Nvidia unveils DSX POD prefabricated AI data centre it says can be installed in 14 days
In short
Nvidia CEO Jensen Huang unveiled DSX POD, a factory-assembled modular AI data centre design, at GTC Berlin on 21 October (local time). Nvidia said it cuts on-site build time from six months to as little as 14 days, with supply planned from early next year. Outlets also reported capacity-expansion plans with substrate makers including Samsung Electro-Mechanics.
Read the full story 2 min read
Nvidia CEO Jensen Huang unveiled DSX POD, a modular design for AI data centres, at GTC Berlin on 21 October (local time), the outlets reported. Servers, cabling and cooling equipment are assembled and verified at a factory and then shipped to the site. Nvidia said that where infrastructure is already in place, equipment installation time falls from six months to as little as 14 days. [ 1 , 2 , 3 ]
Asia Economy said the design is based on the Vera Rubin platform and includes compute units, network equipment, cabling, rack-level power distribution and liquid cooling, delivered as a finished system that is connected to existing power, cooling piping and communications on site. Chosun Ilbo said DSX POD is supplied in container form. Supply is planned from early next year, together with partners; Seoul Shinmun named Dell, HPE and Lenovo, and Chosun Ilbo said Nvidia has partnerships with 17 server, power and cooling companies. [ 1 , 2 , 3 ]
Huang said a 1 GW AI data centre costs about $50 billion to build. Assuming $40 billion in annual revenue, a three-month delay in start-up means $10 billion in forgone revenue, he said. Asia Economy quoted him saying the time from equipment arrival to token production could fall from four to five months to about one month. Chosun Ilbo quoted him saying the value of an AI factory arises when token production begins, not when construction ends. [ 1 , 2 , 3 ]
Chosun Ilbo reported that tech-industry observers see the launch as part of Nvidia's strategy to set standards for the AI industry. It said Nvidia holds more than 80% of the global AI accelerator market and has recently set a liquid cooling standard for Vera Rubin and launched a certification scheme for power and cooling equipment called DSX Ready. [ 3 ]
Seoul Shinmun said Nvidia is also widening its business into AI agents, robots and autonomous driving, and that Vera Rubin has entered full production. It reported that Nvidia, with Samsung Electro-Mechanics, Japan's Ibiden and Taiwan's Unimicron, plans to raise package substrate capacity 2.5 times over three years, and that Samsung Electro-Mechanics will invest jointly with Nvidia in dedicated facilities at its Sejong and Busan sites. Asia Economy said Nvidia announced capacity plans with six substrate, packaging and test partners, including Amkor Technology raising chip packaging supply fourfold by the end of 2028. [ 1 , 2 ]
Seoul Shinmun added that borrowing burdens at data centre operators such as CoreWeave have raised concern about the profitability of AI infrastructure investment, and that securing power and financing remain separate problems. [ 1 ]
Why it matters
Nvidia argues that faster data centre start-up shortens the time to revenue on very large AI investments. Chosun Ilbo cites tech-industry observers who see the move as Nvidia trying to set standards for how AI data centres are built, beyond chips and servers.
Key facts
- Jensen Huang unveiled DSX POD at GTC Berlin on 21 October (local time). [ 1 , 2 , 3 ]
- DSX POD is a modular design in which servers, cabling and cooling equipment are assembled and tested at a factory before being delivered to the site. [ 1 , 2 , 3 ]
- Nvidia said on-site build time falls from six months to as little as 14 days where infrastructure such as power and cooling is already in place. [ 1 , 2 , 3 ]
- DSX POD is based on Nvidia's Vera Rubin platform, and supply is planned from early next year. [ 1 , 2 , 3 ]
- Huang said a 1 GW AI data centre costs about $50 billion to build and that, assuming $40 billion in annual revenue, a three-month delay forgoes $10 billion in revenue. [ 1 , 2 ]
- Chosun Ilbo reported that Nvidia has built partnerships with 17 server, power and cooling companies, including Dell Technologies, HPE, Lenovo, Foxconn and Vertiv. [ 3 ]
Confirmed by several sources
- Nvidia unveiled the DSX POD modular AI data centre design at GTC Berlin on 21 October (local time), saying it cuts installation from six months to as little as 14 days. [ 1 , 2 , 3 ]
- Huang put the cost of a 1 GW AI data centre at about $50 billion and said a three-month delay on $40 billion in annual revenue would forgo $10 billion. [ 1 , 2 ]
- Nvidia is working with Samsung Electro-Mechanics, Ibiden and Unimicron to expand supply of chip package substrates. [ 1 , 2 ]
- Supply of DSX POD is planned from early next year. [ 1 , 2 , 3 ]
Still unclear
- Whether Nvidia and Samsung Electro-Mechanics will make a joint investment in new production facilities at Sejong and Busan. Only Seoul Shinmun reports the joint investment and the sites; Asia Economy mentions only a capacity-expansion plan with partners.
- Whether faster start-up resolves concerns about the returns on AI infrastructure spending. Seoul Shinmun says borrowing burdens at data centre operators such as CoreWeave have raised profitability concerns and that power supply and financing remain separate problems; the other outlets do not address this.
- Huang's claim that the time from equipment arrival to token production can fall from four to five months to about one month. Reported only by Asia Economy.