Nvidia DSX POD and the supply-chain capacity deal: containerized Vera Rubin AI factories, 2.5x substrates and 15x optics

Nvidia DSX POD is a reference design for complete Vera Rubin AI datacenters that partners build inside modular containers and validate before delivery. Nvidia introduced it at GTC Berlin on October 21, 2026, saying it cuts deployment at a site that already has power and cooling from six months to "as little as 14 days." The same day, Nvidia said six suppliers are expanding the manufacturing capacity behind its chips, with targets of 2.5 times more substrate capacity and 15 times more co-packaged optics capacity. Every performance and capacity figure in both announcements is Nvidia's own.

This story is based on three Nvidia press releases and its GTC Berlin keynote blog, all dated October 21, 2026, plus DatacenterDynamics' report, checked October 23, 2026. Vera Rubin itself, which reached customer production in September, is covered in our Vera Rubin story.

The short version

• DSX POD: a containerized, pre-validated Vera Rubin AI factory design from 17 build partners; expected worldwide in early 2027.
• The 14-day claim: applies only "at a prepared site," with power, cooling and networking already in place. Nvidia doesn't define its six-month baseline.
• Supply chain: six named suppliers expanding substrates, packaging, testing and optical engines, with dates from February 2027 to 2028.
• Efficiency: up to 100% less cooling water and up to 30% less power, both vendor-reported "up to" figures.
• Not disclosed: prices, GPU counts or megawatts per POD, and the total value of the supply deals.

What is Nvidia DSX POD?

DSX POD is part of Nvidia's DSX AI Factory Platform, its set of designs and software for building and running AI data centers. Nvidia's release describes DSX POD as "complete NVIDIA Vera Rubin AI factories" that partners build "inside modular containers." Computing, networking, cabling, rack-level power and liquid cooling are integrated at the factory. Partners then finish rack-to-rack cabling, wiring, piping and functional validation before shipping.

Nvidia's release doesn't give a rack or GPU count. In his keynote, Jensen Huang said "seven liquid-cooled compute racks connected with 17 km (11 miles) of Spectrum-X cables form one Vera Rubin DSX POD," DatacenterDynamics reported. DCD identifies the racks as Vera Rubin NVL72 systems.

The 17 named build partners are Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology, Schneider Electric, Vertiv and Wistron. No customers or sites were named. Nvidia expects DSX POD AI factories "worldwide in early 2027."

Huang's pitch was speed: "An AI factory creates value when it starts producing tokens — not when construction is complete." The 14-day figure covers onsite deployment once the site's utilities are ready, Nvidia's keynote blog says, not building the power and cooling plant, which is usually the slow part. Our gigawatt datacenter analysis found that power, gas turbines, transformers and financing, not chips alone, decide how much announced capacity gets built.

Which suppliers are expanding capacity, and by how much?

Nvidia's second release names six companies expanding capacity for its products. Huang said: "Demand is extraordinary. Capacity is the constraint."

Last verified: October 23, 2026

Across suppliers, Nvidia says it plans 2.5 times more substrate capacity over three years and expects total testing capacity to double from calendar 2026 to 2027, with buildouts in the US, Taiwan and Singapore. An unnamed Singapore testing expansion is due to start high-volume production in February 2027. The only dollar figures are KYEC's $1.1 billion and IBIDEN's unquantified "multibillion-dollar" advance, which the release says was agreed earlier in 2026.

Substrates are the layered boards that chips and memory sit on, and interposers link GPU dies to high-bandwidth memory inside one package. Co-packaged optics put optical transceivers next to the switch chip, replacing pluggable modules.

What efficiency does Nvidia claim, and on what basis?

A third Nvidia release, on European partners adopting DSX, carries the efficiency numbers. All are vendor-reported.

• Water: "up to 100%." DSX designs use closed-loop liquid cooling with outdoor dry coolers, which reject heat without evaporating water. At 2 gigawatts, Nvidia estimates savings of 54 million liters a day compared with conventional cooling towers, which it equates to the daily use of 180,000 households. Nvidia notes that local climate decides whether extra mechanical cooling is needed in the hottest periods, and says the designs support coolant temperatures up to 45°C.
• Power: "up to 30% less." Nvidia says DSX MaxLPS, which shifts power between workloads within a facility's fixed budget, can produce the same tokens with up to 30% less power. The release gives no test conditions for that figure. It cites two smaller results: a Lambda proof of concept on HGX B200 systems with 24% more token throughput and 19% more GPU capacity within a fixed power budget, and an Nscale evaluation in Iceland on four GB300 NVL72 racks with 49% higher inference throughput and 37% more GPU capacity within 264 kilowatts.
• Grid flexibility. In a trial with Emerald AI, EPRI, Nebius and the UK's National Grid, DSX Flex cut a Blackwell Ultra cluster's power draw by up to 40% during simulated grid events, Nvidia says.

None of these results comes from Vera Rubin hardware, and none was published with a method. Nscale, one of the evaluation partners, is the company whose Essex project lost Microsoft as an anchor, as our Nscale story reported.

Why it matters for AI data center buyers

Vera Rubin is the platform the biggest buyers are waiting for. Meta, for one, says Vera Rubin NVL72 racks come next in its Prometheus cluster, per our Prometheus story. DSX POD is Nvidia's attempt to sell more of the finished building, not just the chips, through partners who integrate everything off-site. The supply deals show where Nvidia sees bottlenecks: substrates, packaging, testing and optics, with most added capacity arriving in 2027 and 2028.

What's confirmed, reported and uncertain

• Confirmed (Nvidia releases, October 21, 2026): DSX POD's design, 17 partners and early-2027 timing; the six suppliers, their targets and dates; the $1.1 billion KYEC advance.
• Vendor-reported: the 14-day deployment time, the water and power savings, and the Lambda, Nscale and grid-trial results.
• Reported (DatacenterDynamics): Huang's description of seven racks and 17 km of Spectrum-X cable per POD.
• Uncertain: DSX POD pricing, power per POD, GPU count, named customers, and the total value of Nvidia's supply commitments.