At NVIDIA GTC Berlin, October 20–22, 2026, World Wide Technology (WWT) showcased the AI Factory Digital Twin, a planning twin of a real, running data center built with NVIDIA Omniverse on Omniverse DSX Blueprint conventions. The WWT and NVIDIA co-branded demonstration gave visitors a guided look at the experience at the WWT booth. In short, the twin let us place a proposed three-rack AI pod in a real room and see where power and cooling would run short, as a planning-grade screen before anything is ordered. 

Row F in Building 60, Data Center 1, at WWT's Advanced Technology Center in St. Louis. Row F in Building 60, Data Center 1, at WWT's Advanced Technology Center in St. Louis.
Row F in Building 60, Data Center 1, at WWT's Advanced Technology Center in St. Louis. 

Why we built it 

Most AI infrastructure doesn't land in a purpose-built hall. It lands in a room that already exists, full of equipment and shaped by years of changes. And the questions that matter most about that room, such as where the new pod fits and whether that spot can cool and power it, tend to get answered late. Often, it's after the order, on the floor, when changing course costs the most. 

We built the twin to do three things before any hardware is ordered or moved: 

  • De-risk new hardware. A proposed system is checked against the room's real power and cooling limits, so surprises show up on screen instead of on the floor.
  • Plan digitally before deploying physically. Candidate locations are compared on the same physics and ranked by a stated rule, so the placement decision is made before racks are moved.
  • Play out what-ifs safely. Load changes, a failed cooling unit, a lost power feed and a second pod can all be run without touching production equipment.

What the twin does 

Our test case was a proposed AI pod: three racks of NVIDIA DGX™ B300 enterprise servers, modeled from NVIDIA's own hardware model because they are not installed in the room. A rack of the pod would draw about 17 kW at 60% of full load and up to 29 kW flat out, while the racks around it average about 3.3 kW. That is five to nine times the power and heat load. 

The twin is interactive. At the booth, visitors could click any of the 120 racks, swing its door open and see every device the asset system holds for it, at its slot. 

 

Rack E04 in the twin, door open, with its device list from the asset system. 

The walkthrough told one story: the room as metered today, three candidate bays for the pod, the pod at 60% of full load, the pod at full load on a busy day, a cooling unit or a power feed lost, and finally a second pod and the cooling needing to support it. Airflow appeared as CFD streamlines, and every number carried a label saying where it came from. 

 

The twin's guided walkthrough opens on the room today. 

What decisions it enables 

The walkthrough played out the decisions that usually get made too late, each answered with the room's own numbers: 

Where does the pod fit? Three candidate bays were surveyed, solved on the same physics, and ranked by a stated rule: First, the fewest existing racks with an inlet predicted above 27 °C, the ASHRAE-recommended limit; then the coolest pod inlet; then the least to move. The room has 14 racks and is predicted to exceed that limit today. At Bay C, the best of the three, only one existing rack runs warm, against six at Bay B and fifteen at Bay A. Bay C's warmest pod inlet, predicted at about 22 °C, is the highest of the three but well under the limit. The model is not yet calibrated against measured temperatures, so the 14 racks flagged today are predictions to verify on the floor, not confirmed hot spots; the pod's in-row cooling units appear to be what reduces the count at Bay C.  

 

The pod at bay C, with the placement comparison on screen. 

Can the room cool and power itself? At an assumed 60% of full load, the pod's cooling has about 40 kW to spare, and the room would draw about 451 kW of the 500 kW power ceiling, the site states. Push the pod flat out on a busy day, a what-if that scales the room's load by a sister building's peak-to-mean ratio rather than a forecast, and power becomes the tighter limit: about 560 kW, roughly 60 kW over that ceiling. At this room's own metered peak, it is about 496 kW, just under. The thin margin at the metered peak and the overage in the busy-day case are the open items, and the twin surfaced them before a purchase order. 

The proposed AI pod in the twin. Hardware proposed. 

What if something fails? The twin runs the what-ifs before anything is bought. Lose one of the pod's three cooling units at 60% load, and the pod holds, with every pod inlet predicted under 27 °C. Run it flat out with a unit down, and the cooling comes up 24 kW short. Lose one of a rack's two power feeds, and the other carries the rack within its breaker rating.

One of the pod's three cooling units out of service, shown in red. 

What about a second pod? The twin's answer is to grow the chilled-water plant first. Each pod brings its own three in-row cooling units, and the six together would require about 180 kW of chilled water from a 100 kW chilled-water plant. 

All of this took about two weeks from the first survey visit to a CFD model solved for every scenario, with no hardware moved and no order placed. Once the room is surveyed and modeled, additional scenarios for that room can be run faster. A new room needs its own survey, and timelines depend on site access and data availability. 

How we built the twin of our own room 

The twin models Building 60, Data Center 1, at our Advanced Technology Center in St. Louis (120 racks), which is the room our AI Proving Ground is planned to grow into. It is full, mixed and imperfect, like most rooms our customers run, which is why we chose it. We are starting with NVIDIA DSX in our own data center first, so we can learn on our floor before we bring it to yours. 

