How Stellantis Built a Virtual Factory Digital Twin: 2026 Analysis

When Carlos Tavares stepped down as Stellantis CEO at the end of 2024, he left behind one less-discussed legacy: the most aggressive virtual-factory rollout in the legacy auto industry. The “Dare Forward 2030” plan that Tavares championed quietly funded a multi-year build-out of plant digital twins across Stellantis’ Mirafiori, Melfi, Eisenach and Windsor sites, all stitched to Dassault Systèmes’ 3DEXPERIENCE platform and, increasingly, to NVIDIA Omniverse OVX for real-time rendering. By 2026, the program has stopped being a press-release talking point and started showing up where it matters: in commissioning timelines, OEE numbers and the quarterly cost-per-vehicle slide.

This post is an industry analysis, not a vendor brochure. We will walk through the actual stack Stellantis has assembled, where it integrates with PLM, what changes on the shop floor, the publicly defensible ROI math, and where this kind of program tends to crack. We will also compare it to BMW iFACTORY, Ford’s Heinz Center and Mercedes MO360 so you can place it in context. Where we cite numbers, we cite them. Where we offer opinion, we flag it as such. This is the 2026 read on the virtual factory digital twin automotive 2026 conversation, from a factory-tech perspective rather than a financial one.

Context: why Stellantis went all-in on the virtual factory

Stellantis was born in January 2021 from the merger of PSA and FCA, inheriting 14 brands and roughly 130 plants worldwide. Tavares’ brutal arithmetic was that the group could not afford to digitize all of them at the same rate; he chose instead to standardize on a single platform and ramp it plant-by-plant. The chosen platform was 3DEXPERIENCE on cloud, which both PSA and FCA already used in patches for product engineering. The strategic bet, telegraphed at the March 2022 Dare Forward 2030 investor day, was simple: take the four legacy STLA platforms (Small, Medium, Large, Frame), engineer them once in a shared MBSE backbone, and use a virtual factory to validate every NPI before tooling steel was cut.

Three forces made the urgency real. First, EV ramp pressure: Stellantis committed to 100% BEV sales in Europe by 2030 and 50% in the US, which means the BIW (body-in-white) and battery-tray lines at every plant have to be re-engineered, in many cases without losing more than a handful of weeks of ICE production. Second, the giga-press wave: Stellantis announced its first 6,400-ton IDRA giga-press for the STLA Large platform in 2023, and giga-presses are notoriously brittle to commission — every parameter shift cascades through downstream cells. Third, labor: post-COVID, post-UAW-2023, the cost of an extra week of commissioning labor or a botched ramp went up sharply in both North America and Europe. The virtual factory is, in part, an insurance policy against those costs.

Tavares’ replacements have not walked any of this back. The 2025 capital markets day reiterated 3DEXPERIENCE as the group’s “single source of truth” for product and process. That is the strategic frame the rest of this analysis sits inside.

It is worth noting one piece of organizational scaffolding that often gets overlooked: Stellantis built a centralized “manufacturing engineering digital core” team that sits between the corporate engineering office and the plant manufacturing engineering groups. That team owns the platform standards, the change workflows, the simulation templates and the integration playbooks. Plants do not each pick their own tooling; they consume from a curated catalog. This pattern — central platform team, plant-level implementation teams — is the same one BMW and Mercedes have converged on, and it is in our view the single most important non-technical decision in any virtual-factory program. Without it, every plant rebuilds its own twin from scratch and the investment never compounds.

The virtual factory stack

The Stellantis virtual factory is not one product. It is a layered stack where each tier owns a specific question and hands off curated data to the tier above. The diagram below shows the conceptual layout.

At the top sits the business and strategy layer — program leadership, plant directors, and the NPI program managers who actually consume twin output. They do not touch the platform directly; they consume dashboards and sign-off packages.

Catia Magic: the systems-engineering layer

Catia Magic (the rebranded No Magic / Cameo line that Dassault Systèmes acquired in 2018) is where vehicle and plant requirements live as SysML models. For an OEM the size of Stellantis, this matters because the same battery-tray subsystem ships across multiple brands (Jeep, Dodge, Ram, Peugeot) with different boundary conditions. Catia Magic models those constraints once and traces them down to the CAD authoring layer. In practice, the team uses Cameo Systems Modeler with custom Stellantis profiles to enforce things like “every weld point traced to a structural requirement” and “every robot reach validated against the ergonomic envelope.” This is, in our opinion, the part of the stack that does most of the unglamorous heavy lifting — and the part most other OEMs underinvest in.

