Jensen Huang Unveils Preassembled AI Data Centers: "Installation Takes Just 14 Days"

Unveils Containerized AI Factory "DSX POD"
Global Rollout to Begin in Early 2027
Brings Forward Time to First Token, Enabling Earlier Monetization

Nvidia has unveiled a modular infrastructure design that can cut the time needed to build an artificial intelligence (AI) data center to as little as 14 days. The strategy is to shorten construction time, bringing forward the start of operations and the point at which investments can be recouped.

At GTC Berlin 2026, held in Berlin, Germany, on October 21 (local time), Nvidia announced DSX POD, a modular design standard for AI factories. It is based on Vera Rubin, Nvidia's next-generation computing platform. In collaboration with global partners, the company plans to begin supplying it to the market in early 2027.

Traditional data center construction involved complex on-site design and lengthy civil engineering work. There was also a high risk of delays due to difficulties securing specialized engineers and disruptions in component supplies.

Nvidia CEO Jensen Huang delivers a keynote address at the annual GTC developer conference in Berlin, Germany, on the 21st (local time). AP-Yonhap News

By contrast, DSX POD has computing units, networking equipment, cabling, rack-level power distribution units and liquid-cooling systems preassembled at manufacturing plants. The finished systems are delivered to the site after their functions have been verified. At the site, container units simply need to be connected to pre-installed power grids, cooling pipes and communications networks. Nvidia said the on-site construction period can be reduced from six months to 14 days.

Nvidia CEO Jensen Huang said, "It costs about $50 billion to build a 1-gigawatt (GW) AI data center," underscoring that the speed of completion is a key factor determining a project's profitability. Assuming data center facilities generate $40 billion in revenue annually, a delay of just three months in construction would mean missing out on $10 billion in revenue, he said.

The shorter time needed to prepare systems for operation is also expected to significantly advance the time to first token (TTFT). TTFT refers to the time it takes for a newly built AI computing facility to begin commercial computing operations and generate its first data token. Corporate customers can start training and inference workloads for AI models earlier and recoup service revenue more quickly.

Huang said, "We expect to reduce the time from when computing equipment arrives on site to when it starts producing tokens from the current four to five months to about one month," adding, "This will improve profitability for everyone by reducing idle computing resources and unused capital investments."

Meanwhile, Nvidia also announced plans to expand production capacity with six substrate, packaging and testing partners, including Samsung Electro-Mechanics, to ease bottlenecks in AI chip supplies. Ibiden, Samsung Electro-Mechanics and Unimicron will help secure substrates, while Amkor Technology will quadruple its chip packaging supply by the end of 2028. Nvidia plans to double semiconductor testing capacity in 2027 compared with 2026.

This content was produced with the assistance of AI translation services.