Vertiv and NVIDIA Push an AI Factory Approach, With the Numbers Modelled, Not Measured
The two companies, with Dassault Systèmes, say pre-engineered designs and AI factory-built infrastructure cut weeks from a build. Every headline figure is a model, and the release says so.
Vertiv and NVIDIA used NVIDIA’s GTC Berlin event to introduce an approach to building AI data centres that they say moves a project from kickoff to productive capacity far faster. The idea starts from pre-engineered reference designs rather than a blank sheet, configures them in a shared virtual environment, validates power and cooling in a digital twin, and then builds as much of the infrastructure as possible in a factory before it reaches the site. Dassault Systèmes supplies the virtual-twin platform, and Polar Data Centers applies the approach to the first 41 MW phase of a campus at Herøya in Norway.
The pitch rests on a simple observation that Vertiv’s chief executive, Giordano Albertazzi, puts plainly: an AI factory earns its return in tokens, and every day spent configuring or integrating infrastructure is output that never comes back. That framing is sound, and it explains why the announcement leads with time. It also explains why the numbers need care. Vertiv labels each of them as modelled and states that actual results will vary.
What Vertiv and NVIDIA announced
The approach has three stages. The first is configure and validate. Teams start from pre-engineered AI factory reference designs and adapt them to a site’s compute, power and cooling requirements inside a shared virtual environment. Vertiv built those designs around NVIDIA’s Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint, added its own SimReady power and cooling assets, and integrated Dassault Systèmes’ 3DEXPERIENCE agentic platform. The second stage is deployment. Vertiv shifts assembly, integration and testing into the factory, using prefabricated products such as OneCore, a hybrid-built data centre solution, and SmartRun, an overhead infrastructure system that carries high-density power, liquid cooling and networking.
The third stage is the physical build, where prefabricated modules, SmartMod and Unify software carry the validated design to the site. Vertiv also publishes SimReady OpenUSD assets of its equipment to the SimReady Central Library on Hugging Face, so engineers can reuse validated components in digital twins.
The approach sits inside a wider NVIDIA effort. The company released the Vera Rubin DSX AI Factory reference design and the Omniverse DSX Blueprint in March 2026, with support from Cadence, Dassault Systèmes, Eaton, Jacobs, NScale, Phaidra, Procore, PTC, Schneider Electric, Siemens, Switch, Trane Technologies and Vertiv. At GTC Berlin, NVIDIA also described a European ecosystem building DSX-based factories that cut cooling water to near zero through closed-loop liquid cooling and dry coolers, and it launched the SimReady Central Library on Hugging Face to hold validated equipment assets. Vertiv’s announcement forms one part of that programme, and the reference design underneath it belongs to NVIDIA rather than to Vertiv.
The mechanism, and why the sequence matters
The traditional way to build an AI data centre runs power, cooling and facility systems through separate, sequential engineering processes, and the interfaces between them become the source of delay. Vertiv and NVIDIA invert that sequence. They resolve the interfaces in a virtual twin first, where a conflict costs a software change rather than a construction change, and only then commit to steel, pipework and cable. That answers a real problem. High-density, liquid-cooled AI halls demand far more integrated design than the air-cooled data centres the industry built for two decades, and every pipe, fitting, valve and power interface becomes a potential point of failure. The case for prefabrication also goes beyond speed.
Building a high-density AI hall demands electricians, plumbers and HVAC technicians who understand liquid cooling, and those skills are scarce in most markets. Shifting assembly, integration and testing into a factory moves the work to where the labour exists and where quality control is easier, and it lets site preparation and module manufacturing run in parallel rather than in sequence.
That is why the industry has converged on the approach: Schneider Electric, Eaton and Vertiv all now sell prefabricated modules built around NVIDIA‘s reference designs, and all of them pitch the same two benefits, less on-site labour and less on-site risk. Digital twins and pre-engineered designs are not new, and several infrastructure vendors now sell some version of them. What is new here is the breadth of the partnership and the claim that the whole path, from design to deployment, compresses together.