The twin of Building 60, Data Center 1, from above, with the CFD airflow prediction shown. 

We surveyed the room over two site visits using a laser, tape, camera and LiDAR, placed every rack where the survey indicated and modeled the room's own equipment from our scans of the real units.  

Numbers you can check, clearly labeled 

Every figure in the twin carries its source. The room's load comes from a month of its own power metering (25 August to 24 September 2026), about 400 kW on the busbars. Five racks carry their own PDU metering; the split of the load across the other 115 is synthetic, bounded by statistics from a sister building because live building-management telemetry was not available for this room. The airflow is a CFD prediction solved in advance on the surveyed room. We plan to compare it with measured inlet temperatures, and until then, the results are planning-grade screening, not yet a basis for a purchase decision. 

Built with NVIDIA 

WWT built this twin with NVIDIA's support, in about three weeks, from the first line of code in mid-September to a booth-ready twin. It runs on NVIDIA Omniverse, uses OpenUSD as its format and follows the DSX Blueprint's conventions for putting together an AI factory twin. Those conventions are what let it grow. A twin built this way can take on more of the DSX platform over time, starting with the partner tools and live signals described below. 

At the Berlin booth, the entire scene was rendered in real time on a single NVIDIA DGX Spark™ and streamed to the presenter's screen. The airflow behind it was solved in advance using an open-source solver. It was a twin of a 120-rack data center running on hardware that fits on a desk. 

Why it matters 

What customers get from this is early design screening and optimization before the purchase order, with the sources and limits of every number visible, and a path to validated reference designs rather than assumptions. It also shows WWT's role as an integrator of NVIDIA AI infrastructure, and it is an early proof point for our Physical AI practice. 

Demo Walkthrough

What's next 

The twin answers planning questions, but it is only the starting point. It sits in DSX Sim, the simulation layer that holds the Omniverse DSX Blueprint. Our next steps build on it. 

  • Partner and ISV integrations. Inside DSX Sim, the next step is partner tools inside the twin, so you can change the layout or the load and watch airflow, temperatures and the electrical one-line recalculate on the spot. Schneider Electric is a leading infrastructure partner in this ecosystem, and we are exploring the possibility of bringing their CFD and ETAP tools into the twin to improve thermal and power analysis.
  • An operational twin of Data Center 1. The same model can then move from planning into operations. We are working with companies whose digital twin planning and operations tools already integrate with DSX, and we are exploring how those capabilities could support the DSX-based solutions we help customers design. Through DSX Exchange, live signals and predictive alarms can run on the model that planned the pod. The twin watches and predicts, while alarms go to the people who run the room, and changes continue through the site change control process.
  • The full DSX platform. From there, we are exploring how to bring more of the DSX platform into our own data centers, moving from simulation to deployment, live operations and ongoing optimization as each step is validated.
  • Factory-built deployment with DSX POD. Deployment is the step beyond the room we modeled, and at GTC Berlin, NVIDIA introduced DSX POD, a reference design for complete NVIDIA Vera Rubin AI factories, built by partners in modular containers and integrated and validated before delivery. NVIDIA says the units connect to a prepared site's power, cooling and networking, reducing deployment from about six months to as little as 14 days, with partner deliveries expected in early 2027. Our twin today models an indoor data hall rather than a containerized site. Extending it to a DSX POD would mean modeling the site's power, cooling and networking readiness before a factory-built system arrives.

DSX POD is one piece of an NVIDIA platform that spans the path from planning to operations. Today, the AI Factory Digital Twin covers the planning step; the other components named here are NVIDIA platform components or planned directions, not WWT-delivered offerings. DSX Sim and the DSX Reference Design help model and validate the AI factory; DSX Air accelerates deployments by providing a simulation of a full AI factory ecosystem; DSX POD provides a factory-built deployment option; DSX Exchange and DSX OS connect operational data and support lifecycle management; and DSX MaxLPS and NVIDIA DSX Flex help optimize power and respond to grid conditions. WWT can help determine which approach fits the site; prepare the required power, cooling and networking; integrate the technology ecosystem and validate the design in our Advanced Technology Center and AI Proving Ground before deployment, working with NVIDIA and the partner ecosystem. 

How WWT can help 

For customers, WWT can bring this connected method to their own environments and align the partner ecosystem required to deliver it—from compute and networking to power, cooling and modular infrastructure. The result is a practical path from simulation to deployment and operations, with WWT helping connect the technologies and partners behind the AI factory. 

See it next at NVIDIA GTC DC. Following its presentation at NVIDIA GTC Berlin, WWT will bring the AI Factory Digital Twin to NVIDIA GTC DC, November 30–December 3, 2026. Join us at Booth #523 to explore what the twin revealed about our own data center—and how the same approach can help screen questions about AI infrastructure placement, power and cooling before equipment is ordered. 

Looking to develop a digital twin today? Connect with WWT's Physical AI team Contact Today 

Technologies