Physical AI on Korea Invest Insights

Korea Humanoid Robot Value Chain: Samsung, Hyundai, LG Timelines and the Profitable Parts Suppliers

Mon, 11 May 2026 18:00:00 +0900

This is a Korea AI hardware value-chain piece. For broader context, start with the Korean AI Companies Hub and the KOSDAQ complete guide. For related public-market angles, see the Samsung Electro-Mechanics coverage on AI infrastructure re-rating and the Hyundai Mobis EV and robotics deep dive.

🔄 Robotics series follow-ups (2026-05-12): · Part 2 — SPG vs. Halla Cast (reducers vs. structural parts) · Part 3 — Robotis vs. Rainbow Robotics (humanoid whole-robot compare) · Korea Humanoid / Robotics Investment Hub

The useful question in Korean robotics is not whether robots are a promising theme. The market has already assigned a large amount of value to that answer. The more practical question is simpler: where does revenue show up first?

Samsung Electronics is talking about AI-driven factories and physical AI. Hyundai Motor Group has a concrete route through Boston Dynamics and the Atlas humanoid. LG Electronics is building around actuators, service robots and physical-AI components. The direction is real.

The public-market reality is less clean. Some pure robotics stocks already trade as if the 2028-2030 commercialization curve has been solved. Several have small current revenue bases and limited or negative operating profit. Parts suppliers, by contrast, can add robotics demand on top of existing businesses in reducers, actuators, cameras, batteries and precision manufacturing.

This post is not an attempt to hype or dismiss the sector. It is a map of the distance between expectation and earnings.

TL;DR

Samsung, Hyundai and LG are all moving at the same time. Samsung is framing robotics around AI-driven manufacturing, Hyundai has Boston Dynamics and Atlas deployment plans, and LG is pushing actuators and service robots.
The first verification window is 2H26 through 2028. Samsung Electro-Mechanics humanoid camera modules, Hyundai’s RMAC and Atlas deployment, LG actuator production and HL Mando’s 2027-2028 actuator line buildout sit in this window.
Humanoid robots are a parts problem as much as an AI problem. Reducers, actuators, camera modules, sensors, batteries, grippers and assembly quality determine cost and performance.
Many Korean robotics listings still have earnings far behind valuation. Pure-play robotics names may have large market caps relative to current sales.
Revenue may appear first in parts suppliers. SPG, Robotis, HL Mando, Samsung Electro-Mechanics, LG Innotek and Hyundai Mobis have existing manufacturing businesses that can absorb robotics demand as incremental revenue.
Atlas and Optimus supply chains include many estimates. Treat development exposure, confirmed orders and recognized revenue as three different stages.

1. Not All Robots Are the Same

“Robot” is too broad a word for investment work. A welding arm in a factory, a collaborative robot next to a worker, a warehouse robot on wheels, a restaurant serving robot, a surgical robot and a two-legged humanoid all sit under the same label. Their economics are different.

Category	Main Use	Industry Stage	Korean Listed Examples
Industrial robots	Welding, painting, assembly	Mature	Raon Robotics, Hyulim Robot
Collaborative robots	Work beside humans	Growth	Doosan Robotics, Neuromeka, Robotis
Logistics robots	Warehouse and factory movement	Growth	Yujin Robot, T-Robotics
Service robots	Serving, guidance, store operation	Early growth	Clobot, Robotis, LG-related platforms
Medical robots	Surgery and rehabilitation	Product-specific	Curexo, P&S Mechanics
Humanoid robots	Human-like work in factories or homes	Very early	Rainbow Robotics, Boston Dynamics supply chain

The market is currently most focused on humanoids because the addressable market could be much larger if they work. They are also the hardest robots to commercialize.

Humanoids are difficult for four reasons. Balance is the first. Once a robot stands on two legs, real-time stability becomes a core engineering problem. Hands are the second. A humanoid needs to pick up fragile objects, rotate handles and hold tools with the same end effector. Power is the third. Motors, sensors and onboard computers all consume energy. Cost is the fourth. A factory robot only makes sense when it can compete with human labor economics.

That is why the gap between a demo and mass production is large. In 2026, the key signal is not another impressive video. It is whether robot parts begin to turn into orders, production lines and recognized revenue.

2. Physical AI Reopened the Robotics Cycle

Robotics is not new. Factory automation has existed for decades. What changed in 2025-2026 is the AI layer.

Older robots repeated pre-programmed motions. They stopped when the environment changed. New physical-AI robots are meant to see, interpret, plan and adapt. They combine cameras, sensors, foundation models, control software and mechanical bodies.

That matters most for humanoids. A humanoid is only useful if it can operate in messy environments such as factories, warehouses, stores and eventually homes. That requires both machine intelligence and reliable hardware.

