The AI Factory Comes in a Box: What Vera Rubin and DSX POD Actually Change

By Oluwaseyi Ayodeji, Published on oluwaseyiayodeji.com | Regal Stack Newsletter


At GTC Berlin on October 21, Jensen Huang unveiled something that deserves more attention than it received: a steel container holding seven fully cabled racks of NVIDIA’s newest computers, integrated and tested in a factory before shipment.

The internet, naturally, was busy with the robot ducks.

But that container caught my attention. Not because it was a particularly attractive piece of hardware, but because of what it represented. NVIDIA was showing an approach to building AI infrastructure that takes a familiar manufacturing principle and applies it at an entirely different scale.

I started my career in semiconductor manufacturing, where a chip’s value is measured in yield, quality and production cycle time. Later, I worked on equipment acceptance testing for server assembly and on qualifying contract manufacturers responsible for integrating server racks before they ship. Across more than 17 years in technology, manufacturing and infrastructure, one lesson has stayed with me: work that can be completed and tested in a controlled factory is often better done there than on a busy installation site.

NVIDIA’s new DSX POD brings that principle to AI infrastructure. Vera Rubin, the computing platform inside it, provides the horsepower. Together, they point to a different way of getting AI factories from a construction site to production.

To understand why that matters, it helps to look at what is changing beneath the AI applications everyone can see.

Understanding the AI infrastructure stack

Jensen Huang has previously used a five-layer cake analogy to explain the AI industry. It is a useful way to understand how the pieces fit together, even if a multibillion-dollar infrastructure project is not usually what comes to mind when someone mentions dessert.

Working from the bottom up, the layers are:

  1. Energy: The electricity required to power AI systems.

  2. Chips: GPUs, CPUs and networking processors that perform and coordinate the computing.

  3. Infrastructure: The data centers where those components are installed, connected, powered and cooled.

  4. Models: The AI systems that use the computing capacity.

  5. Applications: The products and services people interact with.

Most public conversation focuses on the top two layers: models and applications. But some of the most difficult constraints sit much further down.

The International Energy Agency estimates that global data center electricity consumption reached 485 terawatt-hours in 2025 and could roughly double to 950 terawatt-hours by 2030. AI-focused data center consumption is expected to grow even faster. That is a substantial demand increase for electricity systems already dealing with grid constraints, equipment shortages and lengthy connection timelines.

There is another constraint that receives less attention: the time between completing a data center and getting its computing systems into production.

During the Berlin keynote, Huang illustrated the economics with a hypothetical one-gigawatt AI factory costing around $50 billion and generating approximately $40 billion in annual revenue. In that scenario, a three-month delay could mean roughly $10 billion in deferred revenue.

Those figures are illustrative, not a guaranteed business case. But the underlying point is difficult to ignore. At this scale, time is not just a project-management metric. It is an economic variable.

Vera Rubin and DSX POD address different parts of this challenge. One advances the computing platform. The other changes how that platform is assembled and deployed.

Vera Rubin: when the rack becomes the computer

Vera Rubin is NVIDIA’s next-generation computing platform, named after the astronomer whose work provided important evidence for the existence of dark matter. It combines the Rubin GPU, which performs the AI computation, with the Vera CPU, which handles general-purpose processing and coordinates workloads.

The important development is not simply that the individual chips are faster. It is that NVIDIA continues to design the system around the rack as a unified computing unit.

The progression helps explain why.

Hopper, best known through the H100, brought powerful GPUs into the mainstream of accelerated computing. Blackwell pushed the architecture further by connecting 72 GPUs through NVIDIA’s NVLink technology, allowing them to operate as a tightly integrated rack-scale system.

Vera Rubin keeps that 72-GPU configuration but significantly upgrades the components and the connections between them.

A Vera Rubin NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs. Each Rubin GPU has 288GB of HBM4 memory and delivers up to 50 petaflops of NVFP4 inference performance under NVIDIA’s published specifications. Across the rack, NVIDIA specifies up to 3,600 petaflops of NVFP4 inference performance.

NVIDIA also claims that Vera Rubin can reduce inference cost per token by up to ten times compared with Blackwell. That is an important claim, but it should be understood as a vendor-reported comparison, not a universal outcome. Actual economics will depend on the workload, model, utilization, power costs and system configuration.

One of the less glamorous changes is just as interesting. NVIDIA has redesigned the system to reduce the complexity of physical assembly. If you have ever tried to untangle a box of Christmas lights, you have a reasonable starting point for understanding why fewer connections matter.

In a high-performance computing rack, every cable, hose and connection is another opportunity for a fault, a delay or a maintenance problem. Simplifying those connections makes assembly and servicing easier. Anyone who has worked on a manufacturing line understands the value of that.

