Stor analyse fra Goldman: Næste bølge bliver humane robotter: Disse aktier kan blive vinderne

Uddrag fra Goldman Sachs:

According to Goldman’s Asian trading desk, AI’s next leg is shifting from chips to real-world deployment, and humanoid robotics is emerging as the clearest monetization frontier; in this context, structural tailwinds (labor shortages, automation demand) are accelerating adoption, and driving investor rotation into robotics-linked names across Korea, Japan and China. The bank’s long-running investment thesis is that with structural demand accelerating and Asia trading at a compelling valuation discount despite stronger growth, “this is an early-cycle opportunity to position ahead of a multi-year capital rotation into the robotics ecosystem.”

We will have more to say on how Goldman is trading the Asian “shift from the infrastructure to the application layer” shortly, but first let’s take a closer look at the fundamental research that compelled Goldman to make this call.

In a note published overnight by the bank’s equity strategist Jacqueline Du (“Mid-year check-in: Several steps closer toward commercial reality” ), who has been leading the bank’s robotic coverage beat (much more on this below), during the GS Asia Communacopia + Technology Conference in Hong Kong on 18–19 May and the bank’s China AI Robotics Trip in Shenzhen and Beijing on 20–22 May, Goldman met with 14 robotics companies across private and public markets. These firms represent a broad cross-section of China’s embodied AI, robotics, and automation ecosystem, including: Daimon Robotics (private), Dobot (2432.HK), Estun Automation (2715.HK/002747.SZ), Galaxea Dynamics (private), Galbot (private), Geek+ (2590.HK), LimX Dynamics (private), Linkerbot (private), Mech-Mind (private), One Robotics (6600.HK), PaXini (private), Spirit AI (private), UBTech (9880.HK) and X Square Robot (private).

According to Du, this trip provided “a mid-year check-in on the latest industry developments” and the bank came away encouraged to see the fast integration of VLA/VTLA (Vision-Language-Action/Vision-Tactile-Language-Action) models with world models for improved planning and robustness, with model scales quickly trending larger despite ongoing iteration needed for deployment-ready quality.” That said, high-quality,
real-world data remains the primary bottleneck, driving a higher consensus on scalable, human-centric data acquisition, with investments in both centralized data factories and distributed deployment-loops (via UMI and egocentric POV etc), and an expectation of significantly  increased data-related revenue.

Commercialization scope is broadening though is mostly at the proof-of-concept (POC) stage (especially for industrial and logistics applications) rather than large-scale deployment, which most industry players expect to materialize by 2027-2029, once they have accumulated tens of millions of hours of high-quality data on top of a deployment-ready model. In the meantime, cost reduction is primarily being achieved through scale and full-stack R&D control.

Despite the current challenges, Goldman views the long-term investment outlook for this sector as “highly promising.” The advancements discussed, particularly in multimodal AI stacks and sophisticated data acquisition, indicate that the industry is moving several steps closer to practical, widespread deployment. However, “this journey still requires patience, as companies navigate the complex transition from proof-of-concept to large-scale commercialization, with consistent quality and cost reduction being key milestones.”

Here are the key takeaways:

