Nvidia launched the Data Center World 2025 event this week in Washington, DC, with a daring vision of the future of IA infrastructure.
In his speech, Wade Vinson, chief engineer of the Nvidia data center, presented the concept of data centers at AI scale; These massive and energy efficient installations would meet the demand for calculating accelerated computers. NVIDIA is considering large “AI factories” powered by Blackwell GPUs and DGX superpods, supported by cooling and feeding systems advanced by Vertiv and Schneider Electric.
“There is no doubt that AI factories are a major trend in the world of the data center,” said Vinson.
Finish the first phase of an AI factory in Texas
Vinson underlined the Lancium clean campus that Crusoe Energy Systems built near Abilene, Texas. As he explained:
- The first phase of this AI factory is widely completed: 200 MW in two buildings.
- The second phase will extend it to 1.2 GW. It should be completed in the middle of 2026.
- The design includes direct liquid cooling with a chip, rear door heat exchangers and air cooling.
- It will include six additional buildings, installing four million square feet.
- 10 gas turbines will be deployed on site to provide on -site power.
In addition, each building will operate up to 50,000 GPU NVIDIA GB200 NVL72S on a single integrated network fabric, advancing the border of the design and scale of the data center for training and workloads of IA inference.
Vinson said that some AI factories will take advantage of energy on site, while others will benefit from the sites where electricity is already available. He underlined the old factories, manufacturing sites and retail facilities which are already connected to the grid.
For example, an old shopping center in San Francisco can be converted to the AI factory in months, rather than the many years necessary to finish the construction of new constructions and obtain interconnections and utility permits. These sites often have large roofs that can be used for solar energy networks.
Reconfigure existing data centers in AI factories
What about existing data centers? Aging structures can find it difficult to adapt to Nvidia equipment and AI applications. Vinson thinks that many roommate installations (colos) are in good position to be transferred to AI factories.
“All colors built in the past 10 years has had enough power and cooling to become an AI factory,” he said. “AI factories should be considered a income opportunity rather than expenses.”
He believes that AI could stimulate business and personal productivity 10% or more, adding 100 billions of dollars to the world economy.
“It represents a more important change in productivity than what happened due to the wave of electrification in the world that started about 100 years ago,” said Vinson.
Planning is the key to the success of the IA factory
Vinson warned those interested in building or managing their own AI factories on the importance of planning. It is important to consider the different factors involved and the modeling is vital.
He praised Nvidia's Omnivese simulation tool as a way to properly plan an AI factory. It uses digital Twin technology to allow complete modeling of the data center infrastructure and design optimization. Not modeling in advance and simulating many possible scenarios can lead to ineffectiveness in fields such as energy consumption and can extend construction times.
“Simulations allow data centers to improve operational efficiency through holistic energy management,” said Vinson.
SEE: Data centers can reduce energy consumption up to 30% with approximately 30 lines of code
For example, many veterans in the data center can find it difficult to pass traditional concepts of racks, alleys and servers to GPU equipment surrounded by liquid cooling and with adequate energy and energy distribution equipment.
The AI factory conceptions will have much more power and cooling equipment inside than server racks; Therefore, the provisions will be radically different. After all, the amount of heat generated by the superpods fueled by GPU is more than that generated by typical data centers.
“Expect a significant consolidation of racks,” said Vinson. “Eight old racks could well become a future rack with GPU inside. It is essential to develop a simplified power and a cooling configuration for racks inside AI factories, because these will be very different from what most of the data centers are used to. ”
