Chapter 9: Cosmax Physical AI Strategy — Lab, Factory, Dark Factory, Micro-Factory

9.1 Where Cosmax Stands — A Physical AI Readiness Self-Assessment

Cosmax is the world's #1 cosmetics ODM (14–18% Korean market share, 7.1% global) and a K-beauty ODM leader with 1,100 dedicated researchers ^[8]. Beyond Incheon, Pyeongtaek, Shanghai, Indonesia, and Ohio, Cosmax established its first European base in Italy in 2025, and a $43.9M new factory in Thailand is scheduled for September 2026 startup [Cosmax, 2025a]. Mapping this asset base onto the three-stage evolution model from Chapter 5 makes Cosmax's coordinates clear.

Coordinates against the Chapter 5 model — between Stage 1.5 and the start of Stage 2

Stage 1 (simple automation) is in place. Automated filling machines, capping machines, labelers, and conveyors are deployed in every major plant, and a few lines have introduced camera-based QC. While the automotive Stage 1 reaches "100% body welding by robots," Cosmax sits at "60–70% automation of core processes" — non-standard containers, high-mix small-batch production, and brand-specific label changes still leave 30–40% in human hands.

Stage 2 (digital twin) is partial and local. PLM, MES, and ERP are running, but the OpenUSD/Omniverse integration that BMW Debrecen demonstrated in 2024–2025 with a virtual SOP two years ahead of physical SOP is absent ^[3]. Simulation stops at the level of individual lines; no work has begun on integrating factory–lab–logistics–brand data onto a single USD graph.

Stage 3 (autonomy) is unexpectedly advanced — but on the lab side. The CAI (Cosmetic AI) Research Center, founded in 2021, already runs three pillars: a fragrance prediction AI (8,600+ molecule database), a color matching AI, and an ingredient optimization AI ^[4]. The HelloBiome partnership has materialized into a B2B2C platform that connects microbiome analysis → AI ingredient recommendation → ODM production end to end [Cosmax, 2025b]. In other words, Cosmax's autonomy began on the R&D side, not on the manufacturing side — a path different from the other industries in this book.

Synthesizing these, Cosmax sits at Stage 1.5 in the factory and Stage 2.5 in the lab. This is an asymmetric structure where the late 1990s of automotive coexists with the early 2020s of pharma inside a single company.

Three strengths already in hand

First, the depth of research personnel. The 1,100 researchers form a multi-disciplinary pool spanning chemistry, fragrance, microbiome, clinical, formulation, and packaging engineering. If the biggest barrier to Physical AI adoption is "combining domain expertise with AI capability" (Chapter 8), Cosmax already has the domain pool. What is missing is the data, robotics, and edge infrastructure to layer on top.

Second, diversity of data. Data from thousands of brands, tens of thousands of SKUs, and dozens of categories (skincare, makeup, hair, body) accumulates daily — a breadth no single-brand manufacturer can ever match. Structurally, this mirrors how Foxconn holds EMS data from eight or more OEMs (Chapter 5).

Third, a track record of AI deployment. The fact that the CAI Research Center and HelloBiome are already operating means the organization has lived experience of deploying AI in R&D. The gap between a company that has done this once and a company that has never done it is wider than any single ROI number.

Three things still missing

First, factory data standardization. The Incheon, Pyeongtaek, Shanghai, Indonesia, and Ohio plants do not share a common PLC standard, MES schema, or OpenUSD asset library. Building a single global digital twin requires this prerequisite (the Cosmax-specific version of the "legacy system integration" barrier diagnosed in Chapter 8).

Second, operational experience with robots. The cumulative hours of running AMRs, collaborative robots, and vision-picking robots in production are limited. The know-how Amazon accumulated while operating one million robots took 13 years (Chapter 6). Cosmax cannot leap from year zero overnight, but the gradual accumulation has to start now.

Third, edge AI infrastructure. Just as Audi could not match line speed (one body per second) with cloud inference and instead achieved 25× acceleration by running on Jetson-class edge GPUs (Chapter 5), the cosmetics filling and packaging line that processes hundreds of units per hour requires edge inference. Today's edge GPU and edge vision infrastructure remain at pilot scale.

The conclusion of this self-assessment is clear — Cosmax is a latecomer with the latecomer's advantage. It enters the era when 30 years of tools, standards, and ecosystems built by automotive, semiconductor, and logistics industries can be partially absorbed in three years. This is the essence of what NVIDIA calls "Reindustrialization," and for Cosmax it is the chance to define the ODM version of that wave.