• Model discussion is moving beyond a narrow VLA-only framing toward execution-oriented multimodal stacks: specifically, rapid VLA-world model integration followed by VLA refinement, such as VTLA i.e. adding touch where physical interaction quality matters most. World models were discussed less as a standalone model category and more as a functional layer alongside action models: VLA or VTLA handles policy and action generation, while world-model capabilities improve whether, when, and how actions should be executed through next-state prediction, action validation before commitment, and stronger planning and robustness under real-world uncertainty. Companies that clearly framed the next step as some form of VLA/VTLA + world model combination include: Galaxea, Galbot, Spirit AI and One Robotics. Against that backdrop, model scale is moving higher, with discussions around larger stacks in the c.40B-80B range vs. smaller single-digit-billion-parameter pre-training systems. Players emphasized that multiple rounds of iteration are still ahead before these stacks are at deployment-ready, consistent quality.
• High-quality, real-world, multi-dimensional data remains the primary bottleneck to practical deployment, but the discussion has clearly moved beyond broad “data recipe” debates toward what scalable data acquisition architecture can reliably produce that data. Within that shift, human-centric and ego-centric collection increasingly looks like the preferred real-world collection method for high-fidelity data, especially where companies care about preserving natural motion, contact-rich interaction, and cross-embodiment transfer. In terms of on-the-ground investment, some players have moved toward purpose-built data infrastructure at factory scale with government support; these include PaXini, which currently operates five data factories nationwide. Others such as Galaxea, Spirit AI and One Robotics appear to be building more distributed deployment-loops through deployed systems, wearables, VR, and client-side collection. Multiple companies are guiding toward a higher data-related revenue mix into 2026, including government data factory demand that remains supportive (2026 data factory demand to remain as strong or stronger, per UBTech).
• Commercialization is broadening across industrial handling, logistics-style workflows, and a narrower set of structured commercial scenarios; industrial adoption remains staged from POC to pilot to small-batch rollout, and only then to larger-scale deployment. Key near-term opportunities that stood out include: sorting, material handling, pick-and-place, inspection / testing, and other standardized or semi-structured workflows. Per multiple companies’ comments, industrial deployment necessarily follows a multi-step conversion starting with POC  (typically 3-6 months, avg. 2-3 rounds), small-batch testing (typically sub-50 unit scale per factory order), and around 12 months of validation, before finally pilot deployments begin to move order sizes toward roughly 50-100 units per customer.
• Cost reduction continues and is being pursued through company-specific choices around architecture, component scope, and deployment form factor, but scale remains the primary driver of savings. Among full-humanoid players, full-stack R&D control remains the most common cost-control approach. It is worth highlighting that, likely on both model capability constraints and cost considerations, Goldman noted that many players seem to prefer wheel-base robot plus two-to-three finger grippers as the more reasonable choice for the time being which can cover 70%-90% of industrial applications, while they don’t rule out potential bipedal humanoids plus five-finger dexhands into the future.

Below we excerpt from Goldman’s detailed takeaways by company. For the full breakdown, including a much more indepth analysis of each company’s unique characteristics, please see the complete report available to pro subs.

Daimon Robotics (private) 

• Model/embodied intelligence stack: On model strategy, instead of building the comprehensive VLA model, Daimon focuses on tactile small models or a plug-in/pre-training layer that integrates into other VLA/VTLA frameworks. Daimon’s products will not solely be sensors but a solution including tactile sensing hardware, data and tactile small models. Daimon believes the visuotactile technology route is more compatible for model integration as it shares the same image frame with VLA models.

Dobot (2432.HK)

• Model/embodied intelligence stack: Dobot’s technology route emphasizes an one-brain-multi-form approach, aiming for one model to generalize across 50+ scenes and cover industrial needs across multiple robot forms. The high-level model will act as an intelligence hub for data reading, task distribution and coordinated control, e.g. dispatching a humanoid for handling and a quadruped for logistics. Dobot has in-house VLA while a world model is also in its R&D pipeline, developing for model tuning. The overall strategy runs from pre-programming 1.0, to deep-learning 2.0, to AI-driven embodied intelligence 3.0, with mgmt arguing body-hardware upgrades in return will push software-capability progress.

Galaxea (private)

• Model/embodied intelligence stack: Galaxea’s full technology stack covers an embodied AI base model (VLA + world model), data and the entire robot body designs, with mgmt emphasizing a one-brain-multi-body route. They added that the model’s ability to transfer across robot bodies is a strong emphasis but not applicable yet as it still requires post-training iteration. On the world model progress, Galaxea just released its Fast-WAM in Mar 2026 with low latency of 190ms and much higher efficiency than many others, per the company.

