Nvidia unveils DSX POD to accelerate AI data center construction

From construction to appliances

Nvidia CEO Jensen Huang recently introduced the DSX POD during a keynote in Berlin. The goal of this new reference design is to change how data centers are built. Instead of being slow, complex construction projects, Nvidia wants them to function like appliances that can be plugged in quickly.

The DSX POD is a large, container-sized unit that houses seven liquid-cooled server racks. This design is intended to speed up the buildout of what Huang calls "AI factories." By standardizing the hardware into a single unit, Nvidia hopes to simplify the logistics of scaling up massive computing power.

The company is working with 17 ecosystem partners to bring this to market. These partners include major technology firms such as Dell Technologies, HPE, and Lenovo. Together, they will help roll out the infrastructure needed to support the next generation of AI computing.

Technical specs and connectivity

Inside the enclosure, which weighs approximately 28,000 kg, Nvidia has integrated 17 km of Spectrum-X networking cable. Spectrum-X is the company's Ethernet networking platform. It serves as the essential plumbing that allows thousands of individual chips to communicate fast enough to act as one large computer.

The pod is packed with high-end hardware, specifically 504 Rubin GPUs and 252 Vera CPUs. To keep these components from overheating, the system utilizes built-in liquid cooling. This setup allows for high-density computing within a footprint roughly the size of a standard shipping container.

Nvidia also highlighted the efficiency of the DSX POD design. The system can achieve up to 30% lower power usage through dynamic power allocation. Furthermore, the closed-loop liquid cooling system works similarly to a car radiator, reusing coolant rather than evaporating water into the air.

The economics of speed

The primary pitch for the DSX POD is the reduction in deployment time. Nvidia claims that prepared sites can hook these pods up to existing power and cooling in as little as 14 days. This is a significant jump from the process that previously took up to six months to complete.

Huang provided specific figures to illustrate the financial stakes of these delays. He estimated that a 1 GW AI factory costs about $50 billion to build and can generate roughly $40 billion in annual revenue. He warned that a delay of just three months in deployment could result in approximately $10 billion in lost revenue opportunities.

By shortening the setup window, Nvidia aims to help companies capture market demand more quickly. The faster a company can get its AI factory online, the sooner it can begin generating income. This move targets the high-pressure environment of the current AI infrastructure race.

Sovereignty and global reach

The choice of Berlin as a venue for this announcement was intentional. Huang spent time discussing the importance of AI technology for European countries. Many European governments are currently pushing for AI sovereignty, which means they want computing infrastructure built and controlled locally.

This demand for local infrastructure creates a market for standardized, easy-to-deploy units like the DSX POD. It allows countries to build out their own

Market data and methodology

The financial performance of the company is tracked using specific metrics from SEC filings. For instance, revenue growth is determined by comparing the current fiscal year with the prior year as stated in the 10-K. Additionally, the trailing P/E ratio is marked as not meaningful if the dated trailing diluted earnings per share are zero or a loss.

Market data also distinguishes between different types of assets. The current comparison for NVDAx shows a price of $215.38 for the stock, while the token unit reflects a different value. These figures are based on split-adjusted regular-session stock closes and do not represent fair value, executable premiums, or arbitrage returns. Readers can find more information on content review in the editorial policy.