9.2 Lab Automation — Physical AI for Formulation Development

The integration of lab (formulation development) and factory (production) under one roof is Cosmax's most distinctive asset. The first place to amplify that asset with Physical AI is, without question, lab automation. If BMW captured "30% reduction in production planning cost" through digital twin in Chapter 5, the equivalent ROI window for a cosmetics ODM is the formulation development cycle.

Where it stands today — CAI and HelloBiome as the starting line

Cosmax's R&D automation is not a blank page. Since 2021, the CAI Research Center has been running three engines ^[4]:

• Color matching AI: Recommends optimal color from thousands of combinations against a customer's skin tone. Compresses development time for foundation, lipstick, and eyeshadow.
• Fragrance prediction AI: Predicts olfactory properties from a database of 8,600+ molecules. Compresses fragrance R&D trial-and-error into data.
• Ingredient optimization AI: Multi-objective recommendation across efficacy, safety, and stability.

On top of this, the HelloBiome partnership (2025) has set a one-stop B2B2C platform in motion: microbiome analysis → AI ingredient recommendation → ODM production [Cosmax, 2025b]. This is not a research toy. It is the hardware of a new business model — when a brand has no infrastructure to build its own microbiome line, Cosmax's platform is what they rent.

Next step — applying the Self-Driving Lab concept

The next leap in lab Physical AI is the Self-Driving Lab (SDL). AstraZeneca's High-Throughput Experimentation (HTE) platform for drug-candidate discovery — robotic dispensers, automated synthesizers, automated analyzers, LIMS, and a machine-learning recommendation engine wired into a single closed loop — has accelerated experimentation more than 10× ^[1]. BASF's chemistry R&D, with the same approach cutting new-material development time to one-third, sits in the same lineage ^[2].

Translating this to cosmetics R&D yields a concrete blueprint:

1. Robotic dispensing workstation: Microliter-scale automated mixing of liquid bases, pigments, fragrances, and active ingredients. A combination of Opentrons- or Tecan-class liquid handlers with viscosity-tolerant nozzles tuned for cosmetics.
2. Automated analytical module: Automated measurement of pH, viscosity, particle size, color, and stability, streaming results into LIMS in real time.
3. Closed learning loop: A machine-learning model recommends the next experimental combination. One full cycle iterates autonomously for 12–24 hours without human intervention.
4. Linkage to CAI: The fragrance, color matching, and ingredient optimization AIs feed their recommendations directly into the next experimental combination, enabling true self-driving — "AI proposes the hypothesis; the robot tests it."

Digital twin formulas — virtual formulation testing

Even with maximally fast physical experiments, stability evaluation still requires 4–12 weeks. The way around this constraint is the digital twin formula — representing the physicochemical properties of a formulation as a simulation model and predicting stability, viscosity, emulsion robustness, and sensorial spread virtually. The "AI-driven formulation" reports L'Oréal and P&G released in 2024–2025 point in this direction (Chapter 7). Quantitative targets for digital twin formulas:

• 50% fewer physical experiments (virtual pre-screening)
• 30% shorter stability evaluation cycles (combining accelerated stability data with simulation)
• New product development cycle from 6 months to 3–4 months (the cosmetics version of the one-third reduction BASF achieved in chemistry)

Execution priorities — what to start first

Three steps in sequence:

1. 0–6 months: Build one HTE pilot cell. Inside the CAI Research Center, install one liquid handler and one automated analyzer, integrated with the existing fragrance prediction AI. Investment ₩1.0–1.5B; KPI: "automated cycles per day."
2. 6–18 months: 3–5 HTE cells + Digital Twin Formula v1. One cell per category — skincare, makeup, hair. Release Digital Twin Formula v1 limited to stability prediction. Investment ₩5–8B; KPI: "reduction in new product development cycle time."
3. 18–36 months: Self-Driving Lab v1 in operation. Closed-loop automation. Human researchers focus on hypothesis design and result interpretation. Investment ₩15–25B; KPI: "new formulations discovered per researcher."

The keystone of this roadmap is exploiting the fact that "CAI already runs." The AI is in place; what is missing is the robotic hand to execute the experiments the AI directs. That gap is the 18–36-month investment target.

9.3 Factory Automation — Priorities for Filling, Packaging, and Quality Inspection

If the lab leads in autonomy, the factory remains at Stage 1.5. Mapping the "Physical AI barriers in manual-labor industries" diagnosed in Chapter 8 onto Cosmax factories surfaces three processes where automation ROI is clearest — vision inspection, filling-line assistance, and palletization/material movement.