DELMIA: process simulation and scheduling

DELMIA is the workhorse. It is where process plans, robotic kinematics, line balance, ergonomics and finite scheduling are simulated. The marquee module is DELMIA Quintiq, the constraint-based scheduling engine Dassault Systèmes acquired in 2014 and which has become the de-facto scheduler at several premium OEMs. Stellantis uses Quintiq to balance the takt across mixed-model lines — for example, running Jeep Compass and a BEV Avenger derivative through the same body shop without breaking either build sequence.

For physical process simulation, DELMIA’s robotic-simulation layer hands off to SIMULIA (Abaqus) for FEA on fixtures and to PowerFLOW for paint-shop CFD where wet-out and overspray matter. The handoffs are managed inside 3DEXPERIENCE, so a robot reach failure ties back to the part design that caused it.

3DEXPERIENCE platform: the substrate

3DEXPERIENCE is the underlying PaaS — multitenant, browser-first, USD-aware as of the 2024 release wave. Stellantis runs it on Dassault Systèmes’ OUTSCALE cloud in Europe and on a US region for North American programs. The platform handles identity, access, versioning, change governance and the “compass” UX that surfaces every tool from a single shell. The strategic significance is that every artifact — a SysML requirement, a CAD part, a DELMIA process plan, a Quintiq schedule — lives in a single graph and is governed by the same change workflow. That is the integration property that everyone else is trying to replicate without buying the whole platform.

PLM integration via ENOVIA

ENOVIA is the PLM backbone inside 3DEXPERIENCE. The integration that matters for the virtual factory is the EBOM-to-MBOM bridge: the engineering BOM authored by product teams gets re-structured into a manufacturing BOM that mirrors the actual build sequence at each plant. ENOVIA’s change-management workflow (engineering change notice, manufacturing change notice) is what keeps the twin and the physical line from drifting out of sync. When an ECN is released, the twin re-runs the affected process plans, Quintiq re-schedules, and the diff is reviewed by the plant team before the change reaches the line. Without that loop, you have a 3D animation, not a digital twin.

The data flow from CAD through PLM into the live virtual factory and out to the shop floor is shown below.

Two opinionated observations. First: the USD scene graph layer (powered by NVIDIA Omniverse OVX, which Dassault Systèmes formally integrated with 3DEXPERIENCE in late 2023) is what makes the twin feel real-time rather than batch. Without it, you have nightly process simulations; with it, you can stream PLC tags into a rendered scene and let an operator scrub timeline. Second: the Quintiq scheduling tier is undersold in public Stellantis communication. It is, arguably, what makes the difference between “we have a twin” and “we changed the production schedule because the twin told us to.”

What changes on the shop floor

The shop-floor experience is where the program either earns the investment or becomes shelfware. Stellantis’ rollout, observable in public press visits to Mirafiori and Melfi in 2024-2025, has three changes worth flagging.

Operator HMIs now carry a live twin view. At the body shop, andon stations show not just a red/green status but a small 3D view of the offending cell, rendered from the same USD scene the simulation engineers used. When a weld station throws a fault, the operator sees the robot pose, the part orientation and the predicted root cause (e.g., “fixture clamp 3 out of tolerance, see ECN-2026-0142”) rather than a four-digit fault code. The 3D scene is pushed down from the 3DEXPERIENCE cloud via the plant DMZ and cached at an edge historian, which buffers about 15 seconds so the HMI does not stall if the WAN hiccups.

Work instructions are animated. DELMIA-authored work instructions, exported to a lightweight 3D viewer, replaced the PDF-based instruction books at several stations during the STLA Large body shop conversion. Operators report shorter time-to-competence on new variants; supervisors report fewer “I didn’t know about that ECN” miscommunications. The flip side: the instructions only stay accurate if the twin stays in sync with the released MBOM, which puts pressure on the ENOVIA change workflow.

AR overlays on tablets are used in quality cells for torque verification and bin-pick confirmation. This is a smaller deployment, and frankly the part most prone to over-promise — Stellantis has not made big public claims about AR adoption, and we would advise treating any vendor case study on this with skepticism.

The underlying integration pattern is standard: MQTT broker plus OPC UA aggregator in the plant DMZ, edge historian for short-term retention, and a tag-bridge that maps PLC names to twin entity IDs. The robots (mostly ABB and Comau in Stellantis plants, with some Fanuc and Kuka in former FCA sites) and the AGV fleet (a mix of MiR and SEER) all stream joint state and position into the broker. The Giga press cell — high-value, high-risk — gets its own dedicated tag path and its own redundancy story.

