Nvidia DSX Pod Packs Seven Vera Rubin Racks Into One Shipping Container

Photo by CHUTTERSNAP on Unsplash

Nvidia used its GTC Berlin keynote on October 21, 2026 to unveil the DSX Pod AI Factory, a reference design that stuffs an entire Nvidia Vera Rubin rack cluster, seven of them, into a shipping container that leaves the manufacturing plant pre-wired, pre-plumbed, and liquid-cooled. The company says the DSX Pod can cut the time between a prepared site receiving the box and that site serving AI tokens to roughly 14 days, down from the six months a conventional buildout typically takes.

The unit is heavy, dense, and expensive. It weighs about 62,000 pounds and can ship by truck on normal roads, and Nvidia did not disclose a price, though analysts estimate a single Vera Rubin rack can cost more than $7 million. Multiply that by seven racks per pod and the arithmetic turns ugly fast.

What Is Actually Inside the 28-Tonne Box

504 Rubin GPUs. Each container holds seven Vera Rubin NVL72 racks, or 504 Rubin GPUs, and Nvidia's VP of hyperscale and HPC Ian Buck said in a press briefing that a single unit packs about 10 million components and weighs around 28 metric tons. The containers also come with roughly 20 kilometers of copper cabling and 17 kilometers of scale-out networking preinstalled and tested, so once a container arrives, crews set it in place, connect it to its neighbors, and run the final burn-in tests.

The pod inherits the specs of the underlying rack. According to thundercompute.com's architecture breakdown, each Vera Rubin NVL72 rack houses 72 Rubin GPUs and 36 Vera CPUs, aggregating 20.7 TB of HBM4 memory and 1,580 TB/s of memory bandwidth, with Nvidia rating the full rack at 3,600 PFLOPS of NVFP4 inference and 2,520 PFLOPS of NVFP4 training. A fully populated DSX Pod therefore carries about 145 TB of HBM4 memory and, on Nvidia's own numbers, around 25.2 exaFLOPS of NVFP4 inference compute inside one steel container.

The silicon doing the work is the largest reticle-scale GPU Nvidia has shipped. Silicon Report notes that the standalone Rubin GPU is built on a TSMC N3-class process and packs 336 billion transistors, up from Blackwell B200's 208 billion, and that an NVL72 rack delivers 1.26 exaFLOPS of dense FP8 compute, 260 TB/s of NVLink 6 scale-up bandwidth inside the rack, and 28.8 TB/s of scale-out networking.

Why Nvidia Is Selling Steel Boxes, Not Just Chips

The pod is a hedge against the one bottleneck Nvidia's chip advantage cannot solve on its own: the physical data center. Hyperscalers are power-constrained, not GPU-constrained, and traditional buildouts still move at the pace of concrete, permits, and electricians. Nvidia said DSX PODs won't replace traditional data centers, but will serve as a means to start running AI applications before brick-and-mortar facilities are ready.

There is a water-use angle buried in the announcement that most of the day-one coverage glossed over. According to the company, the DSX POD supports liquid cooling, but because it uses dry-cooler designs, it can save up to 54 million liters of water per day versus a standard data center. That matters because data-center water consumption is now a live political issue in markets like Ireland, the Netherlands, and the US Southwest, where new AI campuses have run into permitting resistance over cooling-tower draw. A dry-cooled modular pod sidesteps the water-rights fight that has stalled several conventional builds.

There is also a channel-strategy signal here. Shipments start in early 2027 through 17 partners including Dell, HPE, Lenovo and Foxconn, which means Nvidia is not building these itself. It is standardizing a reference design that OEMs manufacture to spec, the same playbook Nvidia used to industrialize the DGX line. The result lifts Nvidia's revenue per gigawatt of deployed capacity without turning Nvidia into a construction contractor.

The $7 Million Rack and the Memory Bill That Changed Everything

The reason the DSX Pod's unpriced sticker is being quoted in the tens of millions traces back to the memory bill of materials. According to wccftech's reporting on Morgan Stanley's teardown, a full Vera Rubin NVL72 rack now carries an estimated $7.8 million bill of materials, with the Rubin GPUs alone accounting for nearly $4 million, a 57% increase over the Blackwell NVL72 B300 rack's roughly $2.5 million GPU cost.