Priority 1 — AI vision inspection (the Audi model adapted to cosmetics)

Audi's case in Chapter 5 — automating five million daily weld-point inspections in the body shop, achieving 25× edge-inference acceleration and "less than one hour of training on twenty samples" — transplants almost directly to a cosmetics ODM. Defect inspection items on a cosmetics line are concrete:

• Label print errors: Broken characters, color drift, misalignment — detectable in milliseconds via vision.
• Underfill/overfill: Combine weight sensors with vision-based fill-level inspection.
• Loose or incomplete capping: Vision plus torque sensors.
• Foreign material contamination: Vision for transparent containers, X-ray or acoustics for opaque containers.
• Cosmetic defects: Container scratches, dents, color flaws.

For each item, validated algorithmic patterns already exist. Implementation requires per-line one or two Jetson-class edge GPUs, four to eight cameras, and an Inspekto-style few-shot learning toolkit. The Audi-Siemens benchmark of ₩50–100M investment per line, 12–18 month payback is likely to come even faster on a cosmetics line — although the cost of a single defect is smaller than in automotive, new-product launch frequency is much higher and the cumulative effect is larger.

Vision inspection is priority #1 for three reasons. First, risk is low — it adds a camera next to the line, it does not stop the line. Second, data accumulates immediately — hundreds of thousands of products pass the camera daily, so training data is generated automatically. Third, visible ROI — defect rate, rework rate, and customer-complaint metrics move within a single quarter.

Priority 2 — Filling-line assistance robots (the Sparrow principle adapted to cosmetics)

Amazon Sparrow uses suction cups plus a seven-axis gripper to recognize and pick 65% of two hundred million SKUs (Chapter 6). On a cosmetics line, standard containers account for 60–70%, so a Sparrow-style vision-picking cell transplants directly. Three insertion points:

• Bulk → filling-machine feed: Sorting and feeding empty containers of non-standard shape — currently a simple repetitive human task.
• Insert (sub-component) placement: Automatic placement of pumps, nozzles, and dispensers at exact positions.
• Small-pack box packing: Precise SKU-aware arrangement of finished products into boxes.

The limit is also clear — the remaining 30–40% (irregular containers, fragile materials, low-volume SKUs) still need human hands (Amazon's "remaining 35%" lesson, Chapter 6). The deployment strategy therefore is "standard-container lines first, irregular-container lines later." Per-line investment ₩200–500M; payback 18–30 months.

Priority 3 — AMR deployment (low risk, high visibility)

Deploying AMRs for in-plant transport of materials, semi-finished goods, and finished products is the lowest-risk and highest-visibility of the three priorities. The cases of Locus Robotics passing six billion cumulative picks and DHL Supply Chain operating 8,000 collaborative robots with 80% reductions in worker training time and 30–180% productivity gains (Chapter 6) translate directly to Cosmax factories.

Deployment strategy — gradual introduction at 30–50 units per facility. Year one: 30 units in Incheon. Year two: simultaneous expansion to Pyeongtaek and Shanghai. Begin with "one worker, one robot" pairing; after operational know-how accumulates, expand to "one worker, three robots." Per-facility investment ₩3–8B, payback 24–36 months. Two by-products of AMR deployment matter especially: (a) factory data infrastructure gets paved naturally (AMRs only move when location, inventory, and route data are standardized), and (b) workers accumulate experience working alongside robots.

The ODM-specific constraint — reconciling brand-formulation security with AI training

The ODM model carries one constraint that ordinary manufacturing does not — the security of brand formulations. Cosmax produces under contract on the basis of confidential formulations and process know-how from thousands of brands. Training AI on this data carries two risks: (a) leakage of one brand's formulation into another brand's output, and (b) IP-attribution disputes over models trained on combined data.

The solution is a two-part combination. First, a federated learning structure — each brand's data resides in an isolated partition; only model weights are integrated. Second, explicit contractual consent — adding "anonymized process data may be used for ODM-common model training" as a standard clause in new ODM contracts. Both must work for the diversity data created by 1,100 researchers to convert into AI training assets. This is not a compliance issue — it is the core IP strategy of the ODM foundry model.

9.4 Dark Factory Roadmap — When and Where to Begin

The dark factory — a fully autonomous plant operable without lighting — is the symbolic terminus of industrial automation. Foxconn's "smart factory" lines, Mujin's automated warehouses, and JD.com's Zhilang system (Chapter 6) all point in this direction. Two questions face Cosmax when it considers a dark factory: what is realistically possible, and where to begin.

