The Goliath Era: Physical AI, Humanoid Robots, and Korea's Last Shot at Industrial Leadership
The Goliath Era: Physical AI, Humanoid Robots, and Korea's Last Shot at Industrial Leadership
For three years, the most consequential AI story was a software story: chatbots, coding assistants, and the trillion-dollar buildout of data centers to train and run them. That story has not ended, but a second one has started underneath it, and it is a hardware story. In January 2026, Boston Dynamics stood a fully electric, production-ready version of its Atlas humanoid on stage at CES and said the robot's entire 2026 output was already sold — committed to Hyundai's own factories and to Google DeepMind, with the first commercial deployment landing on a Georgia assembly line in 2028. A few months later, Korea's Finance Ministry confirmed it would put KRW 5.1 trillion of public money behind humanoid robots and "physical AI" pilots through 2030. Neither of these is a demo. Both are capital allocation decisions by people who do not make them lightly.
Having watched several waves of "this changes everything" technology arrive, peak, and get bid back down to reality since the 1990s, I want to be precise about what is and is not established here. What is established: humanoid robots are moving off the trade-show stage and into named factories on named timelines, backed by named capital. What is not established: which country, which company, and which business model captures the value created when that happens. This piece is about Korea's position in that second, unresolved question — where its manufacturing base gives it a real structural advantage, where its demographic and regulatory realities threaten to erase that advantage, and what the evidence suggests companies and policymakers should actually do about it.
All figures cross-verified against at least two independent sources where possible; single-source figures are flagged inline.
1. From the Training Era to the Goliath Era
The cleanest way to understand what changed in 2026 is to separate two phases of the same AI buildout. The first phase, roughly 2022 to 2024, was about training: acquiring enough GPU compute and high-bandwidth memory (HBM) to build ever-larger language models, a race in which Korea's Samsung Electronics and SK hynix captured extraordinary margins as the HBM supply bottleneck's primary beneficiaries. The second phase, underway now, is about inference and application — putting those trained models to work inside real enterprise systems and, increasingly, inside physical machines that act on the world rather than merely answering questions about it. That shift changes what "winning the AI race" even means, and it changes who is positioned to win it.
The scale of the infrastructure behind even the software side of this shift is worth grounding in a concrete comparison. A single hyperscale AI data-center campus now under construction in Utah is targeting roughly 9 gigawatts of capacity — compared with Seoul's entire municipal power consumption of roughly 10 gigawatts, and Utah's own statewide grid capacity of about 4 gigawatts. A facility built to do nothing but generate machine intelligence is being sized to consume as much electricity as one of the world's largest cities. That is the infrastructure base the "playground era" of chatbots quietly built. Physical AI is what gets bolted onto that base once the compute exists to run it in real time, inside a robot, on a factory floor, with no tolerance for the multi-second delay a cloud round-trip would introduce.
1.1 Why memory's HBM windfall was always going to normalize
Korea's chip windfall over the past two years rests on a specific, temporary condition: demand for HBM has outrun supply because building new fabrication capacity takes four to five years, not months. That mismatch is real and it has been extraordinarily profitable for Samsung and SK hynix, but it is a supply-timing story, not a permanent moat. Once the capacity investments already announced come online — Samsung's Pyeongtaek P5 cleanrooms phasing in from 2027, SK hynix's Cheongju M17 fab targeting its first cleanroom for December 2028 — the pricing premium built into current earnings should compress. Samsung's own public guidance points to a supply-demand imbalance persisting through 2028; SK hynix has guided to a fuller normalization by 2030. Either way, the memory windfall has a visible expiration window, which is precisely why the more durable question is what comes after it: not who supplies the parts for other people's AI factories, but who owns and operates the physical layer where AI actually does the work.