AI does not remove the hardware bottleneck. If reducers are noisy, actuators are expensive, batteries last too briefly or cameras cannot support precise gripping, commercialization still slows. Physical AI should therefore be analyzed as a body-plus-brain value chain, not as software alone.

3. The Parts That Matter

A humanoid robot breaks into several core hardware layers.

3.1 Reducers: Precision Gears for Robot Joints

Motors spin fast. Robot joints need slower motion with high torque. A reducer converts motor speed into controlled force.

Reducers are costly and technically demanding because small errors create vibration and imprecision. Japan’s Harmonic Drive has long been the reference company in this layer. In Korea, SPG is the most visible listed reducer supplier. It appears repeatedly in Korean robotics value-chain maps across Boston Dynamics, Rainbow Robotics, Samsung and LG-related platforms.

3.2 Actuators: The Muscle Module

An actuator combines motor, reducer, driver and sensor into a joint module. A humanoid can use 20-40 or more actuators, so both price and reliability matter.

Korean names in this layer include Robotis, HL Mando, Hyundai Mobis and LG Electronics. HL Mando is trying to extend automotive steering and braking know-how into robot actuators. LG Electronics has emphasized actuator production because actuators represent a large share of robot cost.

3.3 Cameras and Sensors: The Eyes

Physical AI needs perception. Cameras, lidar, inertial sensors and force-torque sensors allow robots to locate objects, move around people and perform gripping.

Samsung Electro-Mechanics and LG Innotek are natural Korean names here because both already mass-produce camera modules for smartphones, vehicles and adjacent applications. Samsung Electro-Mechanics has discussed high-resolution recognition modules for humanoid customers and small, thin camera technology for precise gripping. Media reports have pointed to humanoid camera-module production in the second half of 2026.

3.4 Batteries and Power

Humanoids consume power through motors, compute and sensors. Battery density and safety therefore matter.

LG Energy Solution and Samsung SDI are the natural Korean power-layer names. Samsung SDI has discussed solid-state batteries for physical AI, while LG Energy Solution has deep cell and pack manufacturing capabilities that could extend into robotics over time.

3.5 Assembly and Manufacturing

A robot is also a precision-manufactured product. Frames, covers, wiring, heat management, quality testing and final assembly all matter. Intops is often mentioned as an OEM assembly candidate because of its electronics manufacturing background.

This is where Korea’s broader manufacturing base becomes relevant. Robotics sits between electronics and automotive components. Korea has both.

4. What Samsung, Hyundai and LG Are Actually Doing

4.1 Samsung: Physical AI Starts in the Factory

Samsung Electronics has described a path toward AI-driven manufacturing by 2030 and progressive deployment of humanoid and task-specialized robots across production lines. The practical starting point is manufacturing, not consumer homes.

Samsung Electro-Mechanics sits in the sensor layer. In 2026, the company discussed high-resolution recognition modules for global humanoid customers and small/thin camera technology for precise object gripping. Multiple media reports have also pointed to second-half humanoid camera-module production.

Samsung SDI is a longer-term power-layer story through solid-state batteries for physical AI. This is less about the first wave of humanoid revenue and more about solving later battery constraints.

4.2 Hyundai Motor Group: Boston Dynamics and Atlas

Hyundai has a different starting point because it owns Boston Dynamics. At CES 2026, Hyundai Motor Group outlined an AI robotics strategy and said Atlas would be introduced at HMGMA in Georgia from 2028, beginning with tasks such as parts sequencing before moving to more complex work.

Hyundai also describes RMAC, the Robot Metaplant Application Center, as a training and validation environment for AI robotics. This matters because robots improve through real-world data. RMAC is not just a showroom; it is a data and deployment infrastructure layer.

Hyundai Mobis and HL Mando are important parts names in this context. Mobis has the group-level electromechanical component base. HL Mando is building a long-term actuator strategy from automotive control and motion know-how.

4.3 LG: Actuators, Service Robots and Sensing

LG is building a broader physical-AI hardware ecosystem that includes service robots, home robots, actuators, sensing, batteries and assembly.

LG Electronics is especially focused on actuators. The logic is straightforward: actuators are a large share of robot cost, and LG has decades of motor design and production experience. LG Innotek adds camera and sensing modules; LG Energy Solution adds batteries; Intops is often discussed on the assembly side.

The near-term commercial opportunity is more likely to be industrial and component supply than a mass-market home robot. Homes are messy, price-sensitive and hard to standardize.

5. The 2026-2028 Verification Window

Robotics is a long-duration theme, but public markets need milestones.