Cosmax's realistic target — start with a "dark line"

In a cosmetics ODM, fully darkening an entire factory does not deliver short-term ROI. The reason traces to the "five barriers to Physical AI in manual-labor industries" diagnosed in Chapter 8 — material variability, SKU diversity, cleanliness requirements, short batches, and frequency of brand changes. When all five are present, the cost of automating the last 5% of manual work overwhelms the labor cost of leaving it to humans (the cosmetics version of the Tesla 2018 lesson, Chapter 5).

So the realistic target is a "dark line" — automating one line completely, not the whole factory. The candidate line should satisfy the following:

• High-volume single-SKU line (low-volume high-mix lines excluded)
• Standard containers and standard labels (minimal handling of irregularities)
• Long-term contracted brand (low frequency of line reconfiguration)
• Continuous operation possible (the ROI of dark operation comes from 24/7 utilization)

Across Cosmax Incheon and Pyeongtaek, only 1–2 lines satisfy all four conditions. Designating one of these as the first dark line and completing it by 2030 is the realistic target.

The Thailand new factory — Physical AI from day one

The $43.9M Thailand new factory scheduled for September 2026 startup gives Cosmax a "clean-sheet design opportunity" [Cosmax, 2025c]. While the existing Korean and Chinese plants require retrofitting Physical AI on top of legacy assets, Thailand can include OpenUSD digital twin, standardized PLCs, and edge GPU infrastructure in the design phase from the start. This is exactly the pattern by which BMW Debrecen achieved a virtual SOP two years ahead of physical SOP (Chapter 5).

A Thailand-factory design checklist:

• [ ] OpenUSD asset library — Express every line, equipment, warehouse, and logistics flow in USD. The reference point for future global integration.
• [ ] Standardized PLC stack — Single Siemens or Rockwell standard. Compatible data communication with other plants.
• [ ] Edge GPU infrastructure — Pre-wire one or two Jetson-class edges per line.
• [ ] AMR-friendly layout — Aisle width, floor finish, and charging-station locations all designed AMR-friendly from the start.
• [ ] Data-lake connection — Real-time synchronization with the Korea HQ data lake.
• [ ] First dark-line candidate selection — Designate one line in Thailand as a dark-line candidate from day one.

When this checklist works, the Thailand factory becomes a showroom of "Korean factories five years from now." It also turns into a directly usable asset in global brand sales — "we have dark-line operating experience" becomes the ODM differentiation message.

Korean legacy plants — a three-stage transition roadmap

Incheon and Pyeongtaek are retrofits and need a different approach. A three-stage transition over four years (2026–2030):

Stage 1 (2026–2027): Data standardization + vision QC + AMR deployment. Unify all line data into a single MES schema, deploy vision QC on five to eight lines, and bring in 30–50 AMRs. By the end of this stage, the plant is "a factory where data flows." Cumulative investment ₩20–30B.

Stage 2 (2027–2028): Digital twin v1 + filling-assistance robots. Build a partial Omniverse digital twin of the Incheon plant; install two to three vision-picking cells. New and expansion lines designed with twin-first principles. Cumulative investment ₩30–50B.

Stage 3 (2028–2030): First dark line live. Convert one line at the Thailand plant to 24/7 unmanned operation. Extend the same model to one or two lines at Incheon. Cumulative investment ₩50–80B.

When does the dark factory ROI actually pencil out

The threshold where the dark factory crosses from "romance" into "capital decision" is the simultaneous satisfaction of four conditions:

1. 24/7 operation — Night and weekend utilization above 60%; otherwise labor savings cannot justify ROI.
2. Single-SKU long runs — A line must run the same product for at least six months for reconfiguration costs to drop.
3. Standardized inputs — Containers, labels, and caps must be standardized so that vision and robotic algorithms stabilize over years.
4. Yield ≥ 99.5% — Defects require human intervention; therefore dark operation presupposes very high first-pass yield.

Across all of Cosmax, at most 5–10% of lines satisfy all four. The dark line is therefore a "premium line" in the Cosmax factory, not a "standard line." This precise positioning is the safeguard that keeps dark-factory investment from drifting into fantasy.

9.5 Custom Micro-Factories — The Future of Personalized Cosmetics

If the dark factory is the "automation terminus of mass production," the micro-factory is the opposite — the "automation starting point of small-batch custom production." That K-beauty's next ten-year frontier is personalization (custom foundation, custom skincare, custom fragrance) is already confirmed in market data ^[8]. The question of section 9.5 is what position Cosmax should take on that frontier.