The modelled numbers
The headline figures are where the release needs careful reading, and Vertiv deserves credit for labelling them clearly. The company says the configure-and-validate stage falls from about 30 days to three, a 10x improvement. It says factory-built deployment cuts the on-site period from about 90 days to 45, a reduction of up to 50 per cent. Combined, it says the two stages fall from roughly 120 days to 48, which makes productive capacity available about 10 weeks sooner and generates up to 30 per cent more cumulative token output in the first 12 months at the same installed capacity. The footnotes describe every one of those as modelled, with the token figure tied to a specific assumption: 317 productive days against 245, measured from design kickoff. Vertiv states plainly that actual results will vary.
That distinction matters more than it might appear. A modelled result projects from assumptions; it does not observe a site. The 30 per cent token figure follows from the schedule saving rather than from anything measured in a hall, and it holds only if the assumptions hold. The claim is coherent, because capacity that comes online ten weeks earlier produces more tokens over a year, but a buyer should treat the percentage as an illustration of the arithmetic rather than a guaranteed gain, and should ask what drives it. The one piece of real-world evidence in the release is the Polar deployment, and that is a project in progress rather than a completed comparison.
The Polar deployment
Polar Data Centers applies the factory-built model to the first 41 MW phase of its campus at Herøya, Norway, a site it plans to scale to 100 MW. Vertiv supplies prefabricated power and thermal modules, SmartRun overhead infrastructure and SmartMod, supported by Unify software. Polar’s chief executive, Andy Hayes, frames the appeal in operational terms: customers measure the company on how quickly capacity is ready, and the factory-built model gives a more disciplined, repeatable deployment while preserving flexibility. Norway also offers cheap renewable power and cool ambient conditions, which suit liquid-cooled AI halls, so the site makes a reasonable place to prove a build method that leans on prefabrication and long supply chains.
The market context
The announcement lands in a market defined by speed and scarcity. Operators race to bring AI capacity online because demand outstrips supply, and the constraint has shifted from chips to the buildings that house them: power, cooling and the skilled labour to integrate them on site. That is why prefabrication, digital twins and reference designs have moved from a niche to the centre of the industry’s thinking. Vertiv’s approach is not unique. Eaton sells a Beam Rubin DSX platform built on the same NVIDIA reference design, and Schneider Electric ships prefabricated AI pods and a modular DSX reference design.
Schneider has also published a measured time study with Compass Datacenters that found onsite installation falling from 75.5 hours to 15, an 80 per cent reduction, which is the kind of observed evidence Vertiv’s release does not yet contain. NVIDIA gains from all of it, because a GPU that sits in a warehouse produces no tokens either.
What buyers should ask
For an operator weighing the approach, the questions are practical. Which of our sites suits prefabrication, given transport limits and local labour? How closely does a reference design match our compute mix, and how much reconfiguration does it need? What did the modelled figures assume, and how do they hold under our schedule and supply chain? Does the digital twin stay useful after construction, or does it end at handover? How do Vertiv’s prefabricated modules interoperate with equipment we already own? And what happens to the promised time saving when a site hits a permitting delay or a power connection slips, since those dominate many projects?
What it signals
The Vertiv and NVIDIA approach answers a real bottleneck with a coherent method, and the partnership behind it carries weight. Pre-engineered designs, virtual validation and factory-built deployment each remove a genuine source of delay, and combining them is the logical next step for an industry that must build faster than it ever has. The caveat is the one the release states: the numbers are models. That is no reason to dismiss them, and Vertiv deserves credit for labelling them rather than presenting them as results. It is a reason to watch the Polar deployment, and the next few like it, for the measured figures that will show whether the modelled saving survives contact with a construction site. Schneider has already published that kind of evidence from its own deployment, which sets the bar the rest of the industry now has to clear.
None of that diminishes the announcement. The bottleneck is real, the method addresses it, and the partnership gives Vertiv a place inside the reference design that the rest of the industry now builds around. But the industry has reached the point where the argument no longer needs to be made and the evidence does. Prefabrication works; the open question is how much, under which conditions, and at what cost. That is a question measured deployments answer and models cannot.