Timing	Event to Track	Why It Matters
2Q26	Samsung Electro-Mechanics robotaxi camera module supply	Extends sensing modules into AI mobility
2H26	Samsung Electro-Mechanics humanoid camera module production	First visible humanoid parts revenue signal
2H26	LG Electronics actuator production preparation	Tests internalization of a core robot component
Around 3Q26	Hyundai RMAC operation	Training and validation infrastructure for Atlas
2027	Samsung SDI solid-state battery target	Long-term power constraint
2027-2028	HL Mando humanoid actuator line buildout	Auto-parts-to-robot-parts conversion
2028	Hyundai Atlas deployment at HMGMA	Factory deployment proof point
2028-2030	LG Innotek large-scale robot sensing opportunity	Sensor revenue scaling window

The timeline suggests a three-stage structure: first revenue signals in 2026, production validation in 2027-2028 and scale validation closer to 2030. Stocks that already price in the final stage can react sharply to delays in the first two.

6. Value-Chain Map: Separate Confirmed Orders From Estimates

Robotics supply chains move through research notes before official disclosures. The map below should be treated as an observation framework, not a confirmed order list.

Platform	Korean Names Often Discussed	Interpretation
Boston Dynamics Atlas	Hyundai Mobis, HL Mando, SPG, Robotis, Samsung Electro-Mechanics, LG Innotek, LG Energy Solution, Korea PIM	Hyundai group integration plus Korean component opportunity
Tesla Optimus	LG Innotek, LG Energy Solution, Samsung Electro-Mechanics, HL Mando, Robotis	Estimate-heavy; needs order confirmation before treating as revenue
Rainbow Robotics	SPG, KH Vatec	Samsung connection raises expectations, but revenue scaling is early
LG CLOi / Bear Robotics	Robotis, LG Innotek, LG Energy Solution, Intops	Service-robot and component-internalization path
Samsung manufacturing robots	Samsung Electronics, Samsung Electro-Mechanics, SPG, Intops	Factory deployment may arrive earlier than home robots

Repeated appearances are worth monitoring. SPG appears across reducers, Robotis across actuators and LG-related robots, Intops in assembly, LG Innotek in sensing and LG Energy Solution in batteries.

But repetition is not recognized revenue. There are three stages:

Stage	Meaning	Evidence
Supply-chain candidate	Development, samples, validation	Research notes, industry checks
Confirmed supplier	Contract or customer reference	Filing, call comment, customer disclosure
Recognized revenue	Revenue appears in accounts	Quarterly revenue, segment disclosure, order backlog

Much of the Korean robotics chain is still between stage one and stage two.

7. The Earnings Reality Across Listed Robotics Names

The main public-market issue is the gap between expectation and income statements. Based on 2025 financials and May 2026 market data, only a small number of robotics-related listed companies show meaningful operating profit. Many humanoid and collaborative-robot pure plays are still loss-making or very early in revenue scale.

7.1 Pure Robotics: The Highest Expectation, Earliest Earnings

Company	2025 Revenue	2025 Operating Profit	Market Interpretation
Rainbow Robotics	About KRW 34.1bn	Loss-making	Large Samsung-related expectation, high market-cap-to-sales burden
Doosan Robotics	About KRW 33.0bn	Loss-making	Global cobot brand, profitability still unproven
Robotis	About KRW 38.9bn	Turned profitable	Actuator plus service-robot optionality
Neuromeka	About KRW 19.0bn	Loss-making	Cobot growth story, small revenue base

For a name like Rainbow Robotics, the key issue is time. The market is not valuing only current robots. It is valuing a successful later-stage humanoid platform. That may be possible, but it requires revenue to rise by many multiples.

7.2 Parts Suppliers: Existing Business Plus Robotics Optionality

Company	Main Layer	2025 Revenue	2025 Operating Profit	What to Watch
SPG	Reducers	About KRW 341.7bn	Profitable	Repeated appearances across robotics platforms
Samhyun	Motors / actuator-adjacent parts	About KRW 95.0bn	Profitable	Auto-to-robot parts conversion
Haigen RNM	Actuators / motors	About KRW 70bn+	Loss-making or low margin	Speed of actuator commercialization
Korea PIM	Precision metal injection parts	About KRW 36.9bn	Profitable	Precision parts and internal robot materials

Parts suppliers have a different risk profile. If robotics is delayed, existing industrial, automotive or electronics businesses provide some cushion. If robotics accelerates, the new demand can be incremental.

7.3 Medical, Industrial and Logistics Robots

Company	Field	2025 Revenue	2025 Operating Profit	Interpretation
Raon Robotics	Industrial robots	About KRW 53.8bn	Profitable	Has earnings base, but humanoid exposure is indirect
Curexo	Medical robots	About KRW 74.5bn	Turned profitable	Export approvals and procedure volume matter
Hyulim Robot	Industrial / service robots	About KRW 168.3bn	Loss-making	Revenue exists, profitability needs verification
Yujin Robot	Logistics / service robots	About KRW 28.2bn	Loss-making	Needs autonomous logistics revenue traction

This group should not be forced into the humanoid basket. Medical robots, logistics robots and industrial robots have different customers, pricing and adoption cycles.

