Part I: The Physical AI Manufacturing Revolution — Platforms and Strategy

Chapter 3: Siemens, Rockwell, ABB — Legacy Industrial Automation Embraces Physical AI

Written: 2026-04-28 Last updated: 2026-04-28

3.1 OT/IT Convergence — The Incumbents' Dilemma

While NVIDIA repositioned itself as "the company that builds the entire factory virtually," what about the OT (Operational Technology) incumbents who have been wiring those factories for the last fifty years? Siemens (founded 1847, Germany), Rockwell Automation (1903, USA), and ABB (1883, Switzerland) each carry over a century of industrial-automation assets — they actually built and sold the cables under the concrete, the PLCs in the cabinets, and the robot arms beside the lines. Their installed base is worth tens of trillions of dollars.

These companies face a sharp dilemma. Decades of installed infrastructure is both an asset and a liability. While NVIDIA rewrites the standard for green-field factories with OpenUSD, Omniverse, and Isaac, OT vendors cannot rip and replace the PLCs and SCADA systems running tens of thousands of factories overnight. Yet if they fail to compete with the new entrants, they slide down the value chain into a "shop-floor body, NVIDIA brain" division of labor. The path the three companies chose in 2025–2026 was not head-on competition but deep cooperation — each signed a multi-year partnership with NVIDIA and repositioned its domain assets (field data, control SW, robot firmware) as the industrial-domain adapter on top of the NVIDIA stack.

This chapter examines how the three companies play the same NVIDIA partnership card differently. Siemens emphasizes digital enterprise + industrial metaverse, Rockwell focuses on edge generative AI + Omniverse digital twins, and ABB attacks the sim-to-real gap for robots. From the vantage of a cosmetics ODM such as Cosmax, the most direct decision variable is what AI capabilities the PLC and MES vendors already inside their factories (mostly Siemens or Rockwell) are starting to offer.

3.2 Siemens — The Digital Enterprise Strategy

Co-building an Industrial AI Operating System with NVIDIA

At CES 2026, Siemens and NVIDIA announced the joint construction of an "Industrial AI Operating System" [8]. Siemens contributes its hundreds of industrial-AI experts and PLM/CAE/EDA/MES software assets, while NVIDIA contributes AI infrastructure, simulation libraries, and blueprints. The combination puts the entire industrial value chain — design → engineering → manufacturing → operations → supply chain — onto a single operating system. The first blueprint is Siemens's own electronics factory in Erlangen, Germany, which in 2026 becomes the world's first fully AI-driven, adaptive manufacturing site. NVIDIA NIM and Nemotron get directly integrated into Siemens EDA software, and the entire Siemens simulation portfolio becomes GPU-accelerated through CUDA-X.

The significance of this announcement is that the Omniverse, Isaac, and Jetson stack from Chapter 2 becomes one body with the SW that actually runs the floor. Past NVIDIA digital twins risked staying at the "pretty visualization" level, but once they connect directly with Siemens TIA Portal, Teamcenter, and Opcenter MES, PLC-logic changes get verified inside the digital twin and MES work orders flow immediately into Omniverse simulations. This is exactly the kind of asset a new entrant cannot create alone, and it is the strongest hand Siemens brought to the partnership.

Digital Twin Composer and the PepsiCo Case

At the same CES 2026, the three-party announcement by Siemens, NVIDIA, and PepsiCo introduced Digital Twin Composer, the most concretely quantified case of the industrial metaverse to date [10]. A PepsiCo beverage plant was fully digital-twinned with Digital Twin Composer + NVIDIA Omniverse + computer vision. Results:

  • Up to 90% of potential quality issues identified ahead of time — before any physical change is made, AI agents simulate, test, and refine to catch the risks of line modifications.
  • Throughput up 20% — purely from simulation-driven optimization on the same line.
  • Capex down 10–15% — over-engineering eliminated during line additions and expansions.
  • Near-100% design-validation coverage — virtually no change reaches production without prior validation.

These numbers matter to a cosmetics ODM because PepsiCo's beverage process — "high-speed, high-mix filling + labeling + packaging" — is structurally similar to a Cosmax filling/packaging line. The biggest pain point on lines that run many SKUs in short batches is the cost of changeover, and Digital Twin Composer's "catch 90% of issues before the change" is a tool that decisively reduces the change-management cost of an ODM.

Mendix Low-Code Lets Floor Staff Build AI Apps

Another Siemens card is Mendix — the AI-assisted low-code application development platform Siemens acquired in 2018 [4]. Mendix is an embedded environment of Siemens Xcelerator that lets workflow, inventory, production-schedule, quality-management, SCM, predictive-maintenance, and data-visualization apps be built without code. In 2025 Gartner named Mendix a Leader in its Magic Quadrant for Enterprise Low-Code Application Platforms, and Siemens Energy alone runs more than 200 internal apps on Mendix. Opcenter Execution MES also integrates with Mendix so that adaptive MES apps can be built without coding.

