Nvidia unveils DSX POD modular AI factory and plans supplier capacity expansion
In short
At its GTC conference in Berlin on 21 October, Nvidia introduced DSX POD, a modular AI factory reference design it says can cut deployment on a ready site from about 6 months to as little as 14 days, with delivery planned for early 2027. The same day Nvidia announced capacity expansion with supply chain partners, including a plan to raise substrate capacity to 2.5 times the current level over three years.
Read the full story 3 min read
Nvidia introduced DSX POD, a reference design for modular AI factories, at its GTC conference in Berlin on 21 October, 36Kr and Jiemian reported. IT Home said CEO Jensen Huang made the announcement. The design puts a complete AI factory based on Vera Rubin systems into modular containers, with system integration and validation completed before delivery. [ 4 , 5 , 17 ]
According to IT Home and cnBeta, work on computing, networking, cabling, rack-level power and liquid-cooling piping is moved into the factory. After the equipment arrives, customers connect it to power, cooling and network facilities already in place. Nvidia said this reduces on-site construction, coordination of specialists and logistics, and lowers the effect of installation errors and part delays. It said that on a site that is ready, deployment can drop from about 6 months to as little as 14 days. cnBeta noted that the 14 days excludes the full cycle from site selection to completion, and customers must still prepare power, cooling and network support. [ 17 , 19 ]
IT Home quoted Huang as saying an AI factory creates value only once it starts producing tokens, not simply when it is built. Nvidia said partners plan to begin delivering the systems worldwide in early 2027. The partner lists in the documents differ: 36Kr, Jiemian and CLS name Dell, HPE, Lenovo, Schneider Electric and Vertiv; IT Home also lists Gigabyte, Foxconn Industrial Internet, Pegatron, Quanta's QCT and Wistron; cnBeta names Bull. [ 4 , 5 , 6 , 17 , 19 ]
Later on 21 October Nvidia announced capacity expansion with supply chain partners covering substrates, advanced packaging, chip testing and optical engines, Jiemian reported. Nvidia plans to raise substrate capacity to 2.5 times the current level over three years. CLS reported that Ibiden's Ogano plant in Japan targets three times its current capacity, with the added capacity starting to ramp in the first quarter of 2028. Unimicron's expansion in Taiwan is expected to contribute capacity from the second quarter of 2027 and reach full output by mid-2028. Samsung Electro-Mechanics will join the substrate supply chain, supplying first from existing capacity in South Korea. [ 8 , 9 , 10 ]
Jiemian reported that Amkor plans to raise interposer wafer capacity by more than 300% by the end of 2028. Nvidia expects total test capacity in 2027 to double the 2026 level and will give KYEC a USD 1.1 billion capacity-reservation prepayment for a new test facility in the United States, targeting mass production in April 2027. With SPIL, Nvidia plans to raise light engine capacity for co-packaged optics to 15 times the current level by mid-2027. CLS said large-scale test production in Singapore is to start in February 2027. [ 7 , 9 ]
CLS also reported that Amkor and Nvidia added a USD 1.3 billion agreement to expand capacity in South Korea, bringing their multi-year agreements to USD 3 billion; Amkor shares rose as much as 12% intraday. Jiemian said the total includes earlier cooperation in Arizona. [ 13 , 14 ]
Caixin reported that Nvidia announced the expansion in an attempt to dispel market concerns, as worries about an AI demand bubble weighed on global shares. It said the Philadelphia Semiconductor Index had fallen about 14% since 6 October, South Korea's KOSPI by more than 18% and Nvidia's shares by nearly 8%. Yicai wrote that the expansion shows Nvidia still betting on rising demand. [ 16 , 20 ]
Why it matters
Caixin reported that Nvidia announced the expansion while AI-related shares were falling worldwide, and Yicai wrote that it signals Nvidia is still betting on higher demand. The documents give no outcome for either announcement beyond Amkor's intraday share gain.
Key facts
- Nvidia introduced the DSX POD reference design at GTC Berlin on 21 October; it puts a full AI factory based on Vera Rubin systems into modular containers, integrated and validated before delivery. [ 4 , 5 , 17 ]
- Nvidia says DSX POD can cut deployment on a ready site from about 6 months to as little as 14 days. [ 4 , 5 , 17 , 19 ]
- Partners plan to deliver the modular AI factories worldwide in early 2027. [ 4 , 5 , 6 , 17 , 19 ]
- Nvidia plans to raise substrate capacity to 2.5 times the current level over the next three years, with partners including Ibiden, Samsung Electro-Mechanics and Unimicron. [ 7 , 8 , 9 ]
- Nvidia is to give KYEC a prepayment of USD 1.1 billion for a new test facility in the United States. [ 7 , 9 , 12 ]
- Amkor Technology and Nvidia added a USD 1.3 billion agreement to expand capacity in South Korea, bringing their multi-year agreements to USD 3 billion. [ 13 , 14 ]
- Nvidia plans to double semiconductor test capacity from 2026 to 2027. [ 7 , 9 ]
Confirmed by several sources
- Nvidia presented the DSX POD modular AI factory reference design at GTC Berlin on 21 October. [ 4 , 5 , 17 ]
- Nvidia says DSX POD cuts deployment on a prepared site from about 6 months to as little as 14 days. [ 4 , 5 , 17 , 19 ]
- Nvidia and its partners plan to deliver DSX POD-based modular AI factories worldwide in early 2027. [ 4 , 5 , 6 , 17 , 19 ]
Still unclear
- Which companies are DSX POD partners. The documents list different partners: 36Kr, Jiemian and CLS name Dell, HPE, Lenovo, Schneider Electric and Vertiv; IT Home also names Gigabyte, Foxconn Industrial Internet, Pegatron, Quanta's QCT and Wistron; cnBeta names Bull.
- Whether the 14-day figure covers the whole build. cnBeta says it refers only to deployment on a prepared site and excludes the full cycle from site selection to completion; customers must first prepare power, cooling and network.
- Whether the expansion will ease market concerns about AI demand. Caixin describes it as an attempt to dispel those concerns and Yicai says it shows Nvidia still betting on demand, but the documents report no market response to the announcement itself apart from Amkor's share move.
- Whether DSX POD is aimed at Europe. Two CLS headlines say the product accelerates European infrastructure deployment, while other documents describe delivery worldwide.
- Nvidia plans to raise substrate capacity to 2.5 times the current level within three years. Reported by a single source so far
- Nvidia will pay KYEC a USD 1.1 billion prepayment for a new facility in the United States. Reported by a single source so far
- Amkor and Nvidia agreed a new USD 1.3 billion deal to expand capacity in South Korea, bringing their multi-year agreements to USD 3 billion. Reported by a single source so far
What local media are saying
Timeline, local time
- CLS reports that Nvidia has released the Vera Rubin DSX POD modular AI factory reference architecture. [ 1 ]
- 36Kr, carrying a Jiemian report, gives details of the DSX POD design and the 14-day deployment figure. [ 4 ]
- CLS reports that Nvidia and supply chain partners are expanding capacity. [ 7 ]
- CLS reports a new USD 1.3 billion Amkor-Nvidia agreement; Amkor shares rise as much as 12% intraday. [ 13 ]
- Yicai publishes an analysis of the DSX POD and the expansion news. [ 16 ]
- Caixin reports on the expansion against falling AI-related shares. [ 20 ]