The micro-factory concept — small cells, fast turnover

A micro-factory is a production unit with the following characteristics:

• Small footprint: 50–200 m². One-tenth to one-fifth of a typical factory line.
• High flexibility: Hundreds of SKUs switched on hourly basis within the same cell.
• High turnover: Batch sizes of dozens to thousands (one-hundredth to one-thousandth of a typical line).
• Direct B2C linkage: Orders from stores, apps, and D2C channels feed directly into the cell.

The precedent is Shiseido VOYAGER (Chapter 7). Inside or adjacent to the store, a mini-factory receives the customer's skin-measurement data and produces and delivers customized skincare on the spot. Similar attempts followed in L'Oréal Perso and P&G Opte.

Cosmax ODM application — a foundry for D2C brands

Where Shiseido runs micro-factories under a single-brand model, Cosmax can take a different path — becoming the micro-factory foundry of dozens of D2C brands. Concretely:

1. Shared micro-factory cells: A single cell switches between custom products of multiple brands by time slot or by day.
2. Brand SaaS interface: An API that routes orders from each brand's app or store directly into Cosmax micro-factories.
3. Personalization formulation engine: The CAI Research Center's fragrance, color, and ingredient AIs serve as the back-end brain of personalized formulation.
4. Small-batch fast delivery: A 24–72 hour goal from order to delivery.

The keystone of this model is the division "Cosmax provides infrastructure; brands own the customer interface." Structurally, it is the same as TSMC playing foundry to every fabless chip designer in semiconductors. And for this structure to work, both the Self-Driving Lab in 9.2 and the vision/filling robots in 9.3 are required — the micro-factory is the union of R&D automation and factory automation, not a separate domain.

The risk — collision with existing mass-production lines

The micro-factory carries two internal collision risks.

First, resource competition. The same researchers and the same automation budget get split between mass-production efficiency improvements and new micro-factory build-outs. Without explicit separation by management (separate budget line, separate accountability), micro-factories drift to chronic deprioritization.

Second, data contamination. The small-batch, high-variance data of micro-factories, mixed with the stable mass-production data, distorts statistical models. Clear labeling separation between micro-factory and mass-production at the data layer is essential.

The remedy is "separation of the micro-factory business unit." Distinct P&L, distinct KPIs, distinct sales channel, separate from the existing ODM business unit. The same structure under which Shiseido VOYAGER ran independently of HQ R&D.

Why now — the size of the opportunity

Three reasons Cosmax should enter the micro-factory now:

1. The explosion of K-beauty D2C brands: Hundreds of new Instagram- and TikTok-born brands appear every year. Their common pain point is "the inflection between small-batch trial and mass production." The micro-factory smooths that inflection.
2. Acceleration of personalization: As AI skin analysis, DNA analysis, and microbiome analysis go mainstream, demand for "one person, one formulation" rises rapidly.
3. First-mover effect: When the world's #1 ODM enters this space first, the K-beauty D2C ecosystem grows on top of Cosmax infrastructure. Network effects emerge that latecomers cannot easily catch.

9.6 Short, Mid, and Long-Term Execution Roadmap

This section compresses every analysis in this chapter into a single execution grid. The time axis splits into short term (2026–2027), mid term (2028–2029), and long term (2030+); the four areas of 9.2–9.5 (R&D automation, factory automation, dark factory, micro-factory) populate each band.

Short term (2026–2027) — Validate and lay foundations

The message of this stage is "five things shallow," not "one thing deep." Place a small footprint in every area to find out where the real ROI lies, with data. The total investment of ₩17–25B is roughly 1% of Cosmax revenue — a recoverable bet even if it fails.

Mid term (2028–2029) — Scale and differentiate

The message of this stage is "decide which is real." Concentrate investment on the two or three areas whose ROI is validated from the small footprints of the short term; reduce or maintain the rest. Total investment ₩55–81B.

Long term (2030+) — Position and external recognition

• First dark line live (2030) — Achieve 24/7 unmanned operation on one line at the Thailand plant.
• Prepare WEF Lighthouse Network application — Global Lighthouse Factory certification is the strongest differentiation lever in ODM sales. Curate the data, cases, and benchmarks needed for application.
• Spin off or rebrand the micro-factory business unit — Run with a P&L and KPIs distinct from the Cosmax core.
• Contribute Cosmax datasets to global standards — Donate Cosmax data to a public benchmark for "cosmetics ODM vision inspection" or "custom-formulation AI" → secure academic and industry standard influence.