8. Why Korea Has a Robotics Parts Angle

Korea is not guaranteed to lead finished humanoid robots. The stronger starting point is parts and manufacturing.

Automotive manufacturing created motors, actuators, control systems and durability testing. Electronics created camera modules, batteries, miniaturization and production quality. Smartphone and appliance manufacturing created assembly and exterior-process know-how. Robotics can reuse all three.

Existing Industry	Robotics Capability	Korean Examples
Auto parts	Motors, actuators, controls, durability	Hyundai Mobis, HL Mando, Samhyun
Electronics components	Cameras, sensors, miniaturization, substrates	Samsung Electro-Mechanics, LG Innotek
Batteries	Energy density, safety, pack design	LG Energy Solution, Samsung SDI
Precision manufacturing	Reducers, gears, metal parts	SPG, Korea PIM
Assembly / casing	OEM production, quality control	Intops

The logic is similar to the Why Korea semiconductor substrate thesis. Korea is strong when a global theme becomes a manufacturing, parts and yield-learning problem. Humanoid robots could become that kind of problem if production moves beyond prototypes.

9. Three Public-Market Baskets

This is not a buy/sell framework. It is a way to avoid mixing different risk profiles.

9.1 Pure Robotics

Rainbow Robotics, Doosan Robotics and Neuromeka offer the cleanest thematic exposure, but current earnings are early. This basket is close to an option on faster-than-expected humanoid or collaborative-robot adoption.

Key checks: revenue growth, operating-loss reduction, strategic-customer orders and actual deployment.

9.2 Parts Suppliers

SPG, Robotis, Samhyun, Korea PIM and HL Mando sit closer to the first revenue layer. Reducers and actuators are high-content items in a humanoid. The advantage is existing business support; the disadvantage is lower pure-play torque.

SPG is the cleanest example of this logic because reducers appear across multiple robotics platforms. But sample supply, customer qualification, production order and recognized revenue are different events.

9.3 Large-Cap Robotics Options

Samsung Electro-Mechanics, LG Innotek, HL Mando, Hyundai Mobis, LG Energy Solution and Samsung SDI are not pure robotics stocks. Existing businesses dominate current earnings. Robotics is a long-term option attached to a larger manufacturing platform.

This lowers single-theme risk, but it also means robot revenue must become large before it moves the whole company.

10. What to Track Next

Variable	Why It Matters
2H26 production starts	Confirms whether Samsung Electro-Mechanics cameras, LG actuators and Hyundai RMAC move on schedule
Robotics revenue at parts suppliers	Shows whether SPG, Robotis, HL Mando and LG Innotek are converting exposure into sales
Pure-play robotics revenue scaling	Tests whether Rainbow Robotics and Doosan Robotics can grow into market expectations
Official orders vs. supply-chain estimates	Separates Atlas / Optimus development exposure from real contracts

The strongest signal is revenue in quarterly results. The second is a customer or company disclosure. The third is production-line investment. Headlines come last.

Final Note

Korea’s robotics industry is entering a more concrete phase. Samsung, Hyundai and LG are all talking about physical AI, and the 2026-2028 window includes real milestones for sensors, actuators, RMAC, Atlas deployment and component production. Korea also has the right industrial base: automotive parts, electronics components, batteries and precision manufacturing.

But public-market discipline matters. Many robotics listings remain early, and some pure plays already price in a successful 2028-2030 commercialization curve. The better question is not “should investors buy robot stocks?” It is “which layer shows revenue first?”

Reducers, actuators, cameras, batteries and assembly are the practical starting points. If humanoid robots move from demo to production, these parts must be bought before the full robot market becomes visible in revenue. The first test is 2H26 production signals, followed by 2027-2028 factory deployment and parts-supplier revenue recognition.

Sources Reviewed

Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.

AI PCB and Substrate Thesis: GPU, CPU, NIC and CCL Demand Are One System Bottleneck

Tue, 05 May 2026 23:15:00 +0900

Sector map: This is the upper-layer AI PCB and substrate thesis behind the Pamicell work. The company comparison continues in Korea AI PCB Ecosystem: 10 Companies, and the full reading path sits in the AI PCB hub. Read it together with Pamicell Part 1, Pamicell Part 2, and the Samsung Electro-Mechanics AI infrastructure re-rating note.

The company-level question is narrower: which Korean names have the best position in AI PCB, FC-BGA and low-loss material supply? This note moves one level up. It asks why capital should enter the whole ecosystem in the first place.

The answer is not “substrates are the next theme after GPUs.” That is too linear. AI infrastructure is no longer a GPU card. It is a rack-scale system. A modern AI rack has GPUs, CPUs, DPUs, NICs, switch ASICs, memory modules, power delivery, cooling control and high-speed boards. Each layer adds silicon. Each piece of silicon needs a package substrate, a module board, a motherboard, a switch board or a low-loss material stack.