Memory is the single line item that moved most. The same Morgan Stanley analysis found HBM4 and LPDDR5X costs surged 435% versus the Grace Blackwell platform, climbing from about $374,000 to more than $2 million per rack, while Vera CPUs account for roughly $180,000 of the total, or about $5,000 per chip. The remaining $2 million covers NVLink switches, networking silicon, cooling, PCBs, substrates, and power supply, meaning cost pressure is compounding across the entire stack rather than isolated to one component.

Straight multiplication puts seven racks of BoM at around $54.6 million before Nvidia's margin, Foxconn's integration labor, the container shell, the factory burn-in testing, or shipping. The pod is not a discount SKU. It is a logistics product dressed as a hardware product.

Nvidia Vera Rubin NVL72 Specs at a Glance

• Compute per rack: 3,600 PFLOPS NVFP4 inference / 2,520 PFLOPS NVFP4 training, per thundercompute.com
• Silicon: 72 Rubin GPUs plus 36 Vera CPUs; each Rubin die packs 336 billion transistors on TSMC N3-class, per Silicon Report
• Memory: 20.7 TB HBM4 aggregate, 1,580 TB/s rack memory bandwidth
• Fabric: NVLink 6 at 260 TB/s scale-up inside the rack, 28.8 TB/s scale-out
• Pod-level: Seven racks, 504 Rubin GPUs, 252 Vera CPUs, ~28,000 kg, per CNBC
• Production status: Vera Rubin entered full production June 1, 2026 per Silicon Report; DSX Pod deliveries targeted for early 2027 per CNBC

Who Actually Gets One in 2027

Hyperscalers first. According to VRLA Tech's roadmap tracking, first cloud deployments of Vera Rubin are confirmed from AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale, with channel allocation for enterprise on-premises buyers following after the hyperscalers and realistic enterprise delivery timelines extending into 2027 due to allocation queues.

That queue is the quiet caveat on the "14 days to tokens" marketing line. The 14 days starts when the container arrives at a prepared site. The 2027 shipping window starts after whoever ordered the pod clears Nvidia's allocation list, which already has every named hyperscaler ahead of them. Enterprise buyers hoping the DSX Pod shortcuts them past the queue are reading the announcement wrong.

European customers are the explicit targeting signal. Nvidia said the pod is intended for customers worldwide and could help address deployment challenges faced by European customers, which maps directly onto the EU AI sovereignty push and the regional datacenter construction bottlenecks that have slowed Nvidia's European growth. Announcing in Berlin, rather than at the main GTC in San Jose, was not an accident. For broader context on how hyperscalers are responding to Nvidia's vertical integration, our coverage of Microsoft's Azure Maia 300 inference chip walks through the Nvidia-alternative strategy playing out in parallel.

The Power Problem the Pod Does Not Solve

The DSX Pod ships cooling, networking, and compute. It does not ship power. A Vera Rubin NVL72 rack draws roughly 190 to 230 kW, which puts a seven-rack pod in the 1.3 to 1.6 MW range before PDU and cooling overhead. According to VRLA Tech, the next-generation Rubin Ultra Kyber rack (NVL576) targets about 100 PFLOPS NVFP4 per GPU and roughly 1 TB HBM4e per GPU, with rack power consumption jumping to approximately 600 kW, raising data center power density concerns well beyond the current Vera Rubin NVL72.

That trajectory undercuts the pod's own value proposition on a two-year horizon. If a 2028 Rubin Ultra rack draws three times the power of today's unit, the limiting factor stops being how fast you can wire the racks and starts being how fast the local utility can bring a new substation online. The DSX Pod optimizes the integration step. It cannot optimize the grid interconnect queue, which in several US ERCOT and PJM markets now runs 4 to 7 years.

For comparison, CKPillaiAI noted on X that "pre-integrated delivery can remove a lot of deployment friction" but questioned how the 14-day timeline accounts for site power, cooling, networking, and acceptance testing. Nvidia's own fine print concedes the pod needs "a prepared site," which is doing most of the work in that sentence.

What to Watch Next

The concrete signal to watch is the first named DSX Pod customer announcement, which Nvidia has indicated will arrive alongside the early-2027 shipment window. Pricing disclosure from one of the 17 OEM partners, most likely Dell or HPE given their enterprise-channel pricing conventions, will confirm whether the pod carries a logistics premium above the stacked rack BoM or whether Nvidia is subsidizing integration to accelerate deployment. Either outcome will reset expectations for what "AI factory" pricing actually means. For readers tracking the broader shift in datacenter silicon economics, our running coverage in the AI Infrastructure section and the Semiconductors & Chip Supply section will follow the pod's first field deployments as they land.