Nvidia and Schneider test a digital twin to rehearse AI data center upgrades

An AI data center in Portugal is getting a dress rehearsal. Nvidia and Schneider Electric are testing a digital twin that connects AI workloads to models of a site's power and cooling, so operators can see what an infrastructure change would do before they make it. Shane Snider, a senior news writer at TechTarget, reported the pilot on Oct. 21, 2026.

Each partner brings one piece. AVEVA supplies industrial engineering and operations software. ETAP models and simulates electrical power systems. Nscale is the AI infrastructure and cloud provider, and Start Campus is developing the data center campus in Portugal.

The twin is built for a question every facilities engineer planning denser GPU deployments has to answer sooner or later. Can the existing electrical distribution and liquid cooling carry the next generation of hardware? Or does it mean a costly redesign, a late go-live, or capacity left stranded? Stephen Sopko, who leads semiconductor and deep-tech research at HyperFrame Research, told TechTarget the payoff in the near term is lower commissioning risk and faster deployment. Efficiency gains come later, if they come.

How the SIN01 twin is wired

The system combines Schneider's energy management technology with AVEVA's industrial software, ETAP's electrical engineering tools and third-party technology. Facility engineering data and operational data feed into power and cooling simulations. Those simulations are tied to what the AI workloads need. The practical use is testing higher rack density: an operator can model what a denser layout does to electrical distribution and liquid cooling before anyone moves a rack.

The companies say they want the system to cover the whole AI factory lifecycle, from design through operations, and to make it easier to bring in equipment and models from different vendors. Sopko sees that interoperability as the useful part: it could let facility engineering tools talk to AI infrastructure models instead of living in separate silos.

The idea has been around for a while. On March 18, 2025, Facilities Dive's Joe Burns reported that Schneider and ETAP had released digital twins for designing and simulating the power needs of AI factories. Those twins were built on ETAP's integration of Nvidia Omniverse and drew on mechanical, thermal, networking and electrical inputs.

For teams writing specs for future halls, this suggests vendor twins built on this stack are still at pilot stage, not a mature product. Specs should be written that way.

What CoreWeave's DSX Air shows by comparison

CoreWeave already takes a simulation-first approach. TechTarget notes that the company uses Nvidia's DSX Air, a cloud-based platform, to simulate AI factory environments and validate configurations before the physical systems arrive. DSX Air models compute, networking, storage and software orchestration. Public disclosures don't show CoreWeave using it to control power and cooling in a live facility.

Set side by side, the SIN01 pilot moves the rehearsal from the server stack into the building: switchgear-level electrical behavior and the cooling loops that keep dense racks within their limits. Sopko's point applies to both cases: rehearsing a configuration before the hardware lands can surface problems early, which could shorten validation and cut physical rework.

There's a case for a smaller gap, though. Sopko describes full bidirectional closed-loop control, where the twin acts on the facility as well as reading from it, as a longer-term goal. That suggests the SIN01 twin is, for now, closer to a planning tool than a control system. Operators still make the decisions. That could mean a smaller payoff in the near term, with no change yet to who is responsible on the floor.

Sized for Rubin-class density

The campus figures explain why a twin is worth building here. TechTarget reports that the companies haven't said how many Rubin GPUs are in place at SIN01 today or when the rest will arrive. A build-out of this size, done in phases, is where a model of the next phase earns its keep: every hall added is a chance to find a power or cooling mismatch in software instead of on site.

Schneider's cooling hardware fits the same design pattern. Motivair by Schneider Electric's MCDU-70 coolant distribution unit is rated at 2.5 MW, and six of them can provide 4+2 redundancy for a 10 MW DSX Blueprint deployment.

Nothing in the reporting says SIN01 uses the MCDU-70. What the pairing suggests is that the 10 MW blueprint block is becoming a shared unit of planning, with cooling hardware specified around it. That matters most for operators buying into the Nvidia reference design. Teams with mixed vendors or non-reference halls would get less from that alignment and more from the multi-vendor interoperability the companies describe.

The scorecard SIN01 still has to post

Sopko set out a specific test for whether a twin improves AI computing efficiency. Compare time to first token, commissioning defects, power usage effectiveness, thermal events and measured compute output against a baseline that uses comparable workloads and hardware. Simulation results alone don't count. The baseline is what turns a good-looking model into evidence.

The companies haven't yet published SIN01 results on token throughput, energy use, cooling efficiency or deployment time. They also haven't released validation comparing the twin's predictions with measured operating conditions, which is the comparison that shows whether the model can be trusted.

As Sopko put it, the nearer-term value is lower commissioning risk and faster deployment. Efficiency gains still have to be shown against a baseline using comparable workloads and hardware.

The next thing to watch at SIN01 is that prediction-versus-measurement comparison, plus a Rubin deployment schedule, which the companies have not yet provided. Fewer commissioning defects measured against a comparable baseline would be one sign that the twin is doing its job.