This matters to a mid-sized ODM because the biggest bottleneck in adopting AI is shortage of IT staff. The NVIDIA stack we saw in Chapter 2 is powerful but demands ML engineers. Mendix, by contrast, opens a path for floor-side OT engineers and production staff to pull SCADA data themselves and bolt on dashboards and simple AI inferences. It is a partial answer to the familiar mid-sized Korean manufacturer's complaint of "AI sounds great, but we have no people for it."

Mendix is not omnipotent: complex custom logic eventually requires Java/Python fallback, and the Mendix DSL itself has a learning curve. But combined with native Siemens AI tools — Inspekto visual quality inspection (under one hour of training on 20 samples), Audi body-shop weld inspection (5 million welds/day, 25× edge-inference acceleration), Senseye predictive-maintenance generative AI [9] — the surface area for "AI you can touch on the floor" expands dramatically. The Amberg plant already runs 1,000+ sensors with AI predictive maintenance and has cut maintenance cost by 25%.

3.3 Rockwell Automation — Manufacturing-Specific AI

2025 Smart Manufacturing Report — 95% Plan to Invest in AI

Rockwell's annual State of Smart Manufacturing Report, in its tenth edition (2025), is a global survey of more than 1,500 manufacturers across 17 countries [Rockwell Automation, 2025a]. The single most powerful number is that "95% of respondents are investing in or planning to invest in AI/ML over the next five years." This is no longer simple awareness — budgets and organizations are already moving. Other key findings:

  • Quality management is the #1 AI use case for the second year in a row — 50% of 2025 respondents plan to apply AI to quality. This aligns exactly with the core KPIs of a cosmetics ODM (defect rate, rework rate).
  • Generative and causal AI investment up 12% YoY — moving beyond simple computer-vision AI into LLMs and causal reasoning.
  • Cybersecurity ranks #2 external risk — 49% plan to use AI for cybersecurity in 2025 (up from 40% in 2024). A new threat that emerges as OT networks connect to IT.
  • "Ability to apply AI" jumped from 10% to ~50% YoY in skill importance — perception of the core competency for floor staff grew fivefold in one year.

The message this survey sends to a mid-sized Korean manufacturer like Cosmax is unambiguous. "Whether to invest in AI" is no longer a differentiator. When 95% are already moving in that direction, differentiation shifts to "where, how fast, and how well."

NVIDIA Nemotron Nano and a Manufacturing-Specific LLM

At Automation Fair 2025 in Chicago in November 2025, Rockwell and NVIDIA announced edge-based generative AI [7]. The core is integrating an SLM (small language model) based on NVIDIA Nemotron-Nano-9B-v2 into FactoryTalk Design Studio. This SLM can run anywhere along the following spectrum:

  • HMI panels (operator panels) — operators can query and operate line state in natural language.
  • Industrial appliances (industrial box PCs) — standard industrial computers installed beside the line.
  • Desktop IDEs — PLC-coding assistance on engineer PCs.
  • Servers / private cloud — site-wide unified inference.
  • Air-gapped deployment — same behavior in environments cut off from the internet.

That last point — air-gapped deployment — is the key. Cloud LLMs (GPT, Claude) seen in Chapter 2 are highly capable but require sending data out. In industries where recipes and process parameters are the core trade secret — cosmetics, pharma, semiconductors — the data going to OpenAI or Anthropic servers is in itself unacceptable. The Nemotron-Nano 9B + edge box combination keeps model weights inside the site and processes inference, prediction, and operator response without external communication. Rockwell reports superior reasoning, predictability, and responsiveness compared with other SLMs.

The limits are clear too. A 9B-parameter model cannot match GPT-4 or Claude 4 on complex multi-step reasoning. It needs domain fine-tuning to perform well in industrial settings. But for narrow tasks — "auto-suggest PLC-logic edits," "summarize alarm messages in natural language," "search maintenance manuals" — this size suffices, and it is almost the only path that simultaneously satisfies data-security and immediacy requirements.

Omniverse Integration Extends the Digital Twin

Rockwell's other NVIDIA card is the integration of Emulate3D with the NVIDIA Omniverse API (completed in early 2025) [6]. Emulate3D is the digital-twin SW Rockwell acquired and now offers, and the combination of OpenUSD interoperability and NVIDIA RTX rendering enabled factory-scale dynamic digital twins. Emulate3D Factory Test, unveiled at GTC 2025, demonstrated virtual controls testing and Factory Acceptance Testing (FAT) automation — moving the standard process of validating PLC, SCADA, and robot code together before physical commissioning into the digital twin.

The six explicitly named target industries are CPG, food and beverage, life sciences, semiconductors, automotive, and material handling. CPG and life sciences are the adjacent industries to a cosmetics ODM. One caveat is the Rockwell-SW lock-in risk — Emulate3D is powerful but optimized for Rockwell controllers and Studio 5000, so plants with mixed-vendor PLCs incur additional integration cost. In an environment like Cosmax with diverse line vendors, "which vendor is the PLC on this line?" becomes the first variable in choosing between Rockwell and Siemens tools.

