning capital into revenue-generating tokens quickly is the priority, and this is an Artificial Intelligence (AI) infrastructure play rather than silicon.
- Efficiency is the headline message. NVIDIA claims DSX LPS and DSX Flex can enable providers to deliver the same token output with up to 30% less grid power, and the closed-loop liquid cooling design uses no water.
- AI safety is an engineering & hardware play. OpenShell, the runtime that sandboxes agents, now pairs with Sentry on BlueField-4 Data Processing Units (DPUs), and NVIDIA is betting the combination pulls through DPU adoption.
DSX POD Unveiled to Modularize the AI Factory
One of the main headlines from this week has been DSX POD—a reference design for a complete Vera Rubin system that partners can build and package into modular containers. Each pod holds roughly 1.2 MW across seven liquid-cooled racks, is connected by 17 Kilometers (km) of Spectrum-X cables, and is integrated, cabled, and tested before being shipped. NVIDIA states that on-site deployment can be reduced from ~6 months down to just 14 days.
NVIDIA will not build or sell these pods directly to customers. Partners including HPE, Dell, Lenovo, Schneider Electric, and more are due to deliver them worldwide to customers, and the first DSX POD solutions are anticipated to be in production in early 2027.
Why Did Silicon Take a Back Seat?
Vera Rubin is now in full production, yet it was positioned in the keynote predominantly as just the engine inside a bigger product. NVIDIA Chief Executive Officer (CEO) Jensen Huang put a figure out there on stage that 1 GW of AI infrastructure costs roughly US$50 billion when accounting for Land, Power, Shell (LPS), cooling, and networking, and said that this could generate close to US$40 billion in annual revenue. While this is a vendor-stated figure, the message was clear—the speed at which capital can be turned into revenue-generating tokens is key.
The underlying bottleneck for the AI build-out is not simply access to compute for the companies building the AI factories worldwide, but also in how quickly they can stand up and activate these chips to monetize their investments. These players are procuring compute sometimes years in advance of when the facilities are completed and they must still then integrate, cable, and test these systems. A modular solution such as DSX POD means that—whether deployed in an existing or new facility—the hold-up to monetization should never be the AI system itself.
Efficiency for data centers also got the nod over silicon performance. NVIDIA claims that with DSX LPS and DSX Flex, AI factories can produce the same number of tokens while pulling up to 30% less power from the grid. Furthermore, the closed-loop liquid cooled design of the AI systems means that they have no water use due and allow for heat dissipation through air ducts.
This was clearly a key point of emphasis in messaging as the European and global political systems are placing more scrutiny on the AI build-out, so efficiency gains can compound at scale as more AI factories are built in the region.
Is the Grid the Real Bottleneck?
Yes, and DSX POD makes it harder to ignore. The 14-day installation figure by NVIDIA explicitly conditional on on-site utilities being ready and in Europe can be a problem.
Our own research points to this continuing to be an issue. In Europe, active AI Information Technology (IT) capacity for cloud service providers is set to scale from 1.6 GW in 2025 to over 16 GW in 2035, with growth more heavily skewed toward the end of the time frame as access to stable capacity remains a bottleneck in the region. While the Behind-the-Meter (BTM) power source trend may perhaps proliferate in Europe as in the United States, widespread adoption of this trend is unclear.
AI Safety
Safety has been a huge topic of discussion over the past few weeks. Back in September, NVIDIA announced the Open Agent Safety Platform with more than 100 partners, pairing its OpenShell secure runtime with NVIDIA Sentry, an out-of-band watchdog on its Bluefield-4 DPU. NVIDIA says that Sentry can quarantine misbehaving agents in milliseconds, and has now packaged both into the OpenShell Reference Architecture, where Vera Rubin can run the inference, Vera Central Processing Units (CPUs) run the agents, and Bluefield-4 enforces strict limits the agents cannot see or reach.
The Engineering Problem Behind AI Safety
NVIDIA has held the position of “AI safety is an engineering problem.” The above announcements showcase what it has been doing to solve the safety/security issues for a number of months now. As I argued back in March during GTC San Jose in “The Rise of the AI Claw,” the launch of OpenShell signaled a pivotal moment for Agentic AI.
Now, as OpenShell has matured and Sentry has been introduced, NVIDIA is very bullish on the future of not only this software, but on the adoption of its DPUs as well. One of the most interesting partners with NVIDIA on the use of OpenShell is none other than Anthropic itself, which is leveraging OpenShell for its Claude Managed Agents offering to enterprises.
For further information on ABI Research’s perspective on the recent AI slowdown and safety debate, please read our research report here.
Telco Industry Gets the Physical AI Tag
Jensen spent a lot more time highlighting NVIDIA’s progression in the telecoms industry than in previous years. During his keynote, Jensen claimed that radio antennas are Physical AI devices due to their capabilities with beamforming—a technology employed by Massive Multiple Input, Multiple Output (mMIMO) radios to optimize the spectral efficiency of antennas in a given cell site.
NVIDIA views these towers and other telecoms industry assets as distributed AI infrastructure, and announced how, in partnership with Nokia and T-Mobile, the first AI-Radio Access Network (RAN) call was completed on a commercial network. There has been significant advancement in AI-RAN technology since the formal partnership with Nokia was announced last year, and many European telco operators are now moving toward field trials of the technology, including BT, Elisa, Deutsche Telekom, and Orange.
Conclusion
NVIDIA remains bullish on its position in the market and used GTC Berlin as a call to action for the European market to enable their AI economy. What remains outside of the control of AI factory builds is access to the grid, which is why the bottleneck in a highly regulated market is hard to overcome. I anticipate that NVIDIA will be much more vocal and proactive next year in vocalizing exactly how the grid constraints and efficiency gains can be made by all grid operators in the region, rather than focusing on just what data center operators can do.