Geek+ (2590.HK)

• Model/embodied intelligence stack: Geek+ emphasized the importance of a “scenario-first” data/mindset when it comes to logistics model strength: leverage real warehouse task flows and data loops to guide model/application development. In practice, complex tasks are decomposed into smaller, trainable sub-tasks with defined boundaries (e.g. picking workflow: 2D vision → 3D reconstruction → segmentation → grasp execution), prioritizing reliability over generality. Regarding its embodied intelligence subsidiary (established in Jul 2025, to focus on R&D and related products for embodied intelligence, including AI-powered robotic arm picking), mgmt. emphasized that model/data/technology routes have not converged in the industry; hence, as opposed to making heavy upfront bets, the current focus remains on deployable, bounded-use cases rather than frontier embodied AI.

LimX Dynamics (private)

• Model/embodied intelligence stack: LimX’s model is a three-layer system with system 0 as the whole-body motion control, system 1 as the humanoid VLA skills and system 2 as the embodied agentic OS (COSA). The company’s model thesis is to decompose complex environments into learnable skills rather than rely on remote control. On Oli and TRON 2’s uneven-terrain stability, mgmt. said the action model starts from existing video-generation models, extracts motion information from video, and distills it into a simplified action-guidance model.

Linkerbot (private)

• Model/embodied intelligence stack: Beyond positioning as a dexterous-hand player, Linkerbot is building its own embodied big model that converts 2D objects/scenes into 3D operation. Per mgmt., the AI business splits into a faster-monetizing demonstration-free workstation solution that augments existing hardware via algorithms, and the longer-horizon embodied-intelligence track.

Mech-Mind (private)

• Model/embodied intelligence stack: The company’s model focus is on delivering a fast perception-to-decision loop in industrial settings. Per mgmt., their current 3D vision system runs a ~0.3-0.4s cycle, covering coordinate capture (XYZ), real-time learning, and decision output. In practice, this is commercialized as an integrated intelligence layer bundled with robotics, rather than a standalone vision product.

One Robotics (6600.HK)

• Model/embodied intelligence stack: Target architecture is end-to-end embodied AI, moving from modular systems toward VLA + world model integration (with VLM as a supporting layer). OneModel 1.7 (launched on 21 May) uses a latent world action model, combining world-model generalization with VLA execution. Additional modules include motion-centric control for household tasks and a “success memory” layer that recalls prior successful executions.

PaXini (private)

• Model/embodied intelligence stack: Mgmt. emphasized a focus on touch inn addition to vision for physical interaction, citing better alignment with force-based tasks vs. vision-heavy VLA approaches. Plans to launch a touch-led VTLA (“tactic”) model next month to complement customers’ vision-centric stacks.

Spirit AI (private)

• Model/embodied intelligence stack: Spirit’s core thesis is VLA plus world-model fusion (VLA leaning to understanding, world model to prediction) on a latent-prediction (not frame-by-frame) route. On a verifiable basis, Spirit v1.5 (open-sourced in Jan 2026) topped the RoboChallenge Table30 at 66.09 points / 50.33% success as the first Chinese open-source embodied model to surpass Pi0.5. With a rapid expansion of pre-train data volume and improving foundation model quality, the post-train time has compressed sharply.

* * *

For those relatively new to the topic of humanoid robotics, below we share some key commentary and catch-up charts courtesy of several must-read reports from Goldman Jacqueline Du (the full reports are referenced below).

We start with a snapshot of several robot manufacturers are entering the volume production stage, such as Tesla/Agility Robotics/Figure AI in the US and Unitree/AGIBot/UBTech in China.

Next, a significant number of companies (shown below) have publicly disclosed on their plans to enter and or further progress in humanoid robot related business development.

Behind Goldman’s favorable view is the assumption that Humanoid robots have the potential to become the next widely adopted terminal device (after smartphones/cars) with likely deflationary ASP and BOM costs to drive better product economics.

Within the robotics supply chain, Goldman prefers high content value products, such as harmonic reduction gears for higher growth potential and actuator assembly for higher product adoption certainty.