The cultural change is harder than the technical one. Plant teams that previously trusted their own paper SOPs now have to trust a model authored elsewhere. The change-management investment to make that stick is, in our view, larger than the software bill.

One under-discussed shop-floor effect is on the maintenance organization. With the twin streaming live telemetry from the Giga press, BIW robots and battery-tray assembly cells, maintenance teams at Mirafiori have moved from a strict preventive-maintenance calendar to a hybrid model: PM intervals are still set, but the twin can pull a service window forward (or push it back) based on observed vibration, motor current and cycle-count drift. According to Dassault Systèmes’ 2024 manufacturing summit case studies, this hybrid scheduling alone accounts for a meaningful share of the downtime reduction numbers cited later in this post. The catch is that the maintenance MES (Stellantis uses Siemens Opcenter at most converted FCA sites and a legacy PSA system at a handful of older European plants) has to expose work-order APIs the twin can call — that integration is not trivial and is one of the longest-running streams of work on the program.

A second shop-floor change worth flagging is in quality. Vision systems at the body shop now publish their inspection results into the same MQTT broker that feeds the twin, which means a quality failure can be replayed inside the 3D scene with the exact robot pose and part orientation that produced it. This shortens root-cause analysis from days to hours on the kinds of intermittent defects (e.g., weld spatter under specific thermal conditions) that previously took multiple shifts to reproduce. The quality engineers we have spoken to off the record describe this as the single most valuable day-to-day capability the twin has delivered — more useful, in their view, than the headline commissioning-time numbers the executives cite.

ROI math: what the numbers actually look like

The ROI conversation around the Stellantis virtual factory tends to swing between two unhelpful extremes — the vendor pitch deck (50% commissioning time reduction!) and the skeptic’s shrug (none of this is provable). The honest answer sits in the middle and looks roughly like the breakdown shown below.

On the investment side, a single plant rollout — say, converting the Mirafiori 500e line — runs in the range of USD 40-60 million over five years, dominated by software licenses (3DEXPERIENCE plus DELMIA, roughly USD 18-25M), integration and systems engineering (USD 12-18M), edge plus OT plumbing (USD 8-12M) and change management plus training (USD 4-6M). Those numbers come from triangulating Dassault Systèmes investor commentary, public Stellantis CapEx disclosures by plant, and published case-study ranges from comparable OEM rollouts. Treat them as order-of-magnitude, not point estimates.

On the returns side, the publicly defensible bands are:

• Commissioning time: 30-50% reduction. Dassault Systèmes has claimed up to 50% in DELMIA case studies (cited at the 3DEXPERIENCE World 2024 keynote); the realistic average across OEM rollouts is closer to 30%, per Automotive News reporting on the BMW iFACTORY program in 2024.
• NPI cycle compression: 20-25%. From design freeze to job-one. This is the figure Tavares cited at the 2022 investor day for the STLA Medium platform.
• Unplanned downtime: 15-30%. Driven by twin-guided predictive maintenance and the andon-with-twin pattern described above. The lower end of that range is the conservative read; the upper end requires the twin to actually push live PLC data and not just nightly batches.
• First-time-right rate: +8-15 percentage points. Virtual buy-off of fixtures and tooling reduces physical rework.
• Scrap and rework: 10-18% reduction. Aligned with the first-time-right gain.

The payback profile typically looks like net cash out in years 0-1, break-even in year 2 (usually tied to one or two vehicle programs), and compounding returns from year 3 onward as the platform investment is amortized across successive NPIs. This is consistent with what Dassault Systèmes reports for its top automotive accounts, and consistent with what BMW disclosed for its Debrecen and Munich rollouts.

There is a second-order ROI lever that is harder to quantify but real: option value. A platform-based virtual factory means a future vehicle program can be launched into existing twin infrastructure rather than building bespoke simulation each time. Stellantis has used this argument internally to justify reusing the Mirafiori twin scaffolding for the STLA Small platform that will follow. If you believe the EV-and-software-defined-vehicle wave will keep accelerating product variants, the value of being able to spin up a new twin in months rather than years is meaningful. We would not put a number on it in a public post, but it is the kind of figure that gets defended in capital allocation meetings.