2. Why "Physical AI" Is a Fundamentally Harder Problem Than Chatbots
It is tempting to assume that a company good at building large language models will be roughly as good at building a robot that uses one, since both are "AI." The evidence from 2026 suggests this assumption is close to backwards. A chatbot that hallucinates 10-30% of the time is an annoyance a user can catch and route around; that error rate is compatible with mass commercial deployment, and it is roughly what today's frontier consumer LLMs run at. A robot with a comparable failure rate on a factory floor is not annoying — it is a machine that drops a fifty-kilogram part, crushes a co-worker's hand, or halts an entire production line until a technician can diagnose what went wrong. Physical systems operating under gravity, friction, and inertia need reliability closer to 99.999% before a plant manager will accept the liability of putting them near people and expensive equipment.
2.1 Boston Dynamics' Atlas versus the choreography problem
This gap explains why Boston Dynamics' publicly demonstrated Atlas — a fully electric, 6-foot-3-inch, roughly 200-pound humanoid unveiled at CES 2026 — was framed by the company explicitly around industrial capability rather than performance stunts. Atlas is built with a front-back symmetric structure that lets its joints rotate past human anatomical limits, and it has demonstrated lifting loads over 50 kilograms — a small refrigerator, in one widely circulated demonstration — while keeping the object level. Boston Dynamics says the robot can be trained on a new task in under a day, a claim made possible by an AI partnership with Google DeepMind and Nvidia infrastructure for simulation-based training, announced alongside the CES unveiling. None of that is a dance routine. It is, deliberately, the opposite: sustained, precise, heavy-load handling under real physical constraints, which the company and Hyundai believe is the actual bar for factory deployment.
The commercial commitment behind Atlas is unusually concrete for a robotics announcement. Hyundai Motor Group, which owns roughly 80% of Boston Dynamics, said the entire 2026 production run is already allocated between its own Robotics Metaplant Application Center and Google DeepMind, with additional customers to follow in 2027. Deployment at Hyundai's Metaplant America near Savannah, Georgia begins in 2028, starting with parts sequencing before expanding to component assembly by 2030. Hyundai has stated an ambition to manufacture 30,000 Atlas units annually by 2028. Multiple independent outlets — Axios, Forbes, TechRadar, and Boston Dynamics' own release — corroborate the production and deployment timeline, though the 30,000-unit annual manufacturing target and the precise per-unit pricing (reported by Korean trade outlet KED Global at roughly $320,000) should be treated as company guidance and single-sourced estimates respectively rather than confirmed facts, pending actual production data.
2.2 China's humanoid fleet: fast legs, no hands, no brain
China's robotics sector has produced the most visually dramatic humanoid demonstrations of the past two years — robots running competitive sprints, executing martial arts routines, dancing in synchronized troupes. Unitree's platforms in particular have set records for bipedal running speed and shipped in volume: ABI Research's 2026 humanoid forecast specifically cites Chinese state funding and permissive regulation as a structural growth advantage, and Chinese firm Agibot has publicly stated it has shipped roughly 5,000 units. But industry reporting from CES and China's own AW 2026 robotics expo has repeatedly noted a specific pattern behind the most impressive routines: many of the more complex demonstrated behaviors have relied on human operators controlling the robots by joystick from offstage rather than autonomous execution, and a large share of the showcased hardware uses simple gripper-style end effectors rather than five-fingered hands capable of the fine manipulation industrial assembly requires.
That gap matters more than it might appear. Locomotion — walking, running, maintaining balance — is a comparatively well-solved control problem that mostly requires not falling over. Dexterous manipulation — recognizing an object's texture and center of mass, adjusting grip force in real time, handling irregular or unfamiliar shapes — requires an entirely different order of perceptual and decision-making intelligence, closer to what robotics researchers consider a meaningful step toward general-purpose autonomy. A robot that can sprint but cannot reliably pick up a coffee cup by its handle is, for factory-floor purposes, a much less useful machine than the marketing reels suggest.