A closer look at where Goldman sees the most value in the supply chain:

Actuator assembly: Higher certainty of tech adoption in the high spec robot supply chain

• Auto manufacturers developing humanoid robots: For example, as Tesla targets rapid volume ramp up and cost reduction via design optimization and economies of scale to drive robot commercialization, the company will likely need to cooperate with Asia/China component suppliers, suggesting a need for an intermediate actuator assembler to manage supply chain relationships with various component suppliers. Multiple companies such as LeaderDrive and Best Precision have mentioned sending product samples to Sanhua or to Tesla directly. Actuator assembler’s positioning tend to have much lower risk from potential hardware configuration changes. We expect 70%/30% market share split for Sanhua/Tuopu vs. 50%/50% previously as we expect Sanhua to have a more dominant market share, referencing to its global market share of EV thermal management module assembly in Tesla
• Unlike auto manufacturers that can leverage their auto supply chain to develop humanoid robots, robot start-up companies do not necessarily have a need for an actuator assembler role especially given their current scale is likely still small.

Harmonic reduction gears: Higher technology barrier among key components given high requirements on precision, light weight and torque

• High spec: Harmonic Drive System and Leaderdrive are proven players in this product, which has long been applied in industrial robots/Cobots. We raise LeaderDrive 2025-30E market share prospect to 30% for high spec robots vs. Harmonic Drive System at a more dominant 70% (vs. previously 20%/80% split), and we expect the long-run potential split to reach 50%/50% beyond 2030E mainly considering China humanoid robot players’ growth potential that is likely to benefit Leaderdrive more than Harmonic Drive System given LeaderDrive’s competitive pricing and flexible capacity. That said, HDS will still likely maintain a dominant position among global names;
• Mid & Low spec: Given mid & low spec humanoid robots are more price-sensitive, we expect LeaderDrive to potentially win higher wallet share especially among China robot customers. However, we note that not everyone is adopting the harmonic reduction gear design, for example, Zhongda Leader/Guomao Reducer/Shuanghuan are providing planetary reduction gears, which could be better suited for lower-cost humanoid robots with less payload/torque requirements

Planetary roller screw: Rapidly changing landscape, yield rate, production consistency and capacity readiness in question

• High spec: Given Schaeffler’s longer track-record in the product development and with Sanhua also noting that Schaeffler is its main supplier now, we assume that Schaeffler possesses 50-60% market share in planetary roller screws, however, if humanoid robot functionality improvement pace exceeds our expectations, it could demand the entire industry to rapidly expand capacity, which could attract more entrants to this space. Though Seenpin (private)/Zhejiang XCC (603667.SS, Not Covered)/Beite (603009.SS, Non Covered) have shared their views of faster relative development progress as of late, we believe scope to earn significant market share in the long run remains uncertain. For other players, e.g. Hengli Hydraulic has stronger relative R&D capability and manufacturing technique know-how, Best/Sanhua are also purchasing top class grinding equipment to develop in-house manufacturing capabilities of planetary roller screw. Volume production requirements very much emphasize yield rate, production consistency, and capacity ramp-up abilities, which will differ from the current stage of sending samples. Thus, we believe the long-term market share winner in the space remains unclear.
• Mid & Low spec: Due to the lower payload requirements, mid & low spec humanoid robots will not necessarily use planetary roller screws.

Dextrous hands: Technology roadmap remains uncertain

• Moons’ Electric and Zhaowei (003021.SZ, Non Covered) both develop coreless motors for the dextrous hands and are attempting to deliver the entire hand module, including the encoders, gears, sensors, etc. Another potential player is Inovance, who announced previously that they will conduct R&D in the dextrous hand space, where we see Inovance with strong potential to develop such products with their servo motor know-how and technology foundation.
• Goldman notes that there are still a lot of uncertainties surrounding the technology roadmap with current options ranging from 1) 3 degrees of freedom per finger using compact harmonic reduction gears with a total of 22 degrees of freedom per hand; 2) one degree of freedom per finger connected by worm drive and planetary reduction gears with a total of 11 degrees of freedom per hand; 3) motors in forearm and cable-driven fingers with 20+ degrees of freedom.

In order to properly assess the revenue opportunities potential from global humanoid robot development with a varied technology road map and hardware configuration, Goldman has refined its forecast model and segment humanoid robot products into three categories based on system specifications, i.e. high-spec (e.g. Tesla’s Optimus), mid-spec (e.g. UBTech’s Walker-S and AGIBOT’s Expedition A1), and low-spec. The bank then applies its humanoid robot component investment framework in order to identify the likely beneficiaries and gauge the potential financial impact from humanoid-related business to our covered companies.