The platform is also designed with AI agents in mind. Huang described the architecture this way: agents think on Rubin, act on Vera and are governed on BlueField. In practical terms, the GPU runs the model, the CPU helps execute the agent’s tasks, and BlueField provides infrastructure-level networking and security capabilities.

The broader message is that AI infrastructure is evolving beyond a collection of powerful chips. It is becoming a coordinated system designed to train models, serve inference workloads and support increasingly autonomous software.

DSX POD: building the factory inside a factory

Here is where the manufacturing connection becomes especially interesting.

Traditionally, data center equipment arrives at a prepared facility in separate racks and components. The site team then has to position the equipment, connect the racks, install and connect cooling systems, run network cabling, verify power and test the cluster.

That work is complex even when every supplier delivers on time.

In infrastructure supply chains, there is a concept sometimes called a square set: having every component, in the right quantity, at the right place and at the right time. If one critical part is missing, the entire installation sequence can stall. Skilled technicians end up waiting for a pallet that is still somewhere on a truck.

Multiply that across hundreds of deliveries and the logistics become a project of their own.

DSX POD changes the point at which much of that integration takes place. Rather than shipping individual racks to be connected and validated at the data center, NVIDIA’s reference design brings the computing and supporting infrastructure together in modular containers. Partners complete much of the integration and testing before the system leaves the factory.

A DSX POD configuration described at GTC includes seven Vera Rubin NVL72 racks, representing 504 GPUs. The integrated design brings together millions of components, networking, 800-volt power distribution and closed-loop liquid cooling. It also involves substantial network cabling, with approximately 17 kilometers of cable cited for a POD configuration.

NVIDIA is working with a broad ecosystem of partners, including Dell Technologies, HPE, Lenovo, Foxconn, Schneider Electric and Vertiv, among others. Deliveries are planned for early 2027.

The important distinction is that the factory is no longer just shipping a collection of racks. It is shipping a much more integrated and validated computing system.

That is a meaningful change in the level of assembly.

Why moving work upstream matters

Andy Grove, Intel’s former CEO, made a point in High Output Management that remains relevant to modern manufacturing: detect and fix problems at the lowest-value stage of the production process.

His example was a bad egg. It is better to reject it when it arrives from the supplier than after it has been cooked, plated and served. By then, the cost of the mistake has increased.

The same principle applies to an AI cluster.

A faulty cable discovered during factory testing might cost a technician a few minutes to replace. Discovering the same fault after a cluster has been installed and commissioned can mean hours or days of troubleshooting, with hundreds of GPUs sitting idle.

I have seen the difference this makes firsthand. Building acceptance testing into server assembly equipment is not simply about checking a box before shipment. It is about finding problems while the system is still accessible, the right people are nearby and the cost of correction is relatively low.

DSX POD applies that logic at a much larger scale. It moves integration and validation upstream, into a controlled manufacturing environment, before the equipment reaches the data center.

NVIDIA says this approach can reduce deployment time at a prepared site from as much as six months to as little as 14 days. During the keynote, Huang also described reducing the period from the arrival of computing equipment to the production of tokens from four or five months to about a month.

Those are related but different claims. Neither means that an entire data center can be built in two weeks. The site still needs to be developed, powered, connected and prepared. The acceleration comes from reducing the time spent integrating and commissioning the computing infrastructure once those prerequisites are in place.

That distinction matters. It is also what makes the announcement interesting: NVIDIA is not claiming to have solved every data center construction challenge. It is targeting a particularly expensive part of the process.

What it is, and what it is not

There is a risk that the container imagery makes DSX POD sound like a standard product that a customer can order, plug in and immediately operate. That is not quite what NVIDIA has announced.

DSX POD is a reference design. NVIDIA provides the architecture and platform, while ecosystem partners build, integrate and deliver systems based on that design. The exact configuration, commercial terms and delivery arrangements will depend on the partners and the customer.

It also does not solve the electricity problem. A prepared site still needs sufficient power, cooling capacity, network connectivity and the necessary infrastructure approvals. In many markets, those requirements are the hardest part of the project.

Nor does it remove the need for skilled site teams. Equipment must still be transported, positioned, connected to site utilities and commissioned. Factory validation reduces the work required at the destination; it does not eliminate it.

And there is a strategic consideration: DSX POD is built around NVIDIA’s technology stack. That can simplify integration and improve performance, but it also deepens dependence on a single vendor’s hardware, networking and software ecosystem.

For customers, that is a trade-off worth understanding before making a major infrastructure commitment.

