Chapter 1: What Is Physical AI — The New Manufacturing Paradigm of 2025

1.1 The Limits of Digital AI — Why Physical AI Now

Over the past five years, AI has made two enormous leaps. The first was language AI, embodied by ChatGPT (2022). The second was multimodal generative AI, embodied by Stable Diffusion, Sora, and Veo. What both stages share is that the output is digital bits — text tokens, pixels, audio samples — computed in the cloud and ending on a screen.

The problem is that most of the value humanity creates lives off-screen. WEF estimates the global manufacturing and logistics sectors at over $50 trillion in 2025 ^[1]. That value is created by physical actions: moving containers, weighing powders, capping bottles, scanning QR codes. Digital AI can assist these tasks (for example, generating work orders), but cannot perform them. Human hands still fill the last meter.

Physical AI is the attempt to let AI close that last meter directly. Yann LeCun has long argued that LLMs, as "autoregressive predictors of discrete tokens," cannot represent the continuous dynamics of the physical world ^[2]. His AMI Labs, formally launched in March 2026, raised $1.03B in seed at a $3.5B pre-money valuation — Europe's largest seed round ever — and committed that capital not to LLMs but to JEPA (Joint Embedding Predictive Architecture) based continuous world models. In other words, some of the very people who built the first LLM paradigm are already moving capital into the next one.

Cosmax Implication: The core value of cosmetics ODM lies in the precision and speed of physical operations like filling, packaging, and inspection. Digital AI cannot reach this value layer directly. Examining Physical AI is not about "chasing trends" — it is about "raising the automation level of the core business by one full step."

1.2 Definition and Components of Physical AI

There is no single standard definition of Physical AI yet. But three authoritative definitions, all published in Q1 2025, converge.

Distilling the three, Physical AI is the integration of three components:

1. Perception: The ability to convert physical-world states into digital representations via cameras, LiDAR, tactile sensors, microphones, and so on. 2D/3D vision, point-cloud processing, and multimodal fusion are central.
2. Reasoning: The ability to decide the next action from perceived state. This layer is occupied by Large Language Models (LLMs) and Vision-Language-Action (VLA) models, with World Foundation Models like NVIDIA Cosmos rising as physics-aware reasoning engines ^[3].
3. Action: The ability to translate decisions into motors, grippers, and mobile bases that change the real environment. Industrial manipulators, AMRs (autonomous mobile robots), humanoids, and increasingly bimanual collaborative robots fill this layer.

These three layers operate on top of three infrastructure axes: digital twin, edge AI, and cloud. The digital twin supplies simulation data, edge AI handles millisecond-scale control loops, and the cloud takes care of fleet learning and model updates. In Chapter 2 we will see that NVIDIA's Omniverse·Isaac·Jetson stack maps exactly onto these three infrastructure axes.

Cosmax Implication: Treating "AI adoption" as a single product purchase is the failure mode. Physical AI is a system designed across a 3 × 2 matrix of six cells. Even applying it to one filling line forces six choices at once: perception cameras, reasoning models, action robots, digital twin, edge PC, and cloud training.

1.3 Why Manufacturing Became Physical AI's First Battlefield

Physical AI's potential markets span homes, healthcare, defense, and retail, but as of 2025 the fastest adoption is in manufacturing and logistics. Three reasons explain why.

First, structured environments. Unlike homes, factories have fixed layouts, controlled lighting, and finite SKU sets. This means a "narrow data distribution" — the very condition AI training needs. Because identical tasks repeat on identical lines, hundreds of millions of demonstrations can accumulate. Amazon already operates more than 1 million robots across 300 fulfillment centers and trains fleet-level models on the data they generate ^[4].

Second, the size of the prize. WEF·BCG report that Amazon's deployment of the Sequoia, Sparrow, and Proteus systems delivered a 25% lift in fulfillment efficiency, and that mobile robot fleets improved travel efficiency by 10% ^[4]. The $50 trillion figure Huang cited is for manufacturing plus logistics alone ^[1]. Even a 1% efficiency gain is $500 billion. At that multiplier, capital expenditure of hundreds of millions of won per robot becomes rational.

Third, labor-force pressure. Manufacturing labor shortages are intensifying simultaneously in the U.S., Germany, Japan, and Korea. The WEF·BCG AI Agents report finds that 82% of executives plan to deploy AI agents within 1–3 years ^[4]. Adoption intent has already crossed the threshold.

The difference between legacy automation and Physical AI is learning. Legacy PLC-based automation repeats sequences explicitly coded by engineers; an SKU change requires reprogramming. Physical AI adapts to new tasks via demonstration, simulation, and reinforcement learning. For an ODM like Cosmax that runs short batches across thousands of SKUs, this adaptability is decisive.

Cosmax Implication: Cosmetics ODM is high-mix low-volume, so legacy automation has poor ROI. As the retraining cost curve of Physical AI flattens, the per-SKU automation break-even point collapses simultaneously. The proposition "our products are too varied to automate" is likely to break between 2026 and 2028.

1.4 The Year 2025 — What Has Actually Changed

Industrial robots existed in 2015. Computer vision existed in 2020. So why does the word "paradigm" attach only in 2025? Because four technical inflection points arrived in the same window.