Our editorial view: the ROI is real but it is back-loaded, and it depends almost entirely on platform reuse across multiple vehicle programs. A single-program ROI calculation will always look marginal. Boards evaluating this kind of program should insist on a multi-program, five-year window before they sign the cheque — and on a quarterly reporting cadence that tracks the operational metrics (commissioning days saved, downtime hours avoided, NPI weeks compressed) rather than the activity metrics (cells modeled, ECNs processed) that vendors tend to default to.

Trade-offs and gotchas

Three honest caveats every OEM considering this path should sit with.

Vendor lock-in is the elephant. The integration depth between Catia Magic, DELMIA, ENOVIA and 3DEXPERIENCE is also a moat. Once your EBOM, MBOM, process plans, schedules and ergonomic models all live in 3DEXPERIENCE, the cost of moving is staggering. Some OEMs (notably Volkswagen with its Industrial Cloud) have explicitly bet on a more heterogenous stack to avoid this. There is no free option here — heterogenous stacks pay an integration tax instead.

CAD-to-sim translations remain brittle. The promise is a clean Catia-to-DELMIA-to-Omniverse round trip via USD. The reality, as of the 2024 release wave, is that complex assemblies still occasionally lose constraints, kinematic joints or material properties on the way through. Engineering teams report needing a “twin maintainer” role whose job is to babysit these translations, especially on giga-press cells where every parameter matters. NVIDIA’s USD-on-the-wire promise is closing this gap, but it is not closed.

Organizational change is the largest line item that nobody budgets for. A virtual factory only delivers if the plant team trusts the model enough to act on it. That trust is built through painful, public wins where the twin called something the experienced foreman did not, and the twin was right. It is also broken in a single bad call. Stellantis has been relatively disciplined about phased rollouts and pilot cells; not every OEM is. The training and change-management line item should be 2-3x what the procurement team initially proposes.

A fourth, smaller point: the energy footprint of the cloud-rendered twin (Omniverse OVX nodes, the 3DEXPERIENCE compute) is non-trivial and should be on the sustainability ledger. We have not seen any OEM publish numbers, and we suspect that is not an accident.

A fifth caveat, important for boards and procurement teams: the SaaS license model that 3DEXPERIENCE runs on means the OpEx line grows with usage. As more cells, more lines and more plants are brought into the twin, the recurring spend climbs in a way that traditional perpetual-license CAD shops are not used to budgeting for. Several large Stellantis suppliers we have spoken with privately report being surprised by year-three license escalation; this is not a Dassault-specific issue (Siemens and PTC price the same way), but it is a budgeting trap if the finance team has not modeled the ramp explicitly.

Finally, an interoperability caveat: the USD format Omniverse uses is open, but the round-trip from Catia-native parametric geometry into a USD scene and back is lossy in subtle ways that matter for engineering review. A part that looks identical in the rendered twin may have lost the constraint that drove its outer profile in CAD. Teams have learned to treat the twin as a visualization-and-simulation surface, with the CAD model in 3DEXPERIENCE remaining the canonical engineering source of truth. This is a discipline, not a technology, fix.

What other OEMs are doing

Stellantis is not alone. The 2026 competitive picture has three other reference programs worth tracking.

BMW iFACTORY is arguably the most-cited peer. BMW’s program runs on NVIDIA Omniverse as the primary visualization layer, with Siemens Tecnomatix for process simulation and a custom MES tier. The Debrecen plant, which opened in 2025 to build the Neue Klasse BEVs, was the first BMW plant designed twin-first. BMW has been more public about Omniverse, less public about its PLM backbone.

Ford Heinz Center in Dearborn is the smaller, more research-flavored bet. Ford uses a mix of Siemens NX, Tecnomatix and AVEVA for process and operations, with substantial in-house tooling. The Heinz Center is widely reported to be where Ford prototypes plant changes for the F-150 Lightning and Mustang Mach-E, but Ford has not standardized the way Stellantis has.

Mercedes MO360 is the most operations-tilted of the four. MO360 is heavy on real-time MES and quality systems, with Siemens and SAP underneath, and integrates Microsoft Azure Digital Twins for the cross-plant data fabric. Mercedes’ program is, arguably, the most mature in terms of live shop-floor adoption; it is less ambitious on the upstream engineering integration than Stellantis or BMW.

Our take: Stellantis has bought the deepest, most integrated stack of the four. BMW has the prettiest visualization story. Mercedes has the most production-proven operations layer. Ford is hedging. None of them are wrong; they are different bets