Nvidia targets AI factory deployment in 14 days at prepared sites
Key Facts
Nvidia announced an AI factory design called "DSX Pod" in October 2026 that it says cuts deployment time at a prepared site from 6 months to as little as 14 days. The design uses standardized modules manufactured and tested before arriving at the operating site, according to the company. These modules connect to the site's power, water and networks to become part of the factory's operating infrastructure. The modules can be replicated and scaled to form a larger AI factory, according to Nvidia's product description. This approach shifts integration work to the factory, while retaining the need for supporting infrastructure at the operating site, which the company said in its technical blog the design "does not eliminate".
The design includes a standard 800-volt direct-current power input, according to Nvidia's technical blog. Partners handle the engineering of local power generation, conversion and electrical connections upstream of this input. Hardware and firmware validation moves to the factory before the integrated module is shipped. Validation procedures also cover networks, sensors and cooling integrity within the module, according to the company. Nvidia said the design includes "system-level functional testing before the container leaves the factory" as part of preparations before equipment arrives on site.
On operating efficiency, Nvidia says its "Max LPS" suite of technologies helps produce the same number of tokens using up to 30% less electricity. The suite combines chip, system and software technologies, according to the company's product description. It also incorporates cooling at 45 degrees Celsius as part of its design to improve utilization of available power capacity. Nvidia says computing capacity can increase by up to 40% within the same power budget. The company thus presents efficiency improvements in 2 operating forms: reducing the electricity needed to produce the same tokens and enabling additional computing capacity within a fixed power limit.
In a test at Lambda, token throughput rose 24% using "Max LPS", according to results published by Nvidia. Available graphics processing unit capacity increased 19% within a fixed power budget. Performance per watt improved 23% in the same test. In a configuration running inference and training concurrently, the suite enabled Lambda to increase inference throughput by 20%, according to the case study published by Nvidia. Mixed training throughput also rose 17%, while keeping the power budget fixed when running these workloads.
On cooling, Nvidia estimates that the "DSX" platform can save 54 million liters of water daily at the scale of a 2-gigawatt facility. The company compares these savings with facility cooling water use in conventional systems that use cooling towers. The hottest periods may require additional mechanical cooling depending on the local climate, according to Nvidia's statement. For operating software, the company presents "DSX OS" as a modular software suite that can be adopted incrementally. Its components can be integrated with existing systems, allowing phased adoption rather than requiring replacement of the entire operating platform.
Nvidia's partners aim to deliver "DSX Pod" factories globally in early 2027, based on the "Vera Rubin" platform. Dell Technologies aims to make a modular data center design using these modules available in the same period, according to its statement. Dell's solution includes computing, storage, networking, rack power and liquid cooling within the container-based data center. In the supply chain, Nvidia expects high-volume production to begin at expanded semiconductor testing operations in Singapore in February 2027. The company is also making a $1.1 billion capacity reservation payment to KYEC to support new U.S. testing facilities. These facilities target the start of high-volume production in April 2027, according to Nvidia's statement on its expansion with supply chain partners.