What this means for emerging economies

This is where I think the announcement deserves a wider conversation.

For countries such as Nigeria and Kenya, and for emerging technology hubs such as Indonesia and Vietnam, the opportunity to deploy AI infrastructure faster is attractive. But the value of a faster deployment model depends on what a country is ready to do with it.

A factory-integrated system could reduce the amount of complex assembly required at a destination. It could simplify the movement of equipment, reduce the number of separate components that need to be coordinated at the site and make deployment more predictable. For markets where experienced data center technicians and specialist contractors are in short supply, that could be useful.

But faster delivery does not automatically translate into local capability.

If the most complex integration, testing and validation work takes place overseas, the host country may receive a functioning system without developing much of the industrial expertise that made it possible. The economic value associated with those activities may also remain with suppliers and manufacturers elsewhere.

That is not an argument against modular infrastructure. It is an argument for being deliberate about the terms on which it is deployed.

Before committing to a project, governments and institutions should be asking several questions.

First, is the power infrastructure ready? A container full of advanced GPUs is not useful if the site cannot supply reliable electricity. Grid capacity, substations, transmission connections, backup systems and long-term energy costs should be considered before the equipment arrives. The experience of Johor in Malaysia is instructive. Wood Mackenzie estimates that data centers could account for around 40 percent of the state’s electricity consumption by 2035. The issue is increasingly not just how much electricity a region can generate, but whether the grid can deliver it to the right locations.

Second, what capabilities will remain locally? Countries should negotiate for local participation in installation, commissioning, maintenance and ongoing operations. That means building pathways for engineers, electricians, cooling specialists, network technicians and other skilled trades. AI infrastructure is not built by software engineers alone.

Third, what happens to local suppliers? There may be opportunities for domestic firms to provide construction, logistics, electrical services, cooling support, security and facilities operations. Procurement strategies should identify which parts of the supply chain can realistically be developed locally, rather than assuming that every component must be imported.

Finally, how much strategic flexibility does the country retain? A tightly integrated system can deliver real benefits, but policymakers and operators should understand the implications for vendor dependence, future expansion and the ability to work with other technology providers.

These questions matter because the objective should not be to count how many containers have arrived or how many GPUs have been installed. It should be to build durable national capability: reliable compute access, skilled people, stronger local suppliers and a greater share of the economic value generated by the infrastructure.

A country can become a customer of the AI economy without becoming a meaningful participant in it. The difference lies in the capabilities it builds around the technology.

The next bottleneck

Vera Rubin and DSX POD show how quickly the physical architecture of AI infrastructure is changing. The computing platform is becoming more integrated, and the process of deploying it is becoming more industrialized.

That is progress. But it also makes the next set of constraints harder to ignore.

As the time required to integrate and commission computing systems falls, power availability, grid connections, permitting, land, cooling and skilled labor become even more important to the overall schedule. The bottleneck does not disappear. It moves.

For infrastructure builders, the lesson is to treat the factory, the supply chain and the deployment site as parts of one operating system. For technology leaders, it is to understand that AI performance depends on much more than the model or the chip. And for emerging economies, it is to make sure that faster infrastructure deployment also creates lasting local capability.

The AI factory may increasingly arrive in a box. The harder question is whether the country receiving it has built the power systems, skills, institutions and supply chains to make the most of it.

That is the conversation I believe we should be having.


Oluwaseyi Ayodeji is an AI infrastructure Leader and the founder of Regal Stack, a think tank focused on the sovereign participation of emerging markets in the global AI economy. Learn more at oluwaseyiayodeji.com.


References and further reading

  1. NVIDIA, NVIDIA Introduces DSX POD AI Factories to Accelerate Europe’s Infrastructure Deployment, October 21, 2026.

  2. NVIDIA, Vera Rubin NVL72.

  3. International Energy Agency, Key Questions on Energy and AI: Executive Summary.

  4. NVIDIA, Inside the NVIDIA Vera Rubin Platform: Six New Chips, One AI Supercomputer.

  5. NVIDIA, NVIDIA DGX Vera Rubin NVL72 Datasheet.

  6. NVIDIA, GTC Berlin 2026 keynote and DSX POD announcement.

  7. Data Center Dynamics, reporting on NVIDIA’s DSX POD architecture and deployment partners.

  8. Wood Mackenzie, Johor’s Data Centres Could Consume 40% of Electricity Demand by 2035, June 18, 2026.

  9. Andy Grove, High Output Management, 1983.


Previous
Previous

The AI Productivity Paradox: We're Optimizing Tasks. Who's Redesigning the Work?

Next
Next

The Gap Nobody Is Zooming Out To See