Goldman believes a global humanoid robot supply chain is taking shape, but notes that customer relationships are very dynamic at the
moment

Some TAM and valuation thoughts

Based on Goldman’s interactions with investors, it appears that the market consensus is that 1mn global humanoid robots shipments will eventually be achievable, given the vast demand potential, in manufacturing, service industry, elderly care, and so on; that is as long as the functionality and cost of the humanoid robot meets market demand. However, expectations on timeline to achieve this 1mn units level still varies (GSe by 2034E-35E in base case, 2030-31/2028-29 by bull case/blue sky scenarios)

The bank believes the market has likely factored in 500k global humanoid robot shipments by 2027E, assuming a 40X exit P/E multiple; this is on top of our expectations around respective companies’ content opportunity, market share prospects as well as sustainable margins. Below Goldman illustrates how the market is currently approaching potential valuation of humanoid robot supply chain stocks, consisting of the core business valuation and optionality of humanoid robot valuation.

Valuation upside potential would come from the core business plus humanoid robot valuation, which is derived from expectations on global humanoid robot shipments, content value per robot, market share involved with the company, the normalized NPM for the product and an exit P/E multiple. Valuation downside risk would come from the core business only, with zero optionality in humanoid robot valuation. The bank points out that AI robotics technology progress (foundation model primarily, plus certain bottlenecks in hardware) is very sensitive from both shipment outlook as well as exit P/E multiple (40X as typical multiple for high earnings growth ~30% CAGR stocks and up to 120x during 2009-2012 year for early EV development stage stocks).

Goldman’s valuation approach is different from the market: Given a technology inflection point for humanoid robots remains unclear, the bank forecasts 76k/502k units of global humanoid robot shipments by 2027E/2032E as opposed to the high expectations of the market. This forecast is predicated on Goldman’s view that it will likely take a longer period for AI-empowered robots to be ready, i.e. to achieve generalized tasks with high success rate and robustness, while we apply long-term valuations (2030E-based and discounted back to 2025E) considering the structural opportunity.

How to define a potential technology inflection point?

In 4Q24 and 1Q25, Tesla’s Optimus live demo during the ‘We, Robot’ event and Unitree H1’s debut on Spring Festival Gala showcased greatly enhanced hardware dexterity and robustness but the overall technology level has still yet to meet the accuracy/consistency/cost requirements for mass adoption in industrial or consumer settings, mainly due to underdeveloped generalized and autonomous AI capabilities. While a general-purpose AI robot remains challenging from a technology perspective in the upcoming year, we believe the technology inflection could come when humanoid robots are good enough in the following:

• Solving multiple generalized tasks: Generalization is the key difference between humanoid robots and traditional industrial robots. Therefore, we believe humanoid robots have to at least be proficient in several (say 5-10) generalized tasks before starting large scale applications.
• Operating with high success rate and continuous robustness: High success rate and robustness under long operating hours ensures stable output for humanoid robots, which is important for commercialization.
• Quick reasoning: Efficient reasoning and reacting of humanoid robots can greatly improve efficiency and interaction experience for customers.

Industry-wise, the new progress by OpenAI/NVIDIA and top players’ successful cooperation with auto OEMs/manufacturing enterprises likely attracted more humanoid robot players in 2024, especially in China. In 2024,15 new companies entered the humanoid robot industry and 17 new robot prototypes were launched (vs. 10 prototypes in 2023) in China. Top players in tech/auto/new energy industries like HUAWEI/BYD/CATL also announced their initiation of humanoid robot R&D projects. Compared to 2023, humanoid robot technology front runners had mainly started factory testing or initial commercialization (Figure/Agility/Unitree/AGIBOT/Leju) in 2024. With this further progress in humanoid robot and related businesses suggesting the sector may be approaching an industry volume production phase (from the R&D stage) into 2025.