That is the core point: the PCB and substrate layer is not the next stop in a simple rotation. It is the common denominator of the entire AI system bill of materials.

TL;DR

The market’s “GPU to memory to substrate” sequence is directionally useful, but incomplete. The better frame is simultaneous system expansion: GPU, HBM, CPU, DPU, NIC, switch ASIC and memory modules all rise together, and all require substrates or boards.
NVIDIA Vera Rubin NVL72 makes the point concrete. The platform combines 72 Rubin GPUs, 36 Vera CPUs, NVLink 6 switching, ConnectX-9 SuperNICs, BlueField-4 DPUs and Spectrum-X / Spectrum-6 Ethernet scaling. The CPU:GPU ratio is 0.5, or one CPU for every two GPUs. This is no longer a single-chip story.
Agentic AI raises the CPU intensity of inference. Tool orchestration, retrieval, code execution, database access, memory management and security isolation all lean on CPUs, DRAM, NICs and DPUs. If CPU content rises, server CPU FC-BGA, memory module boards, SoCAMM, motherboards and low-loss CCL all get pulled along.
Physical AI expands the thesis outside the data center. Autonomous vehicles, humanoids, industrial robots and space electronics all use more boards, more sensors, more edge AI modules and more reliability-sensitive PCB material. The time curve differs: data center first, autonomous driving second, humanoids later, space as a high-reliability margin premium.
For Korea, the investable map becomes layered. Samsung Electro-Mechanics is the high-end FC-BGA plus MLCC node. Daeduck Electronics is the FC-BGA / MLB / SoCAMM factor candidate. Doosan Electronic BG is the CCL anchor. Kolon Industries and Pamicell sit upstream in low-dielectric materials. Pamicell is not just a standalone stock idea; it is one compressed proxy inside a larger AI substrate system.

1. The Linear Frame: Useful, But Too Small

The market likes sequences because sequences are easy to trade.

First came GPUs: NVIDIA captured the budget because accelerators were scarce. Then came HBM: GPUs could not scale without memory bandwidth, so SK hynix, Samsung Electronics and Micron moved to the center of the discussion. The next step is usually described as substrates or advanced packaging: if GPUs and HBM scale, then FC-BGA, interposers, MLBs and CCL should follow.

That frame is not wrong. A larger GPU needs a larger and more complex package substrate. HBM expansion raises packaging intensity. Server boards become denser and faster. From that angle, substrates do come after GPUs and HBM in the recognition cycle.

But the frame misses the most important change in system architecture. AI infrastructure in 2026 is not a pile of GPUs. It is a rack-scale platform that co-designs compute, memory, networking, security and system control.

The NVIDIA Vera Rubin NVL72 configuration is the cleanest example. Public NVIDIA materials describe a system built around 72 Rubin GPUs and 36 Vera CPUs, with NVLink 6 switching, ConnectX-9 SuperNICs and BlueField-4 DPUs. NVIDIA’s own Rubin platform materials also frame the platform as a six-chip co-design: Vera CPU, Rubin GPU, NVLink switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch.

The arithmetic matters:

Vera CPUs = 36
Rubin GPUs = 72
CPU:GPU = 36 / 72 = 0.5

One CPU for every two GPUs is not a trivial host-processor footnote. It says the rack is a system, not a GPU shelf. Once that is true, the substrate thesis stops being a downstream echo of GPU demand. It becomes a multi-chip system thesis.

2. Why Every Chip Pulls A Board

The easiest way to see the substrate layer is to walk the system bill of materials.

System layer	Chip or module	Board / substrate demand
AI acceleration	GPU, custom ASIC, TPU	Large FC-BGA, advanced package substrate, high-density board
Host and orchestration	Server CPU, Vera CPU, x86 / Arm CPU	Large FC-BGA, CPU socket board, motherboard MLB
Memory bandwidth	HBM, DDR5, LPDDR-based server modules, SoCAMM	Interposer / substrate, memory module PCB, signal-integrity material
Networking	NIC, SuperNIC, Ethernet switch ASIC, InfiniBand switch	Switch board, optical module PCB, low-loss MLB
Data movement and security	DPU, SmartNIC	Package substrate, accelerator card PCB
Power and control	VRM, power modules, BMC and control boards	Power PCB, MLCC, high-reliability board

This is why “GPU demand” is too narrow. A hyperscaler does not buy a GPU in isolation. It buys a working rack. That rack contains compute chips, memory chips, networking chips, control chips and power electronics. The more the system expands, the more the board layer is asked to carry high-speed signals, heat, power and reliability.