3. Hand Intelligence: The Real Bottleneck, and Where Korea Has Quietly Built an Edge
Robotics researchers increasingly describe dexterous manipulation — what gets called "hand intelligence" — as the actual gating technology for physical AI, more so than bipedal locomotion or even large-model reasoning. An estimated 70-80% of real manufacturing tasks cannot be handled by simple gripper automation: picking a smooth, cylindrically shaped tool off a table at an arbitrary angle, supporting a box's weight with one hand while opening its flap with a finger of the other, adjusting grip pressure the instant a material's texture changes. These are tasks human factory workers perform without conscious thought, and they are exactly the tasks that have resisted a decade of industrial-robotics automation because they require perceiving and adapting rather than executing a pre-programmed motion.
This is the specific niche where Korean firms have built a credible, internationally recognized position — not in headline-grabbing humanoid demonstrations, but in the manipulation software underneath them. RLWRLD, a Korean robotics startup, showcased its RLDX-1 robotics foundation model at Computex 2026 in Taipei, presenting technology reported to be hardware-agnostic across different robotic hands and grippers, aimed squarely at the complex, variable-object manipulation problem — sorting irregular parcels, opening containers, handling items whose shape or position changes between attempts. Samsung-affiliated Rainbow Robotics has moved its mobile dual-arm RB-Y1 platform beyond the lab entirely: multiple Korean media outlets report it is now undergoing field trials inside Coupang logistics centers, a genuine industrial deployment rather than a controlled demo. Nvidia's own courtship of Korean robotics makes the same point from the customer side — CEO Jensen Huang has visited Korea repeatedly through 2026 specifically to deepen partnerships on physical AI, forming a Physical AI Startup Alliance with Korean venture capital firms, engaging Doosan Robotics on Isaac and GR00T integration, and partnering with LG on motor technology and mechanical systems for a jointly developed humanoid due in early 2026.
| Company | Physical AI position | Stage (as of Sept. 2026) |
|---|---|---|
| Rainbow Robotics (Samsung-affiliated) | Mobile dual-arm manipulator (RB-Y1) | Field trials at Coupang logistics centers |
| RLWRLD | Hardware-agnostic manipulation foundation model (RLDX-1) | Showcased at Computex 2026; hyperscaler engagement |
| Doosan Robotics | Collaborative robots + Nvidia Isaac/GR00T integration | Partnership development, Nvidia Innovation Center visit Apr. 2026 |
| Hyundai Motor Group / Boston Dynamics | Full humanoid (Atlas) + Google DeepMind AI stack | 2026 production allocated; Savannah factory deployment from 2028 |
| LG Electronics / Nvidia | Home humanoid (CLOiD), Jetson Thor chipset | Unveiled CES 2026; joint humanoid due early 2026 per MOU |
Source: Korea Times, Korea Herald, KoreaTechDesk, Seoul Economic Daily, Korea Business Review, various dates 2026. Company-reported deployment stages; independent third-party verification of production volumes is limited for early-stage programs.
4. The Two-Sided Asset: World-Class Data, World-Threatening Demographics
Korea's case for a leading role in physical AI rests on a genuine structural asset that is difficult for any software-first competitor to replicate quickly: five decades of dense, high-precision manufacturing operating data across semiconductors, shipbuilding, automotive assembly, petrochemicals, and batteries. A shipyard's welders and an automaker's line workers have generated, through their bodies, the exact kind of fine-motor, contact-rich, variable-condition task data that trains a robot's manipulation model — if it can be captured, standardized, and fed into a learning system. No US or Chinese software company owns that data; it exists only where the physical work has actually been done at scale, which is disproportionately Korea, Japan, and to a lesser extent Germany.
The uncomfortable other half of that same story is that Korea is running out of the workers who generate that data and who will need to be replaced by the robots trained on it. Korea's working-age population (ages 15-64) peaked in 2019 at roughly 37.6 million and has been declining since; the IMF's most recent Article IV consultation on Korea projects it falling to roughly 34 million by 2030 and as low as 16 million by 2070. The proximate cause is well documented: Korea's total fertility rate fell to a record 0.72 in 2023 before a marginal rebound to 0.75 in 2024, still by far the lowest recorded rate among OECD economies and less than half the replacement level of 2.1 children per woman needed to hold a population steady.
