In 2025 Meta’s Reality Labs division reported an operating loss of $19.2 billion, and the word metaverse more or less disappeared from consumer marketing. In the same period, Siemens built an entire CES keynote around simulated factories running on the same underlying technology and did not need to defend the concept once. The gap between those two facts is the story, and it is not really a secret. It is a vocabulary change.
Key takeaways
- Meta’s Reality Labs lost $19.2bn in 2025, against $17.7bn in 2024, and over $4bn again in Q1 2026.
- Siemens launched a Digital Twin Composer at CES on 6 January 2026, shipping mid-2026.
- BMW’s Debrecen plant was the group’s first facility planned and validated entirely virtually.
- The strongest published figures are vendor-reported and have not been independently audited.
The consumer version failed while the industrial version shipped
The financial contrast is stark enough to do most of the explaining. Meta’s Reality Labs unit reported a $19.2 billion operating loss for 2025, worse than the $17.7 billion recorded in 2024, and CNBC reported a further loss of more than $4 billion in the first quarter of 2026. Cumulative losses since the division was broken out in 2020 now run past $80 billion on the reported figures.
Industrial spending on the same underlying technology went the other way, and quietly. There was no rebrand announcement, because there was no brand to defend. Engineering teams had been building simulation models for decades under names like virtual commissioning and plant simulation. What changed was not the ambition but the rendering: real-time, photorealistic, physically accurate, and connected to live sensor data rather than to a static CAD file.
That is why the sector never had to argue about whether the metaverse was real. It was already using the parts that worked and had never bought the parts that did not.
What has actually been built
BMW: a plant that existed virtually before it existed
BMW’s site at Debrecen in Hungary was the group’s first facility planned and validated completely virtually, using NVIDIA Omniverse Enterprise. Presenting the work at NVIDIA’s GTC conference in 2023, BMW production board member Milan Nedeljković described running virtual production more than two years before series production was due to start. Layouts, robot cells and logistics flows were tested in simulation, with suppliers working on the same live model. The plant was scheduled to open in 2025, and BMW has since rolled the approach out across its production network.
Siemens: the tooling, sold as a product
At CES on 6 January 2026, Siemens introduced Digital Twin Composer, which combines its own digital twin data with simulation built on NVIDIA Omniverse libraries and real-time engineering data. It is due on the Siemens Xcelerator Marketplace in mid-2026. The same announcement covered nine industrial copilots across Teamcenter, Polarion and Opcenter, and a collaboration with Meta to bring industrial AI to shop-floor workers through Ray-Ban AI glasses, which is a neat illustration of where consumer hardware ended up finding a job.
NVIDIA: the layer everybody rents
The platform position matters more than any single deployment. NVIDIA has announced digital twin collaborations with a partner list that reads like an industrial index: Belden, Caterpillar, Foxconn, Lucid Motors, Mercedes-Benz, Omron, Toyota, TSMC and Wistron among them. Whoever owns the 3D simulation layer sits underneath every one of those projects, which is a considerably better business than selling headsets.
| Consumer metaverse | Industrial digital twin | |
|---|---|---|
| What is sold | Presence and entertainment | Avoided cost and reduced commissioning risk |
| Who pays | The platform, hoping users follow | The capital expenditure budget of the plant |
| Success metric | Daily active users | Throughput, capex variance, ramp-up time |
| Hardware needed | A headset per user | Usually a screen and a GPU cluster |
The numbers that turn it into a board decision
The figures that circulate are worth quoting precisely, and worth flagging as vendor-supplied. In its CES 2026 material, Siemens cited PepsiCo’s work digitising US manufacturing and warehouse facilities: a 20% increase in throughput on the initial deployment, identification of up to 90% of potential issues before any physical modification, and reductions in capital expenditure of 10% to 15%, with close to 100% of design validation performed virtually.
Those numbers come from a supplier press release describing its own customer, which is the weakest link in the evidence chain even when nothing is exaggerated. No independent audit of industrial digital twin returns at this scale has been published. What can be said with more confidence is the mechanism: catching a clash between a robot cell and a conveyor in simulation costs a week of engineering time, and catching it after installation costs a plant shutdown. That asymmetry does not need a study to be obvious to anyone who has commissioned a line.
What is still unproven
Three things deserve scepticism before this gets written up as settled.
- Interoperability. Shared 3D formats such as OpenUSD are converging, but a twin built inside one vendor’s ecosystem still carries real switching costs.
- Model drift. A digital twin is only as good as the data keeping it current. Plants change constantly, and an unmaintained twin becomes an expensive diagram.
- Attribution. When a site adopts simulation, new sensors and new analytics in the same programme, isolating what the twin contributed is genuinely hard.
None of that invalidates the direction of travel. It does mean that a 20% throughput figure from one deployment is a data point, not a benchmark, and should not be pasted into a business case for a different plant.
Questions this usually raises
Is a digital twin the same thing as a simulation?
No. A simulation models behaviour offline against assumptions. A digital twin is connected to the physical asset and updated with its live data, so it reflects the machine as it is now rather than as it was designed.
Does any of this need a headset?
Rarely. Most industrial work happens on ordinary screens, with the compute sitting in a GPU cluster. Headsets and smart glasses appear mainly in training and maintenance guidance, which is a narrow and well-defined use.
Is this only viable for very large manufacturers?
At full plant scale, mostly yes today, because the modelling effort is significant. Cell-level and machine-level twins are a much smaller commitment and are where mid-sized manufacturers typically start. The broader picture of what separates a genuinely connected plant from a merely automated one is something we mapped out in our look at what makes a factory truly smart.
What happens to the term “metaverse” now?
In industry it is already being replaced by “industrial AI” in vendor messaging, which is the third name for a broadly continuous set of capabilities. The technology outlasts its labels, and reading the underlying capability rather than the branding is the only reliable defence against buying the same thing twice.
The honest version of the headline is this: industrial giants are not secretly powering the metaverse. They are openly building the thing the metaverse was supposed to be, under a name that survives a procurement committee, and the reason it worked for them is that they could put a number on it.
Want the operational payoff rather than the platform story?
The same live-data plumbing that feeds a digital twin also underpins the most proven industrial AI use case to date.
Sources: Meta Platforms quarterly and full-year results for 2024 and 2025 and CNBC reporting of Q1 2026 Reality Labs figures (29 April 2026); Siemens press release, CES 2026, 6 January 2026, including the PepsiCo figures cited therein; BMW Group press material on virtual production at Plant Debrecen, NVIDIA GTC 2023, and NVIDIA Omniverse Enterprise case material. Performance figures for digital twin deployments are reported by the vendors involved and have not been independently audited. Updated August 2026.