The Korean equity translation is also clearer under this frame:

Korean layer	Companies to track	What they represent
High-end package substrate	Samsung Electro-Mechanics, Daeduck Electronics, Korea Circuit	FC-BGA and package substrate exposure
Multi-layer board and module PCB	Isu Petasys, Daeduck Electronics, TLB, Simmtech, Korea Circuit	Server motherboard, switch board, memory module and SoCAMM exposure
CCL anchor	Doosan Electronic BG	High-end copper-clad laminate for AI servers and networking
Low-loss materials	Kolon Industries, Pamicell	mPPO / low-dielectric resin and hardener inputs
Power stability	Samsung Electro-Mechanics and MLCC peers	MLCC content per AI server and high-voltage component mix

Pamicell belongs in this map as an upstream material proxy to Doosan Electronic BG. Samsung Electro-Mechanics belongs as the premium listed Korean node in FC-BGA and MLCC. Daeduck belongs as the broader FC-BGA / MLB / SoCAMM factor candidate. These are not separate stories. They are different points on the same system BOM.

3. Agentic AI Makes The CPU Layer Bigger

Traditional LLM inference looked simple from a hardware lens:

prompt
 -> GPU forward pass
 -> response

Agentic AI changes the workload. The model does not just answer. It plans, calls tools, searches, reads files, executes code, queries databases, manages memory, verifies output and may coordinate with other agents. The GPU is still central, but the non-GPU work becomes much heavier.

Agentic function	Main hardware pull
LLM forward pass	GPU + HBM
Tool orchestration	CPU
Retrieval and search	CPU + DRAM + storage
Code execution	CPU, sandbox, compiler / interpreter
Session memory and state	CPU + DRAM
Networked tool calls	NIC + switch ASIC + PCB
Security and isolation	CPU + DPU

TrendForce has been explicit about this direction. Its April 2026 agentic AI report and related public commentary describe a structural change in CPU:GPU ratios, tight server CPU supply and price increases by Intel and AMD. Tom’s Hardware reported the same direction from the industry side: AI server configurations that once used one CPU for four to eight GPUs can move toward much higher CPU intensity in agentic inference scenarios.

The exact ratio will vary by workload. A code-agent cluster is not the same as a video generation cluster. A retrieval-heavy enterprise agent is not the same as a pure batch inference system. But the direction is what matters for substrate investors: more CPU work means more CPU packages, more memory around the CPU, more networking, and more board-level signal integrity requirements.

The path from agentic AI to Korean substrates looks like this:

Agentic AI adoption
 -> CPU orchestration workload increases
 -> server CPU, DPU, NIC and switch ASIC content rises
 -> server CPU FC-BGA and high-layer MLB demand rises
 -> memory modules, SoCAMM and motherboard complexity rise
 -> low-loss CCL and low-dielectric materials become more important
 -> Korean substrate and material companies receive a share of the BOM expansion

That last line is important. Korea does not own the CPU market. Intel, AMD, Arm, NVIDIA and cloud-service-provider custom CPUs capture the chip value. Korean listed companies capture the board and material content around the chip. This is still meaningful because the substrate content grows with the system, not with a single product line.

4. Physical AI: Outside The Data Center

The data center is the first and largest near-term driver. Physical AI is the second expansion path. The timing is slower, but the direction is consistent: when intelligence moves into vehicles, robots, factories and satellites, more compute moves closer to the edge. More edge compute means more boards.

Autonomous driving

Autonomous driving is the most realistic second leg because cars already use large electronics stacks. A vehicle with advanced driver assistance or autonomous functions contains central compute, sensor fusion, camera modules, radar, lidar, vehicle Ethernet and redundant safety controllers.

Vehicle system	PCB and material pull
Central compute	High-density board, processor package substrate
Sensor fusion ECU	Multi-layer PCB, high-speed signal board
Camera / lidar / radar	Rigid-flex, RF board, module PCB
Vehicle Ethernet	Low-loss CCL and high-speed communication PCB
Safety redundancy	More ECUs and more board area

This does not hit earnings as quickly as AI data centers. Vehicle programs take time, qualification is slow and the revenue curve is model-cycle dependent. But the direction is not ambiguous: a smarter vehicle carries more board content than a traditional vehicle.

Humanoid and industrial robotics

NVIDIA Jetson Thor gives the physical AI argument a concrete hardware reference point. NVIDIA positions Jetson Thor for physical AI and robotics, with up to 2,070 FP4 TFLOPS, 128GB of memory and a 40W to 130W configurable power range. That kind of edge AI module needs high-density boards, power boards, sensor interconnect and flexible PCBs.

Humanoids will not drive Korean substrate earnings tomorrow. They are not yet a mass-volume market. But they extend the option value of the thesis. If edge AI modules become standardized across robots, factories and industrial machines, board content shifts from a data-center-only story to a distributed compute story.

Space and defense electronics

Space is different. It is not a volume story. It is a reliability and margin story. NASA and IPC-related materials for mission hardware emphasize high-reliability PCB requirements, supplier qualification and Class 3 / space-addendum type standards. For listed Korean PCB names, the relevance is not “space will absorb massive capacity.” It is that harsh-environment electronics can justify higher reliability standards and potentially better margins.

The time curve looks like this:

End market	Earnings timing	PCB intensity	Practical confidence
AI data center	Fast	Very high	High
Autonomous driving	Medium	High	Medium to high
Humanoids / robotics	Slow	Medium to high	Medium
Space / defense electronics	Slow	High specification, low volume	Medium

This ordering matters. The near-term model should still be data-center led. Physical AI is not a reason to stretch every valuation today. It is a reason the terminal market may be broader than the current AI server cycle implies.

5. What This Adds To The Pamicell Thesis

The Pamicell work started with a company-specific recognition gap. The market remembered a stem-cell company. The income statement increasingly looked like an AI CCL material supplier. The Doosan Electronic BG link, repeated supply-contract evidence, high-margin biochemicals and KRX industry reclassification all pushed in the same direction.

This sector thesis changes the question from “why Pamicell?” to “why does the upstream CCL material layer matter at all?”

The answer is that CCL and low-loss materials are not tied to one GPU generation. They sit under the system layer:

GPU / CPU / DPU / NIC / switch ASIC expansion
 -> high-speed signal and thermal constraints rise
 -> low-loss CCL demand rises
 -> CCL producers need low-dielectric materials
 -> upstream suppliers such as Pamicell become compressed proxies

Pamicell is still not the same as Doosan Electronic BG, and it is not a PCB manufacturer. It is further upstream. That means customer concentration and qualification risk are real. But it also means the company’s small size can make the same system-level demand look more powerful in percentage terms if orders keep compounding.

Put differently: Pamicell’s thesis becomes more durable if the AI board cycle is not just a GPU substrate cycle, but a multi-chip system substrate cycle.

6. What This Adds To The Samsung Electro-Mechanics Thesis

Samsung Electro-Mechanics was already re-rated around two ideas: high-end FC-BGA and AI-server MLCC. That older framing still works. The system-BOM thesis makes it cleaner.

If the only driver were GPUs, Samsung Electro-Mechanics would be a high-end package-substrate story with MLCC attached. If the driver is rack-scale system expansion, the company sits in more than one lane:

Lane	Why it matters
FC-BGA	Larger and more complex CPUs, GPUs, ASICs and networking chips need high-end package substrates
MLCC	AI servers, networking trays and power delivery all increase component density
Glass substrate option	Future large-package architecture may require new substrate materials and processes
Automotive and robotics electronics	Physical AI increases high-reliability component and board demand over time

This does not remove valuation discipline. Samsung Electro-Mechanics has already been recognized by the market more than Pamicell or some smaller PCB names. The factor question is not “is this a good company?” It is “how much of the system-level substrate expansion is already in the price, and how much must be confirmed through orders, margins and guidance?”

That is why this note treats Samsung Electro-Mechanics as the premium anchor rather than the highest-beta idea. It is the cleanest Korean large-cap expression of FC-BGA plus MLCC, but its future return depends on continued estimate revisions rather than simply discovering the theme.

7. Portfolio Frame: Core, Barbell, Options

The company ranking can change with price, but the factor map is stable.

Role	Candidate exposure	Reason
Premium anchor	Samsung Electro-Mechanics	FC-BGA + MLCC + customer qualification + scale
Core PCB factor	Daeduck Electronics	FC-BGA, MLB and potential SoCAMM / module board sensitivity
CCL anchor	Doosan Electronic BG inside Doosan	High-end CCL body of the domestic chain
Upstream material barbell	Kolon Industries and Pamicell	Low-dielectric resin / material exposure, smaller base and higher operating leverage
Optionality	Korea Circuit, TLB, Simmtech, Isu Petasys	Memory module, MLB, networking and broader AI PCB beta

The point is not to force one answer. The point is to avoid treating all “AI PCB” names as the same asset. Package substrates, multi-layer boards, CCL and low-dielectric chemistry have different margin structures, qualification periods and customer risks.

A useful portfolio view is:

Premium anchor:
 Samsung Electro-Mechanics

Core substrate / board factor:
 Daeduck Electronics, selected MLB names

Upstream material barbell:
 Kolon Industries + Pamicell

Higher-beta options:
 Korea Circuit, TLB, Simmtech, Isu Petasys

When the market says “the PCB cycle is over,” the system-BOM thesis helps test whether that is true. If GPU shipments slow but CPU content, networking ASICs, DPUs and memory module complexity keep rising, the board cycle may cool in one lane while staying tight in another.

8. What Could Break The Thesis

The common-denominator frame is not a claim that the cycle lasts forever. Four risks matter.

First, hyperscaler AI capex can slow. If AWS, Microsoft, Google or Meta guide down for more than one quarter, the entire hardware supply chain will feel it.