4.1 Desperation as a strategic asset, not just a liability
It would be a mistake to read Korea's demographic trajectory purely as a threat. In an industry where the single biggest obstacle to deployment is often social and institutional resistance — unions worried about job loss, regulators cautious about safety liability, a workforce reluctant to trust an unfamiliar machine — Korea's demographic reality removes much of that resistance by making automation an existential necessity rather than a cost-cutting preference. The same logic explains why Japan, facing a comparable if somewhat less extreme demographic profile, has also emerged as an early mover in industrial physical AI. A country that cannot staff its factories in the 2030s regardless of policy choices has little room for the multi-year, committee-driven deliberation that has historically slowed automation adoption in economies with more labor-market slack.
5. Korea's Execution Gap: The Real Competitive Risk Isn't Technology
If Korea's core vulnerability were only demographic, the policy response would be straightforward: fund automation, remove barriers, wait for the robots to arrive. The evidence from 2026 suggests a second, more uncomfortable vulnerability sits alongside the demographic one — a persistent gap between Korean industry's technical capability and its willingness to deploy imperfect technology quickly. Multiple industry analyses of the 2026 physical-AI landscape have identified this pattern explicitly: over roughly the past five years, a meaningful share of Korean corporate and policy discussion around large language models and robotics centered on whether the technology was mature enough, safe enough, or commercially proven enough to justify large-scale investment — a caution that, while individually reasonable, compounds into lost time when a technology's capability curve is moving as fast as generative and physical AI have moved since 2023.
The contrast frequently drawn in Korean industry commentary is with Chinese manufacturers' willingness to deploy visibly imperfect humanoid prototypes directly into real operating environments, accept a higher failure rate, and iterate based on the resulting field data — a strategy that trades short-term reputational and safety risk for a faster real-world learning curve. Whether that trade-off is prudent is a genuinely open, evidence-dependent question that reasonable technologists dispute; the point for Korean strategy is narrower: a five-to-ten-year hardware and manufacturing-data advantage can be substantially eroded by a two-to-three-year deployment-speed disadvantage, because physical AI's most valuable asset — real-world interaction data — compounds only when robots are actually out in the field generating it.
6. The Alliance Model: What Hyundai-Boston Dynamics-Google-Nvidia Gets Right
Korea's most instructive recent case study for how to close a technology gap without pretending to build every layer domestically is Hyundai's approach to Atlas, and the contrast with the country's earlier, more difficult experience in autonomous vehicles is explicit in how the company has talked about the decision. Korean and global automakers collectively spent tens of billions of dollars over the past decade attempting to build full self-driving capability largely in-house, with limited commercial success at Level 4/5 autonomy; the core lesson many in the industry now draw is that a hardware manufacturer's core competency — precision mechanical engineering, safety-certified manufacturing, supply-chain execution — does not automatically transfer into building frontier AI software, which is a different discipline requiring a different kind of talent density and iteration speed.
Hyundai's Boston Dynamics strategy applies that lesson directly: keep the hardware, safety engineering, and factory-deployment expertise in-house and at Boston Dynamics, but source the frontier AI "brain" through a formal partnership with Google DeepMind, and source large-scale simulation infrastructure through Nvidia's Isaac and Omniverse platforms rather than attempting to build a competing foundation-model effort from scratch. That is a deliberate division of labor across three companies with three different core competencies, assembled into a single commercial product on a defined timeline. For a Korean industrial base built on manufacturing excellence rather than frontier AI research — Korea produces world-class chips and cars but has not produced a globally leading foundation-model lab — this alliance structure is arguably the only realistic path to competing at the frontier of physical AI without a decade-long, uncertain bet on building an in-house AI brain that may never catch the leaders.