The message of this stage is "we define the category called Physical-AI ODM." On top of the current position of "world's #1 cosmetics ODM," layer "world's #1 Physical-AI ODM." When this positioning works, the K-beauty D2C explosion of the next decade grows on Cosmax infrastructure.

Investment priorities — where to spend first

Cumulative three-stage investment of roughly ₩150–200B (about 6–8% of revenue, distributed over four to five years) is recommended in the following weights:

• Factory automation and digital twin (40%) — Largest scale, most validated ROI. Vision QC + AMR + Omniverse at the core.
• R&D automation (25%) — Highest differentiation potential. HTE and SDL layered on top of CAI.
• Micro-factory (20%) — Highest new-growth potential. Tied to the K-beauty D2C explosion.
• Dark factory (10%) — Most symbolic, but ROI lands last. A measured bet.
• People, organization, external partnerships (5%) — Strategic partnerships with NVIDIA, Siemens, and specialist robotics firms; recruitment and development of talent.

These weights are a starting point, not an absolute rule. Adjust by ±20% in the mid term based on short-term validation results.

Three questions the executive team must decide now

Three questions for the executive boardroom that demand immediate answers:

Question 1 — Should Cosmax create a Physical-AI executive officer (CDO or CPAIO)? If not, an accountability gap opens between R&D and Production. Every roadmap in this chapter requires cooperation between these two divisions, and without a single accountable head, 9.2 (R&D) and 9.3 (factory) drift apart. Recommendation: create the role. Direct executive line, integrated budget, influence over both divisions.

Question 2 — Will the Thailand new factory be "a clone of the Korean plants" or "a showroom of the future five years out"? The former is safe and ordinary; the latter is risky and differentiated. The 9.4 checklist works only if the latter. Recommendation: the latter. Build the operating team around new digital talent, not around legacy Korean plant operators.

Question 3 — Should the D2C brand micro-factory business sit inside the core or split out as a separate business unit? Inside the core, resources get drained into mass production; split out, short-term inefficiencies appear. The 9.5 analysis points to separation. Recommendation: separate business unit. Distinct P&L, distinct sales channel, sharing only infrastructure with the core.

These three decisions will shape the next five years of Cosmax. Tools, data, and capital ultimately move only when people and organizations move (the BCG-Alpega conclusion of Chapter 6). Physical AI is not a game of tools but a game of organization. That is the single conclusion every analysis in this chapter eventually arrives at.

This survey ends here. Across nine chapters it traced a path from the platform strategies of NVIDIA, Siemens, and Rockwell, through the evolution of automotive, semiconductors, and logistics, to the execution roadmap for Cosmax. But the utility of this book begins not when it lands on a bookshelf but the moment the next executive boardroom asks "what do we decide today." This book hopes to have drawn enough auxiliary lines for that question.

References

1. AstraZeneca (2024). High-Throughput Experimentation Platform for Drug Discovery: 10x Acceleration. AstraZeneca R&D Blog. https://www.astrazeneca.com/r-d/our-approach.html
2. BASF (2024). AI-Driven Materials Innovation: One-Third Faster Development Cycles. BASF Press Release / Innovation Report. https://www.basf.com/global/en/who-we-are/innovation.html
3. BMW Group and NVIDIA (2025). BMW Group Scales Virtual Factory with NVIDIA Omniverse. BMW Press Release (NVIDIA GTC Paris). https://www.press.bmwgroup.com/global/article/detail/T0450699EN/bmw-group-scales-virtual-factory
4. Cosmax (2021). Cosmax CAI Research Center: Color AI, Fragrance AI, Ingredient Optimization. Cosmax Press Release. https://www.cosmax.com/research/cai
5. Cosmax (2025a). Cosmax Europe Entry: Italy as First European Production Base. Cosmax Press Release. https://www.cosmax.com/news/europe
6. Cosmax (2025b). Cosmax HelloBiome: AI-Powered Microbiome Cosmetics Platform. Cosmax Press Release / HelloBiome Partnership. https://www.cosmax.com/research/hellobiome
7. Cosmax (2025c). Cosmax Strategic Overview: Global ODM Beauty Leader. Cosmax IR / Industry Analysis. https://www.cosmax.com/ir/overview
8. Mordor Intelligence (2025). Global Cosmetics ODM Market 2025-2035 Outlook. Market Research Report. https://www.mordorintelligence.com/industry-reports/cosmetics-odm-market