(1) Sim-to-Real maturity. GPU-accelerated simulators like Omniverse, Isaac Sim, and MuJoCo XLA can now run physics-faithful simulation at tens of thousands of frames per second. Policies trained in simulation are transferring zero-shot or few-shot to real robots. This drops data-collection cost by more than 100×.

(2) LLM × robotics. Vision-Language-Action (VLA) models that emerged in 2024 (RT-2, OpenVLA family) translate natural-language instructions ("pick up that red cap") directly into motor commands. The moment the task-instruction interface shifts from code to natural language, non-programmers can define robot tasks.

(3) Rise of world foundation models. Cosmos, announced by NVIDIA at CES on January 6, 2025, is a World Foundation Model that takes text, images, video, and robot sensor data as input and generates physics-grounded video ^[3]. It was released under a permissive license similar to Apache-2.0, and at least 14 early adopters joined — 1X, Figure, Agility, XPENG, and others ^[1]. A reasoning capability was added in March 2025, and at CoRL 2025 in September the platform gained data generation through text, image, and video prompts — meaning quarterly updates ^[3]. Just as autonomous driving leapt forward with ImageNet, Physical AI is poised to leap forward with Cosmos-class foundation models.

(4) Authoritative endorsement. The WEF formally listed Physical AI in its Top 10 Emerging Technologies of 2025 ^[4]. Other entries on the same list include structural battery composites, advanced nuclear, and generative watermarking. This signals that consulting, policy, and capital markets have begun pointing at the same category with the same vocabulary. The WEF·BCG Frontier Technologies in Industrial Operations report formalizes the field as a "two-way taxonomy of virtual AI agents + embodied AI agents" ^[4].

Cosmax Implication: Missing any one of the four inflections would have justified delay. But simulation, VLA, world models, and standardization aligned in the same quarter. The gap between a company that runs a PoC in 2025–2026 and one that starts in 2027 is not a one-year gap — it is a difference in the cumulative data-deficit curve.

1.5 How to Read This Survey — A Guide for Cosmax Executives

The book is organized into nine chapters. Part I (Ch 1–4) covers the definition and platform landscape of Physical AI itself: NVIDIA (Ch 2), Siemens·Rockwell·ABB (Ch 3), and the consulting view from McKinsey·BCG·Deloitte·PwC (Ch 4). Part II (Ch 5–8) tracks adoption case studies by industry: pharma and chemicals (Ch 5), cosmetics and beauty (Ch 6), food and beverage (Ch 7), and apparel and consumer goods (Ch 8). Part III (Ch 9) closes with strategic recommendations specific to Cosmax.

Two traps are easy to fall into while reading.

Trap 1: "That's a different industry from ours." Reading about automotive OEMs or Amazon, it is tempting to conclude "we are cosmetics, so it's different." But the value dynamics of Physical AI are determined by process type, not industry. Filling, labeling, inspection, and palletizing are essentially identical whether the substrate is an auto part or a lipstick. Chapter 6 will examine how L'Oréal, P&G, and Shiseido leveraged this process homogeneity to import technology validated in automotive.

Trap 2: "After the technology matures more." Physical AI is on the steep part of the learning curve. The later you start, the more data and operational know-how gaps accumulate. While Amazon holds the operating data of more than 1 million robots, late entrants begin from 10,000-robot data. Time does not close that gap.

Three core questions to keep in mind:

1. When do you start — does the common roadmap of PoC in 2026, one line in 2027, multi-plant rollout in 2028 fit Cosmax?
2. Where do you start — among filling, packaging, inspection, and lab automation, which process sits highest on Cosmax's ROI curve?
3. How do you start — NVIDIA Omniverse-led digital twin first, Siemens-led OT integration first, or consulting-led roadmap first?

The next chapter begins with NVIDIA's first integrated answer to all three questions.

References

1. Huang, Jensen (2025). CES 2025 Keynote: Physical AI and the Next Wave of Computing. NVIDIA Blog. https://blogs.nvidia.com/blog/ces-2025-jensen-huang/
2. LeCun, Yann (2026). Yann LeCun's AMI Labs Raises $1.03B to Build World Models. TechCrunch. https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/
3. NVIDIA Research (2025). Cosmos World Foundation Model Platform for Physical AI. arXiv:2501.03575. https://arxiv.org/abs/2501.03575
4. WEF (2025). Top 10 Emerging Technologies of 2025. World Economic Forum, Frontiers, Dubai Future Foundation. https://reports.weforum.org/docs/WEF_Top_10_Emerging_Technologies_of_2025.pdf
5. WEF and BCG (2025a). Physical AI: Powering the New Age of Industrial Operations. World Economic Forum White Paper. https://reports.weforum.org/docs/WEF_Physical_AI_Powering_the_New_Age_of_Industrial_Operations_2025.pdf
6. WEF and BCG (2025b). Frontier Technologies in Industrial Operations: The Rise of AI Agents. World Economic Forum Report. https://reports.weforum.org/docs/WEF_Frontier_Technologies_in_Industrial_Operations_2025.pdf