6.1 Beyond individual robots: the "dark factory" endgame
The more ambitious and more Korea-relevant version of this strategy extends past deploying individual humanoids to what the industry calls "dark factory" operation — a facility where AI systems monitor and coordinate not just robots but the entire physical plant: conveyance shuttles, processing machinery, inspection sensors, and maintenance scheduling, diagnosing problems and redirecting resources with minimal human intervention. This is a systems-integration problem, not a single-robot problem, and systems integration at industrial scale — building and commissioning entire factories quickly and reliably — is a competency Korean industrial conglomerates (Samsung, Hyundai, SK, LG) have demonstrated repeatedly over decades of building semiconductor fabs, shipyards, and auto plants around the world. If Korea's advantage in dense operational data can be combined with this existing plant-construction and commissioning expertise, the resulting "dark factory as a service" capability would be difficult for a pure AI lab or a pure hardware manufacturer to replicate independently.
7. Government Response: A Real Strategy, Already Partly in Motion
The Korean government's policy response through 2026 tracks closely with what independent industry analysis has identified as the necessary interventions, which is worth noting explicitly because policy announcements in this space frequently outrun actual budget commitments. On August 27, 2026, Korea's Ministry of Economy and Finance confirmed, through its second Fiscal Strategy Council meeting, a financing plan allocating KRW 2.3 trillion to humanoid robot ecosystem development through 2030 and a further KRW 2.8 trillion to physical AI demonstration projects across eight core industrial sectors — a combined KRW 5.1 trillion in confirmed public investment, with KRW 600 billion budgeted for the first year alone.
The plan's structure follows a demand-pull logic rather than a pure R&D-subsidy model: large industrial "anchor" companies will co-develop humanoid robots specialized for ten target industries — including displays, aviation, logistics, distribution, shipbuilding, automotive, appliances, chemicals, hospitals, and hotels — providing field-testing sites and operational data, while smaller robotics firms handle specific sub-components (welding, painting, polishing, and inspection robots for shipbuilding, for example) that large conglomerates cannot economically develop alone, with mass production targeted from 2030. Separately, the Korea Development Institute has confirmed that a National Manufacturing Data Library project — explicitly designed to integrate and validate manufacturing data scattered across industries into a unified national infrastructure — is under active review for inclusion in the 2027 budget across the Ministry of Science and ICT and the Ministry of Trade, Industry and Energy.
| Policy pillar | Mechanism | Committed scale |
|---|---|---|
| Humanoid ecosystem development | Demand-led co-development with 10 anchor industries; anchor firms supply test sites and data | KRW 2.3T through 2030 (KRW 600B in year one) |
| Physical AI demonstration | Field pilots across 8 core sectors | KRW 2.8T through 2030 |
| National Manufacturing Data Library | Unified national standard for industrial operating data | Under review for FY2027 budget (Sci/ICT + Trade/Industry ministries) |
| K-Humanoid Alliance | Industry-academia-research consortium; robot AI foundation model and mass-production targets | KRW 1T+ targeted through 2030; foundation model by 2028, 1,000+ units/year by 2029 |
Source: Korea Ministry of Economy and Finance (Aug. 27, 2026); Herald Corp; KDI Economic Information and Education Center; Korea Business Review, June 2026.
8. Market Outlook: Two Scenarios Through 2030
What follows is a probability-weighted reading of where the currently available evidence points, not a directional prediction or investment recommendation.
Base Case (more probable given current evidence)
Humanoid deployment scales gradually rather than explosively through 2028-2030, consistent with Hyundai's own phased Savannah timeline (parts sequencing in 2028, assembly by 2030) and with the wide dispersion in market-size forecasts, which itself signals that the market has not yet found a stable growth trajectory analysts can converge on. Korean hardware manufacturers (Samsung-affiliated Rainbow Robotics, Boston Dynamics via Hyundai) and manipulation-software specialists (RLWRLD and peers) continue winning design partnerships with US AI labs (Google DeepMind, Nvidia) precisely because building a competitive foundation model from scratch remains prohibitively expensive and slow, while Korea's hardware and field-deployment access remains genuinely scarce. Government funding scales roughly as currently budgeted, supporting steady rather than dramatic acceleration in domestic deployment, and Korea's National Manufacturing Data Library becomes operational on a multi-year timeline that lands after, not before, several other countries attempt similar national manufacturing-data infrastructure.