Aside from the rapidly rising number of new humanoid robot manufacturing players, Goldman thinks the Physical AI R&D ecosystem launched by NVIDIA is another key accelerant for humanoid technology development. At the 2024 GTC in Mar, NVIDIA mgmt. launched edge computing SoC Jetson Thor and GR00T model plus Isaac platform leveling up their exposure in the humanoid robot industry. The Jetson Thor  is a high performance necessity for humanoid robots powered by AI, allowing robots to run pre-trained edge AI models on their own. Isaac is an R&D platform integrated with different tools and an optimized model development workflow, which generated NVIDIA’s in-house generalized model GR00T.

However, other obstacles remain, such as the lack of original data for training. Similar to LLMs, robotic AI models require a large amount of data to support model training and optimization. But unlike LLMs’ video/image/text data source, which has already been created through internet at a giga-scale, the important physical data for robotic AI models, such as force/torque/movement data remain scarce. Therefore, compensating the original physical data shortage has become a consensus need to boost robotic AI R&D. Given this need, companies are responding e.g. at the 2025 CES in Jan, NVIDIA mgmt. released another two R&D tools to accelerate robotic AI model evolution. Based on the previously launched Isaac and Omniverse, Isaac GR00T Blueprint and Cosmos are a simulation framework, a series of world foundation models and data pipelines that can work together to upscale synthetic 3D to real data. The large amount of synthetic data created by Isaac GR00T Blueprint and Cosmos can act as proxies of original physical data to speed up physical AI model training.

There is much more primary research, and we urge all pro subscribers interested to read the full coverage in the series available at the following links (please reach out directly for several other secondary reports in the series):

• Global Automation: Humanoid Robots III: The supply chain dynamism
• Global Automation: Humanoid Robot: The AI accelerant
• Global Automation: The investment case for humanoid robots
• Global Automation: Futures of Robotics: Apptronik: building versatile robots

Going back to Goldman’s sales desk, here is how Singapore-based trader Peter Sheren frames the trade reco:

Trade setup: AI narrative is shifting from infrastructure to the application layer, with humanoid robotics emerging as the next monetization leg; structural tailwinds (labor shortages, automation demand) are accelerating adoption, driving investor rotation into robotics-linked names across Korea, Japan and China.

Valuations: On a mean basis, both baskets trade at roughly ~29–30x P/E, appearing comparable. However, the Asia-Pac median (22.0x) is materially lower than the US median (28.0x), indicating that the “typical” Asian industrial/robotics stock trades at a ~21% discount to its US peer. The Asia-Pac mean is inflated by outliers like Leader Harmonious Drive (~95x) and Harmonic Drive JP (~85x).  The Asia-Pac basket’s median PEG of 1.5x vs. 2.0x for the US suggests that Asian robotics/automation stocks offer more earnings growth per unit of valuation. This is driven by higher expected growth rates in Chinese and Korean automation names (Ningbo Tuopu, Sanhua, Hengli Hydraulic) relative to their multiples.

Mutual Funds are starting to rotate into robotics-linked supply chains but the positioning remains early and concentrated in components rather than pure-play robotics. Flows are going into Korean and Chinese Auto components, China industrial automation / precision manufacturing and select robotics component suppliers.

Passive rebalancing is driving a sharp Korea rotation (Mobis inflow vs Hyundai Motor outflow), while Japan industrial automation faces broad-based passive selling, with liquidity-adjusted flows pointing to outsized price impact in mid-cap beneficiaries like BizLink and Hengli.

High-quality real-world data is the key bottleneck, driving investment into centralized data factories and distributed, human-centric collection, with data expected to become a meaningful revenue stream.  While commercial use cases are expanding across industrial/logistics workflows, adoption remains at POC/pilot stage, with scale deployment likely 2027–2029, and cost reduction primarily driven by scale and full-stack control.

Here is how the Goldman Asia Humanoid basket (GSXACHUM) has performed in the context of the Asia Power basket (GSXAPOWG). It has a long way to catch up.

Finally, for those who would rather just allocate capital once and forget about it, there is now an ETF for that: