Physical AI and the Humanoid Supply Chain: Korea's Bid to Become the World's Component Factory

By Leo Jaeyoul Lim | CEO, SEJONG DnC
Market observer since the 1990s · September 1, 2026 · Robotics & Industrial Technology · ~14 min read

Robotics has had false starts before. Specialized machines have been promoted as transformative for decades, and the actual economic footprint has usually disappointed relative to the promise. Having watched a number of automation cycles arrive with fanfare and then plateau since the 1990s, I think the more useful question about the current humanoid-robot buildout isn't whether the technology is impressive -- it clearly is -- but whether the underlying unit economics and supply-chain structure are different enough this time to justify the capital now flowing into it. Two threads converge on an answer. First, the vision-language-action models steering these machines have moved robotics out of the single-purpose category that limited prior generations. Second, and less discussed outside Korea, a sixfold upward revision in Goldman Sachs' own shipment forecast between 2024 and 2026 has been accompanied by a specific claim: that Korea's automotive-parts supply chain, not Silicon Valley software, may end up supplying roughly a third of the world's humanoid hardware by 2035.

6.48MGoldman Sachs' revised 2035 global humanoid shipment forecast, up from a prior estimate of 1.38 million units (Source: Goldman Sachs Research, cited by Investing.com, Aug 2026)

30%Share of 2035 global humanoid production Goldman Sachs estimates Korea's supply chain will directly or indirectly support (Source: Goldman Sachs, "Robotics: A Time of Transformation," Jun 25, 2026)

~1.9yrEstimated 2027 factory payback period for a humanoid deployment, measured in equivalent years of human labor cost -- down from roughly 2.8 years in 2026 (industry cost estimate; single-source basis)

All figures cross-verified against at least two independent sources where possible; single-source figures are flagged inline.

1. The Forecast That Keeps Getting Revised Upward

Start with the headline number, because its trajectory is itself informative. Goldman Sachs first sized the 2035 humanoid market at 1.38 million units in an earlier estimate; by mid-2026 that had been revised to 6.48 million units and roughly $138 billion in market value -- a sixfold increase in under two years, driven by faster-than-expected progress in end-to-end AI training and steeper cost declines across the hardware stack. A forecast that moves this much this quickly deserves to be read as a wide-error-bar scenario rather than a precise target, a caveat worth holding through everything that follows.

Bar chart showing humanoid robot shipment forecast growing from 0.016 million units in 2025 to 0.075 million in 2026, 0.89 million in 2030, and 6.48 million units by 2035 — Figure 1. Global humanoid robot shipment forecast, 2025-2035 (Goldman Sachs, 2026 revision).

2. From Digital to Physical: Why the Control Stack Is Splitting in Two

The center of gravity in AI research has been shifting from purely digital systems -- large language models generating text and images -- toward what practitioners now call Physical AI or Embodied AI: systems that perceive a physical environment through vision and touch and act on it directly, under the constraints of Newtonian mechanics rather than a probability distribution over tokens. That shift matters commercially because the cost of error is asymmetric in a way pure software never had to contend with. A hallucinating chatbot produces a wrong answer; a robot arm executing a hallucinated action can damage equipment, halt a production line, or injure a person nearby.

The industry's response has been to split the control problem into two loops running at very different speeds. A high-level cognitive layer -- built on vision-language-action (VLA) models such as Helix, Pi-Zero, and OpenVLA, often paired with large language model reasoning -- interprets natural-language instructions and plans multi-step tasks at roughly 7 to 20 Hz. Beneath it, a low-level dynamics layer running at 200 to 1,000 Hz handles real-time balance, joint torque, and obstacle avoidance using reinforcement learning and model-predictive control, with deterministic safety interlocks that override the upper layer if a proposed action would be physically unsafe. The practical effect is that the probabilistic, occasionally-wrong reasoning layer never gets direct, unfiltered authority over the robot's limbs.

Diagram showing Physical AI's dual-loop control architecture: a System 2 cognition and planning layer running at 7 to 20 Hz using vision-language-action models, connected via a confirmed safety interlock to a System 1 real-time dynamics control layer running at 200 to 1000 Hz using reinforcement learning and model-predictive control — Figure 2. The two-loop control architecture that lets humanoid robots combine language-level reasoning with real-time physical safety.

3. The 24/7 Economics That Specialized Robots Never Had

Purpose-built robots -- automated guided vehicles, collaborative arms, delivery bots -- have circulated through industrial facilities for years without producing the kind of economic transformation now being forecast for humanoids. The reason is structural: a single-purpose robot idles the moment its designated task is complete, which stretches out the payback period on the capital required to buy it. A general-purpose humanoid, built to roughly human proportions, can instead use existing tools, walk existing factory floors, and move through doorways designed for people -- meaning it requires no infrastructure retrofit and can, in principle, run continuously across a rotating set of tasks.

That utilization advantage is only becoming investable now because of a parallel cost curve. Actuators, sensors, and AI inference chips are following a "technology deflation" trajectory familiar from other hardware categories, pushing bill-of-materials costs down each year even as average selling prices fall alongside them. The gap between the two curves is what eventually turns a humanoid from a research showpiece into a line item a factory manager can justify against a headcount budget.

Line chart showing humanoid robot average selling price declining from about 42000 dollars in 2025 to 21300 dollars in 2035, and bill of materials cost declining from about 27700 dollars to 13500 dollars over the same period — Figure 3. Humanoid unit economics: average selling price and underlying manufacturing cost, 2025-2035 estimates.

Table 1. Factory Payback Period, in Equivalent Years of Human Labor Cost
Year	Initial deployment cost (est.)	Payback period
2026	~$45,000 (incl. systems integration)	~2.8 years of labor cost
2027	Falling with BOM cost curve	~1.9 years of labor cost
2030	Falling with BOM cost curve	~1.2 years of labor cost
2035	Falling with BOM cost curve	Under 1 year of labor cost

Source: Industry cost-curve estimates compiled from company disclosures and equity research, 2026. Treat as directional; single-source basis.

4. Three Approaches, One Race: Tesla, Figure AI, and Boston Dynamics

The competitive field has settled into three distinct technical philosophies rather than converging on a single architecture. Tesla's Optimus program leans on a vision-only sensor stack -- deliberately excluding LiDAR -- and repurposes the neural-network infrastructure built for its Full Self-Driving effort, aiming for vertically integrated mass production at a sub-$20,000 target price. Figure AI has taken the opposite path on partnerships, pairing a multimodal vision system with OpenAI's language models under a strategic alliance, betting that System-2 reasoning capability differentiates its robots in complex, less-structured environments such as BMW's Spartanburg assembly line. Boston Dynamics, now under Hyundai Motor Group, has stayed closest to its research heritage: LiDAR and RGB-D sensor fusion, a hybrid of reinforcement learning and model-predictive control, and a focus on dynamic locomotion and industrial-grade durability, with deployment inside Hyundai's own HMGMA logistics operations and a nasdaq listing reportedly under consideration.

Three-card comparison of Tesla Optimus, Figure AI Figure 02/03, and Boston Dynamics Atlas showing their sensor approach, AI partnership strategy, and production targets — Figure 4. Three technical philosophies competing for the humanoid market.

5. Korea's Opening: Tacit Manufacturing Knowledge as the Scarce Input

The most consequential recent development for institutional readers outside Korea may not be a robotics company announcement at all -- it's a June 25 Goldman Sachs report, "Robotics: A Time of Transformation," that assigns Korea a specific, quantified role in the global humanoid supply chain. The bank's reasoning centers on an asset that doesn't show up cleanly in a semiconductor or software forecast: Korea's existing automotive-parts ecosystem, robot density (the country has among the highest ratios of industrial robots per manufacturing worker in the world), and precision-actuator manufacturing capability, particularly in the reduction gears -- the mechanical linkage between motor and joint -- that Goldman analysts argue Korea's supply chain now leads even Japan and Germany in producing for humanoid applications.

Goldman's estimate is specific: Korea's supply chain could directly and indirectly support roughly 74,000 humanoid units in 2030 and 412,000 units by 2035, translating into a 30% share of global production outside China. The bank frames US robotics firms as already courting Korean component suppliers for a structural reason -- American software leadership in the VLA and foundation-model layer has no equivalent domestic hardware-manufacturing partner at the scale required, whereas Korea's parts ecosystem, built over decades of automotive supply, survived in a way the equivalent US ecosystem largely did not. Our recent coverage of Korea's physics-informed industrial AI push traced a related structural argument -- that Korea's edge increasingly lies in converting decades of plant-floor domain knowledge into a defensible data and hardware asset, rather than in racing to build its own foundation models from scratch.

Donut chart showing Korea's projected 30 percent share of global humanoid robot production by 2035, with the remaining 70 percent from the rest of the world excluding China — Figure 5. Korea's estimated share of global humanoid production outside China, 2035.

Table 2. Three Humanoid Strategies Compared
Company	Core Sensor/AI Approach	Primary Deployment Site	Stated Target
Tesla (Optimus)	Vision-only, no LiDAR; FSD-derived stack	Tesla gigafactories	Mass production, sub-$20K
Figure AI (Figure 02/03)	Multimodal vision + OpenAI VLA alliance	BMW Spartanburg plant	B2B commercial supply
Boston Dynamics (Atlas)	LiDAR + RGB-D fusion, RL + MPC	Hyundai HMGMA	Industrial-grade, IPO-linked

6. The Demographic Clock Korea Cannot Wait Out

What separates Korea's position from a purely opportunistic industrial-policy bet is a demographic constraint that makes robotic labor substitution closer to a structural necessity than a discretionary upgrade. Korea's annual births fell from roughly one million in 1970 to 238,300 in 2024, according to national statistics -- a decline of close to 76% that leaves the workforce entering the labor market twenty years from now at a fraction of its historical size. For most developed economies, automation displacing routine labor raises a politically fraught unemployment question. In Korea's case, the more binding constraint over the coming two decades is not too few jobs but too few working-age people to fill the jobs that already exist, which reframes robotic labor substitution as a condition for maintaining existing industrial output rather than a threat to existing employment.

7. What the Task-Based Labor Model Says About Who Wins and Who Doesn't

Economist Daron Acemoglu's task-based model of labor markets offers a more precise framework than a blanket "robots take jobs" narrative. Automating technology simultaneously produces a displacement effect, removing specific tasks from human workers, and a reinstatement effect, creating new tasks -- robot workflow design, field data curation, teleoperation, predictive maintenance -- that did not previously exist. Applied to physical AI, the tasks most exposed to near-term displacement are standardized, repetitive, and high-risk: routine material handling, precision assembly on fixed cycle times, and the "3D" (dirty, dangerous, demanding) roles that have historically been hardest to staff in aging industrial economies. The tasks most likely to expand are supervisory and design-oriented -- overseeing fleets of robots, curating the sensor data that trains next-generation models, and managing the human-robot interaction layer -- alongside a category of high-touch human work (complex problem-solving, in-person care, emotionally sensitive service roles) whose relative economic value may rise precisely because it remains difficult to automate.

For Korea specifically, this reframes the education-and-training question raised in the original policy discussion around physical AI: engineering curricula shift from individual machine control toward robot-foundation-model fine-tuning and physical-data digitization, while vocational training shifts from manual tool operation toward teleoperation and fault diagnosis -- a transition the government's Manufacturing AI 2030 strategy, discussed in our earlier analysis of industrial-software platforms as the defensible layer in advanced manufacturing, is attempting to fund directly.

8. Market Outlook: Two Scenarios for the Next Decade

What follows is a probability-weighted reading of where the currently available data points, not a directional prediction.

Base Case (more probable given current evidence)

Cost curves for actuators, sensors, and inference chips continue their multi-year decline in line with the pattern already visible in the 2025-2035 BOM estimates, converting humanoid deployment from a research expense into a factory-floor line item with sub-two-year payback by the late 2020s. Korea's component ecosystem -- concentrated in precision actuators and reduction gears -- converts Goldman's 30% supply-chain-share estimate into real, growing order volume from Tesla, Figure AI, and Boston Dynamics alike, reinforced by government co-investment through Korea's national manufacturing alliance. Labor-market adjustment in Korea proceeds through the reinstatement channel Acemoglu's model predicts, with engineering and robot-supervision roles absorbing displaced routine-task workers faster than in economies without Korea's acute demographic pressure forcing the transition.

Risk Case Worth Taking Seriously

Humanoid shipment forecasts have already been revised sharply upward multiple times in two years -- Goldman's own 2035 estimate moved from 1.38 million to 6.48 million units -- which is itself a signal that current projections carry wide error bars in either direction. If VLA model reliability plateaus short of what's needed for genuinely unstructured environments, industrial adoption could concentrate in narrow, pre-mapped use cases rather than the broad 24/7 general-purpose deployment the economic case depends on, stretching out payback periods materially. For Korea specifically, the 30% supply-chain-share estimate assumes sustained capital investment and successful technology transfer into actuator and reduction-gear manufacturing at a scale the domestic supply base has not yet demonstrated at volume; a slower-than-expected ramp would leave Korean suppliers exposed to Chinese cost competition in exactly the mid-tier component categories where China's manufacturing base is already strongest.

9. Basic Investment Principles

This section is deliberately general and not a recommendation to buy, sell, or time any specific position, company, or sector.

Diversify across the humanoid value chain rather than a single company or country. Foundation-model developers, component manufacturers, and system integrators respond to different adoption cycles even within the same broad theme, and Korea's opportunity is concentrated in the hardware-component layer specifically, not the AI-software layer.
Weight forecast figures by their revision history. A market where the leading forecaster has revised its own headline number sixfold in roughly two years is one where any single point estimate should be treated as a scenario input, not a target.
Distinguish demonstrated deployment from roadmap ambition. Hyundai's HMGMA logistics use and BMW's Spartanburg pilot are running deployments; multi-million-unit 2035 shipment figures remain forecasts built on cost-curve assumptions that have not yet been tested at scale.
Maintain a long time horizon and avoid overweighting any single company announcement, cost-curve data point, or government policy statement in isolation from the broader multi-year capital and demographic trajectory this piece describes.

This article is also published, with a different angle for Korean retail-investor readers, on Leo Jaeyoul Lim's Korean-language sister blog on Naver (blog.naver.com/woduf00). Both blogs are written and operated by the same author -- readers interested in the Korea-specific retail angle on this same story are welcome to check it out.

Search This Blog

Senior Wisdom: Navigating Capital and Power