Second, substrate technology can change. Glass substrate or other new architectures could alter the FC-BGA cycle faster than expected. This does not remove board demand, but it can shift the winners.

Third, capacity can arrive. If high-end CCL, glass fiber or low-loss material capacity enters faster than expected, pricing power can normalize before volumes fully mature.

Fourth, physical AI can take longer than the market wants. Autonomous driving, humanoids and space electronics all have long qualification and adoption cycles. They are not substitutes for near-term data-center revenue.

These risks do not kill the thesis. They define the checklist.

9. Verification Checklist

The thesis should be tracked at the system level, not just through one company’s quarterly number.

Signal	Why it matters
NVIDIA rack-scale roadmap	More chip types and higher rack density extend the common-denominator substrate cycle
CPU:GPU ratio commentary	A higher CPU ratio strengthens the CPU FC-BGA and motherboard MLB leg
Hyperscaler capex guidance	The first-order demand source for AI data-center boards
CCL and glass-fiber lead times	Confirms whether material tightness is real or easing
Samsung Electro-Mechanics package and component margins	Tests whether premium substrate and MLCC pricing still hold
Daeduck / MLB order commentary	Tests whether broader PCB beta is converting into revenue
Pamicell and Doosan Electronic BG contract cadence	Tests whether upstream CCL material demand is still compounding
Automotive / robotics PCB comments in Korean company IR	Early sign that physical AI is moving from optionality to revenue

If these signals stay aligned, the substrate cycle is not simply a 2025-2027 theme. It becomes a multi-year system architecture shift.

FAQ

What is the AI PCB thesis?

The AI PCB thesis argues that AI infrastructure demand is no longer limited to GPUs and HBM. Rack-scale systems need GPUs, CPUs, NICs, DPUs, switch ASICs, memory modules and power boards. Each layer requires package substrates, multi-layer boards or low-loss materials.

Why does agentic AI increase CPU demand?

Agentic AI uses tools, retrieval, code execution, memory management and orchestration. Those tasks add CPU, DRAM, networking and DPU workload around the GPU. Higher CPU content can increase demand for server CPU FC-BGA, motherboards, memory module boards and low-loss CCL.

Why are Samsung Electro-Mechanics and Pamicell in the same sector map?

They sit at different points in the same AI substrate chain. Samsung Electro-Mechanics is a premium FC-BGA and MLCC node. Pamicell is an upstream low-dielectric material supplier linked to Doosan Electronic BG’s CCL cycle. The same system-level AI board demand can affect both, but with different risk and valuation profiles.

Is Pamicell a PCB company?

No. Pamicell is not a PCB manufacturer. The relevant thesis is upstream material exposure: low-dielectric / low-loss inputs used by CCL producers such as Doosan Electronic BG.

Is this investment advice?

No. This is a sector research framework. The right conclusion depends on valuation, order cadence, margin confirmation, customer concentration and each investor’s risk tolerance.

Selected Public Sources

NVIDIA Vera Rubin NVL72 product page: https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/
NVIDIA Technical Blog, Vera Rubin platform overview: https://developer.nvidia.com/blog/inside-the-nvidia-rubin-platform-six-new-chips-one-ai-supercomputer/
TrendForce, agentic AI and CPU:GPU ratio commentary: https://insights.trendforce.com/p/agentic-ai-cpu-gpu
TrendForce report page, 2026 Agentic AI Wave: https://www.trendforce.com/research/download/RP260408AD
Tom’s Hardware, agentic AI and CPU demand: https://www.tomshardware.com/pc-components/cpus/shifting-need-for-cpus-in-ai-workloads-drives-intensifying-shortages-price-hikes
NVIDIA Jetson Thor product page: https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
NASA quality guidance for high-reliability PCB standards: https://sma.nasa.gov/sma-disciplines/quality
NASA GSFC-STD-8001 high-reliability PCB standard: https://standards.nasa.gov/sites/default/files/standards/GSFC/Baseline/0/gsfc-std-8001.pdf

Final Note

The cleanest version of the thesis is simple:

AI does not buy GPUs alone. It buys systems.

Systems add chips. Chips need substrates, boards and low-loss materials.

That is why the substrate layer should be read as a common bottleneck rather than a late-cycle afterthought. The implication is not that every Korean PCB or material stock deserves the same multiple. It is that the ecosystem should be evaluated as a system-level supply chain: Samsung Electro-Mechanics at the premium FC-BGA / MLCC node, Daeduck and MLB names at the board layer, Doosan Electronic BG at the CCL body, and Kolon Industries and Pamicell upstream in low-dielectric materials.

The work from here is not to repeat the theme. It is to track whether the system BOM keeps getting thicker, whether CPU and networking content continue to rise, and whether Korean companies convert that complexity into orders and margins.