Chapter 7 · Data CentresGTC Berlin
NVIDIA DSX POD: prefabricated, containerised Vera Rubin AI factories
NVIDIA
A reference design for Vera Rubin AI factories that partners build inside modular containers, integrate and test in a factory, then ship to a site whose power, cooling and network are already prepared.
What was announced?
NVIDIA describes DSX POD as “a reference design for complete NVIDIA Vera Rubin AI factories built by partners inside modular containers, integrated and validated before delivery.” NVIDIA does not build the containers itself: seventeen partners are named, from server makers (Dell Technologies, HPE, Lenovo, Foxconn’s Ingrasys, GIGABYTE, QCT, Pegatron, Wistron) to power and cooling suppliers (Schneider Electric, Vertiv, Eaton, Delta) and AI cloud and defence companies (Crusoe, DayOne, Armada, Anduril, Bull).
The compute, networking, cabling, rack-level power and liquid cooling are installed in the factory. Each unit is meant to “arrive ready to connect” to a site that already has its power, cooling and network in place.
Why does it matter?
Most of the time in an AI data-centre build is not the servers, it is the sequence: the hall is finished, racks are set, cabled, filled with coolant and tested on site, one stage after another, by scarce specialist trades. Prefabrication moves the IT integration off site, so it runs in parallel with site preparation instead of after it.
NVIDIA says this cuts deployment “from six months to as little as 14 days”. That figure counts only the time from delivery to running, and only on a site that is already prepared: the substation, the power feed, the heat-rejection plant and the fibre still take as long as they always did.
What has changed?
| Aspect | Conventional on-site build | DSX POD (as NVIDIA describes it) |
|---|---|---|
| Where integration happens | In the data hall, after the building is ready | In a partner’s factory, while the site is prepared |
| Rack-to-rack cabling, manifolds, leak detection | Installed and tested on site | Installed in the container before shipping |
| Power interface | Site-specific distribution to each rack | A standard 800 V DC input, per NVIDIA’s technical blog |
| Time from delivery to running | Months of on-site work | “As little as 14 days”, NVIDIA’s claim |
| What the site still needs | Power, cooling, network, building | Power, cooling, network and a prepared pad — unchanged |
Modular and containerised data centres are not new; what is new is a GPU maker publishing a reference design for a whole Vera Rubin AI factory in that form.
How does it work?
A container (or set of containers) holds Vera Rubin NVL72 racks with their scale-out network, the cabling between racks, power and control wiring, sensors, liquid-cooling manifolds and leak detection. Because it is assembled and validated in a factory, faults found in testing are fixed there, not in a live hall.
On site, the unit is connected to three things the site must already provide: electrical power (NVIDIA’s technical blog cites a standard 800 V DC input design), a liquid-cooling supply and return that can reject the heat, and the network uplinks. NVIDIA publishes no rack count, power rating or cost per POD.
Figures and claims, and what backs them
- Manufacturer claimNVIDIA says DSX POD reduces deployment time “from six months to as little as 14 days”.
- Published specificationNVIDIA’s technical blog states a “standard input 800 VDC power design”.
Impact on AI infrastructure delivery
- Construction
- The construction scope shrinks to a prepared site: foundations or a hall, the electrical and cooling connections, and access for delivery. The critical path moves to utility power and plant, which prefabrication does not shorten.
- Power & Cooling
- The facility must deliver what the container expects at its interface: 800 V DC (or conversion to it) and liquid cooling sized for Vera Rubin rack loads. Agreeing those interface specifications early with the partner becomes the main design task.
- IT Infrastructure
- Factory acceptance testing replaces much of on-site rack build and burn-in; the IT TPM witnesses tests at the factory and plans site acceptance against the same test plan.
- Networking
- Rack-to-rack cabling inside the unit is done; uplinks to the wider fabric and the site’s fibre are still site work and must be ready on arrival.
Commercial availability
- AnnouncedDescribed publicly; no date for customers to buy it.
- Planned (current stage)A stated availability window that has not arrived yet.
- ShippingIn production and reaching early customers or partners.
- Generally availableOrderable by any customer through normal channels.
- DeployedRunning in customers’ production environments, confirmed by them or the manufacturer.
NVIDIA says partners will “deliver DSX POD AI factories worldwide in early 2027.” Nothing has shipped yet.
Not yet confirmed
- The 14-day figure is NVIDIA’s and has not been measured independently.
- NVIDIA gives no rack count, power rating or cost for a POD.
Learn more in BSS/OSS Academy
Official sources
- NVIDIA Introduces DSX POD AI Factories to Accelerate Europe’s Infrastructure Deployment(opens in a new tab)NVIDIA Newsroom · 21 October 2026 · Manufacturer
- Achieve Productive AI Capacity Faster and More Predictably with NVIDIA DSX POD(opens in a new tab)NVIDIA Technical Blog · 21 October 2026 · Technical documentation
Figures are as the manufacturer states them unless marked otherwise. None has been independently tested by BSS/OSS Academy.