3.4 ABB — Robots and HyperReality

RobotStudio HyperReality — 40% Cost Cut, 99% Fidelity

Of the three companies, ABB's strongest domain is robots. ABB robots have held top global share in tasks such as automotive body welding and painting, food and beverage packaging, and chemical and pharmaceutical dispensing. The 2026 announcement of RobotStudio HyperReality is the result of combining that asset with NVIDIA Omniverse [2]:

  • NVIDIA Omniverse libraries integrated into RobotStudio — ray tracing, physics engines, and USD-based assets enter the ABB robot simulation environment directly.
  • Virtual controllers running the same firmware as real robots — the virtual robot executes the actual robot firmware, so a program written in simulation behaves identically on the physical robot.
  • 99% sim-to-real fidelity — virtual-to-real accuracy reaches 99%.
  • Up to 40% reduction in development and production cost — through eliminating physical prototypes and virtualizing iterative testing.
  • Setup and commissioning cut by 80% — new-line cell setup and validation drops to one fifth.
  • Time-to-market cut by 50% — overall duration to bring up a new model or product line is halved.

Release is set for the second half of 2026, with Foxconn as the first pilot customer. One limitation is that it is restricted to ABB's own controllers and robots — cross-vendor generalization on KUKA, FANUC, or Yaskawa robots is not validated.

What makes the 99% fidelity and 40% cost reduction different from generic NVIDIA Omniverse / Isaac sim-to-real numbers is the decisive factor of "vendor's own firmware." Generic simulators approximate dynamics and sensors with general models, but ABB knows the exact firmware behavior of its own robots and can therefore eliminate the largest source of the sim-to-real gap at the root. This is a kind of domain asset that is very hard for new entrants to catch up with.

#

References

  1. ABB Robotics and AMS (2025). ABB-AMS Automotive Manufacturing Outlook Survey 2025. Automotive Manufacturing Solutions / ABB Robotics. https://www.automotivemanufacturingsolutions.com/reports/amsabb-automotive-manufacturing-outlook-survey-2025-cost-pressures-bite-but-ev-and-hybrid-optimism-persists/2588253
  2. ABB Robotics and NVIDIA (2026). Closing the Sim-to-Real Gap: How ABB's RobotStudio HyperReality Enables Industrial-Scale Physical AI. ABB Press Release. https://new.abb.com/news/detail/134178/wbstr-closing-the-sim-to-real-gap-how-abbs-robotstudior-hyperreality-enables-industrial-scale-physical-ai
  3. Hyundai Motor Group and Boston Dynamics (2026). Hyundai Motor Group Announces AI Robotics Strategy to Lead Human-Centered Robotics Era at CES 2026. Hyundai Press / CES 2026. https://www.hyundai.com/worldwide/en/newsroom/detail/hyundai-motor-group-announces-ai-robotics-strategy-to-lead-human-centered-robotics-era-at-ces-2026-0000001100
  4. Mendix and Siemens (2025). Mendix Low-Code Platform within Siemens Xcelerator (Smart Manufacturing). Mendix / Siemens Press. https://www.mendix.com/siemens/
  5. Rockwell Automation (2025a). 10th Annual State of Smart Manufacturing Report 2025. Rockwell Automation Industry Report. https://www.rockwellautomation.com/content/dam/rockwell-automation/documents/pdf/campaigns/state-of-smart-2025-cpg/INFO-BR029C-EN-P.pdf
  6. Rockwell Automation and NVIDIA (2024). Rockwell Automation Brings Autonomous Operations to Life Using NVIDIA Omniverse. Rockwell Press Release / NVIDIA Case Study. https://www.rockwellautomation.com/en-us/company/news/press-releases/Rockwell-Automation-Brings-Autonomous-Operations-to-Life-Using-NVIDIA-Omniverse.html
  7. Rockwell Automation and NVIDIA (2025). Rockwell Automation to Advance Industrial Intelligence Through Edge-Based Generative AI with NVIDIA Nemotron. Rockwell Press Release (Automation Fair 2025). https://www.rockwellautomation.com/en-us/company/news/press-releases/rockwell-automation-to-advance-industrial-intelligence-through-e.html
  8. Siemens and NVIDIA (2026). Siemens and NVIDIA Expand Partnership to Build the Industrial AI Operating System. NVIDIA Newsroom / CES 2026. https://nvidianews.nvidia.com/news/siemens-and-nvidia-expand-partnership-industrial-ai-operating-system
  9. Siemens and NVIDIA Blog (2025). Siemens Makes Factory Floors Smarter With Industrial AI. NVIDIA Blog. https://blogs.nvidia.com/blog/siemens-industrial-ai/
  10. Siemens, NVIDIA, and PepsiCo (2026). Siemens Brings Industrial Metaverse to Life with Digital Twin Composer (PepsiCo Collaboration). Siemens News / CES 2026. https://news.siemens.com/en-us/digital-twin-composer-ces-2026/