Risk Case Worth Taking Seriously
Korea's structural execution gap — the pattern of deliberation outpacing deployment speed identified by multiple domestic industry analysts — persists into the physical AI era just as it did in earlier technology cycles, while Chinese manufacturers' faster, higher-failure-tolerance deployment strategy generates enough real-world field data to close the manipulation-quality gap faster than currently expected. Under this path, Korean firms retain their hardware and component advantages but increasingly supply parts and sub-assemblies into humanoid programs led by US, Chinese, or Japanese integrators, echoing the "supplier rather than protagonist" role Korea occupied in the first, LLM-driven phase of the AI buildout. The confirmed KRW 5.1 trillion government commitment, while substantial in absolute terms, proves modest relative to the scale of capital US hyperscalers and Chinese state-backed programs are prepared to deploy, and the National Manufacturing Data Library initiative remains stuck in budget review beyond FY2027 without securing the cross-ministry coordination the KDI briefing flagged as a prerequisite.
9. What Companies and Policymakers Should Actually Do
This section synthesizes the evidence above into concrete strategic priorities, organized by audience. It is analysis, not investment advice, and none of it should be read as a recommendation to buy, sell, or hold any specific security.
9.1 For Korean manufacturing conglomerates
- Treat operational data as an exportable strategic asset, not a byproduct. The welding, assembly, and inspection data accumulated over decades of shipbuilding and automaking is precisely what foreign AI labs cannot generate themselves at scale; licensing or joint-venture structures around that data, on Korean terms, are a stronger negotiating position than most conglomerates have historically assumed.
- Replicate the alliance model rather than the vertical-integration model. Hyundai's Boston Dynamics-Google-Nvidia structure suggests the winning configuration pairs a hardware-and-deployment specialist with an external AI-brain partner and an external simulation-infrastructure partner, rather than attempting to build all three competencies in-house on an uncertain timeline.
- Bias toward field deployment speed over internal certainty. The gap between Korea's demonstrated hardware quality and its comparatively slower deployment cadence is the single most correctable weakness identified in this analysis, and it does not require new technology to fix — it requires an organizational tolerance for supervised, contained field failures that generate the data needed to improve the next iteration.
9.2 For policymakers
- Prioritize the National Manufacturing Data Library's cross-ministry coordination over its total budget. The KDI briefing on the initiative explicitly flagged the need for coordination among the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, and others as the binding constraint, not funding size; resolving that coordination problem before the FY2027 budget cycle would likely do more to accelerate Korea's data advantage than a larger, less-coordinated appropriation.
- Convert regulatory caution into supervised, geographically bounded experimentation. Designating specific industrial zones (the kind proposed around Ulsan, Changwon, Giheung, and Pyeongtaek in current policy discussion) for negative-regulation, sandbox-style humanoid and dark-factory pilots addresses the deployment-speed gap directly, provided the zones carry real regulatory relief rather than becoming another layer of application and review.
- Fund the manipulation-software layer as deliberately as the hardware layer. Korea's hardware competitiveness is largely inherited from prior decades of industrial investment; its manipulation-AI competitiveness (RLWRLD and peers) is comparatively new and more fragile, and sustaining it requires continued compute access, data-sharing frameworks with anchor manufacturers, and talent retention specifically in the AI research layer where global competition for talent is most intense.
10. Basic Investment Principles
This section is deliberately general and is not a recommendation to buy, sell, or time any specific position, company, or sector.
- Separate demonstrated deployment from demonstrated demo footage. A robot shown dancing or sprinting on a trade-show stage is not the same evidentiary category as a robot logging supervised hours on an actual assembly line; Rainbow Robotics' Coupang trials and Hyundai's Savannah timeline are closer to the latter, and warrant more weight accordingly.
- Diversify across the physical AI value chain rather than concentrating in headline humanoid names. The value chain includes hardware integrators, manipulation-software specialists, simulation infrastructure providers, and the industrial conglomerates supplying field-test data — each with different risk profiles, timelines, and customer concentration.
