Vertiv and NVIDIA have introduced an integrated approach to designing and deploying AI data centres, combining virtual infrastructure simulation with factory-built systems to reduce deployment timelines. The companies estimate that the approach could cut infrastructure deployment time by up to 50% and deliver up to 30% more cumulative AI token output in the first 12 months, with the same installed capacity.
Announced at NVIDIA GTC Berlin on November 16, the approach brings together pre-engineered AI factory reference designs, virtual validation and prefabricated infrastructure. It is intended to address delays associated with separately designing and integrating computing, power, cooling and facility systems during data centre development.
Polar Data Centers is applying the factory-built deployment model to the first 41 MW phase of its AI data centre campus at Herøya, Norway. The multiphase project is planned to scale up to 100 MW.
Under the proposed model, infrastructure teams can configure existing reference designs according to site and computing requirements and validate them in a virtual environment before physical deployment. The companies estimate that this could reduce the time required to configure and validate an integrated AI factory design from approximately 30 days to three days.
The design process combines NVIDIA’s Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint with Vertiv’s SimReady power and cooling assets. These are integrated through Dassault Systèmes’ 3DEXPERIENCE platform, which enables teams to simulate infrastructure components and identify potential integration issues before construction.
The deployment stage involves shifting a larger share of infrastructure assembly, integration and testing to factories before equipment is transported to the site. Vertiv’s OneCore, which integrates power and thermal management systems, and SmartRun, a prefabricated overhead infrastructure system for power distribution, liquid cooling and networking, are among the solutions being used.
According to the companies’ estimates, this approach could reduce the infrastructure deployment period from approximately 90 days to 45 days. Combined with the shorter design and validation process, the two stages could take around 48 days, compared with 120 days under the conventional model.
The companies project that making AI infrastructure operational around 10 weeks earlier could result in up to 30% more cumulative token output during the first 12 months after project kickoff. The projection is based on a model comparing 317 productive days with 245 days at equal installed capacity. Actual results may vary depending on project requirements and deployment conditions.
Commenting on the development, Giordano Albertazzi, CEO of Vertiv, said the time required to configure infrastructure and integrate systems at the site affects how quickly AI computing capacity becomes productive. He said the use of pre-engineered reference designs and factory-built deployment could help operators reduce infrastructure complexity and bring capacity online sooner.
Vladimir Troy, Vice president of AI infrastructure at NVIDIA, said the NVIDIA Vera Rubin DSX AI Factory reference design is intended to align computing, power and cooling requirements during the design stage. He added that the Omniverse DSX Blueprint allows companies to simulate these systems together and identify potential conflicts before physical deployment.
Stéphane Sireau, Vice president, High-Tech Industry at Dassault Systèmes, said the use of virtual twins allows infrastructure components to be co-designed and simulated before construction, helping teams address integration issues earlier in the process.
At the Herøya facility, Vertiv is supplying prefabricated power and thermal modules, SmartRun overhead infrastructure and SmartMod systems, supported by its Unify software. Andy Hayes, CEO of Polar Data Centers, said the factory-built model is intended to help manage the complexity of deploying power, cooling and data centre infrastructure for high-density computing while maintaining flexibility for customer requirements.