- Weight the wide dispersion in market-size forecasts as a signal of genuine uncertainty, not analyst error. When five reputable research firms produce 2026 market-size estimates ranging from roughly $1.4 billion to $10.7 billion, that dispersion itself is informative about how immature current measurement and forecasting methodology is for this specific market.
- Maintain a long time horizon. Every credible deployment timeline referenced in this piece (Hyundai's Savannah plant, Korea's National Manufacturing Data Library, the K-Humanoid Alliance's mass-production target) lands in 2028 or later; near-term share-price enthusiasm and underlying commercial reality are likely to diverge repeatedly before then.
This article is also published, with an independent, retail-investor-focused angle for Korean-language readers examining which specific listed companies stand to benefit, 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.
Related Reading
• SK Hynix Hits All-Time High: Is the HBM Super-Cycle Just Getting Started? — on the memory-chip windfall this piece argues is a temporary, not permanent, foundation for Korean tech leadership
• AI Boom vs Energy Crisis: The Battle Shaping Global Markets in H2 2026 — on the energy infrastructure constraints underlying the AI-factory buildout referenced in Section 1
- International Federation of Robotics, "Robot Density Surges in Europe, Asia, and Americas" / "World Robotics 2025" report, Apr. 8, 2026
- The Robot Report, "IFR reports robot density increase across Europe, Asia, and the Americas," Apr. 9, 2026
- InvestKOREA (KOTRA), Korean automotive-sector robot density data, cited via Seoulz industry analysis, May 6, 2026
- IMF, "Republic of Korea 2024 Article IV Consultation," Annex: "Addressing Korea's Declining Labor Force," Feb. 5, 2025
- CNBC, "South Korea's birth rate collapse threatens growth," Sept. 27, 2025, citing Statistics Korea 2024 fertility data (0.748)
- Statistics Korea (KOSTAT), total fertility rate historical series, cited via CEPR VoxEU and PMC/NCBI demographic analyses, 2025-2026
- Boston Dynamics, "Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry," press release, Jan. 5, 2026
- Axios, "Hyundai plans 30,000 humanoid robots a year for factories by 2028," Jan. 5, 2026
- Forbes, "Atlas Humanoid Robots Production 'Fully Committed' For 2026, Factory Will Build 30,000 Per Year," Jan. 6, 2026
- TechRadar, "Boston Dynamics' Atlas humanoid robot is now a product and heading to factories in 2028," Jan. 2026
- SiliconANGLE, "Hyundai and Boston Dynamics join forces to bring humanoid robots to factories," Jan. 6, 2026
- The Robot Report, "AW 2026 features Korea humanoid debuts as industry seeks digital transformation," Mar. 4, 2026
- Korea Herald, "Why Nvidia is betting on Korea for physical AI," Jun. 3, 2026
- Korea Times, "Korea's robotics alliance with Nvidia boosts major tech firms," Jun. 3, 2026
- Korea Business Review (KBR), "한국 휴머노이드 로봇 기술, 어디까지 왔나," Jun. 22, 2026
- KoreaTechDesk, "NVIDIA and Korean VCs Launch 'Physical AI Startup Alliance,'" Nov. 1, 2025
- Seoul Economic Daily, "Koo and Huang Forge 'Physical AI' Alliance; LG-Nvidia Humanoid Due Early 2026," Aug. 14, 2026
- Herald Corp, "휴머노이드 2.3조·피지컬AI 2.8조 투입...2030년 10대 업종 특화로봇," Aug. 27, 2026, citing Korea Ministry of Economy and Finance
- KDI Economic Information and Education Center, "휴머노이드 생태계 구축 본격화... '국가제조데이터라이브러리 사업', '27년 범부처 예산 반영 추진," 2026
- Fortune Business Insights, Grand View Research, MarketsandMarkets, The Business Research Company, Future Market Insights — humanoid robot market forecast reports, various dates 2026 (see Figure 1 for consolidated range)
Comments
Post a Comment