The Humanoid Robotics Boom

Opportunities and Threats as the Industry Grows

The Humanoid Robotics Industry

For decades, industrial automation has been dominated by stationary robotic systems designed for high precision and repeatability within tightly controlled, task-specific workflows. Today, some companies are attempting to shift the industry towards mobile and general-purpose robotic machines, driven by the desire for greater flexibility across less structured environments. The form that has been receiving the most attention as of late is the humanoid (a non-human entity with human form or characteristics). Some of this fascination is certainly due to dreams of recreating something similar in function and form to ourselves, but there is straightforward logic from an economic and productivity front. The built world was designed for human bodies. Doorways, staircases, tools, workstations, vehicles, warehouses, all of which are optimised for a bipedal form (using only two legs for walking) roughly 1.5 to 1.9 metres tall with two arms and two hands. A robot that can operate in these environments without modification to the environment itself unlocks an addressable market orders of magnitude larger than traditional industrial automation.

The convergence of several enabling technologies such as large language models for natural language instruction, transformer-based vision systems, improved battery energy density, and cheaper high-torque actuators (component of a machine responsible for moving or controlling a system by converting energy into physical motion), has made humanoid robots technically plausible for the first time. The question is no longer whether they can be built, but whether they can be built at a cost and reliability level that justifies commercial deployment. That question remains open.

Research firms are producing enormous numbers and they disagree wildly. Projections for the global humanoid robot market by the mid-2030s range from as little as US$4 billion (Grand View Research) to US$182 billion (Future Market Insights), a gap of roughly 45x. MarketsandMarkets forecasts growth from US$2.92 billion in 2025 to US$15.26 billion by 2030, while other estimates are far more aggressive, reaching US$38 billion by 2035. At the conservative extreme, DIGITIMES Research sees humanoid robots at just 0.2% of the global robot market in 2025, rising to 2% by 2030. Morgan Stanley anticipates slow adoption until the mid-2030s, accelerating only in the late 2030s and 2040s. The difference between the floor and ceiling is the difference between a niche industrial subsegment and a civilisation-altering technology. The range should indicate to investors a clear situation: in this field, nobody really knows anything for certain. The most basic commercial questions of who will buy these robots, at what price, what tasks can they perform and what will be their functional capabilities all remain unanswered. As a result, we believe that investors must be prudent and grounded, while also ensuring they embrace the reality of progress and not hide from it.

Global humanoid robot installations reached an estimated 16,000 units in 2025 and the market is currently dominated by Chinese manufacturers as 4 of the top 5 players in global market share (AgiBot, Unitree, UBTECH, Leju Robotics) come from China. These players and more Chinese firms are shipping in volume at price points between US$5,000 and US$100,000 that Western competitors struggle to match. The cost advantage is structural, underpinned by lower labour costs, government subsidies, and vertically integrated domestic supply chains for actuators, batteries, and sensors. However, before we get ahead of ourselves, it’s important to note that these numbers really don’t mean much for now. Humanoids in the automotive sector, the first major deployment vertical, remain in early pilot testing, performing basic tasks like badge labelling, material handling, visual inspection. These are tasks that existing industrial robots already perform capably. The humanoid form factor’s genuine value proposition of operating in environments designed for humans, using human tools, interacting safely alongside human workers has not yet been demonstrated at commercial scale. In other words, the technology and use cases benefits are not genuine yet, but rather the question is can it be achieved in the future and if so, how much is it worth?

The current humanoid field extends beyond these volume leaders with Western companies also investing aggressively across the sector. Tesla’s Optimus program is transitioning from internal prototyping toward commercial production. On its Q4 2025 earnings call in January 2026, Tesla announced it would end Model S and Model X production by mid-2026 and repurpose the Fremont, California assembly lines for Optimus manufacturing, with a long-term target of one million units per year. The Gen 3 Optimus prototype is expected to be unveiled in Q1 2026, with start of production targeted before the end of 2026 and consumer sales aimed for late 2027. Tesla’s aspirational unit price remains US$20,000-30,000 at scale, though most analysts expect initial commercial pricing in the US$100,000-150,000 range. However, progress has been slower than Musk’s forecasts: on the same earnings call, he acknowledged that no Optimus robots were yet performing useful autonomous work at Tesla facilities, despite having predicted thousands would be operational by the end of 2025.

Figure AI has become one of the highest valuation pure-play humanoid companies, raising over US$1 billion in its September 2025 Series C at a US$39 billion valuation, bringing total funding past US$1.9 billion. It’s Figure 02 completed an 11-month deployment at BMW’s Spartanburg plant, running daily 10-hour shifts and loading over 90,000 parts. The company operates a Robot-as-a-Service model at approximately US$1,000 per robot per month and has ended its OpenAI collaboration, now building AI entirely in-house through its Helix platform.

Agility Robotics leads in commercial deployment where its bipedal Digit robot reached 100,000 totes moved at a GXO Logistics facility in late 2025, widely regarded as the first full-time commercial humanoid deployment. Agility closed a US$400 million Series C in March 2025 (total funding approximately US$641 million), with unit pricing estimated at US$250,000-plus alongside a RaaS subscription option. Its RoboFab facility in Oregon has capacity for 10,000 units per year.

Boston Dynamics unveiled the production version of its all-electric Atlas humanoid at CES 2026. Atlas is an enterprise-grade bipedal robot designed for industrial tasks such as material handling and order fulfilment, standing 1.9 metres tall with a 2.3-metre reach, 30 kg lifting capacity, four hours of battery life with autonomous self-swapping, and 56 degrees of freedom enabling movement beyond human range. All 2026 production is committed to Hyundai’s Robotics Metaplant Application Center and Google DeepMind, with a DeepMind partnership integrating Gemini Robotics foundation models for improved perception and autonomy. Hyundai (approximately 80 per cent owner) plans broader factory deployment by 2028 and is building a dedicated robotics facility with 30,000-unit annual capacity. Unit pricing has not been disclosed.

Sanctuary AI’s Phoenix, now in its eighth generation, is powered by Carbon, a proprietary AI system combining symbolic reasoning with large language models. Partnered with Magna International for automotive manufacturing, Phoenix has demonstrated industry-leading dexterity (21-DOF hydraulic hands, near-human tactile sensitivity), though at approximately US$140 million in total funding it remains significantly less capitalised than competitors. 1X Technologies (OpenAI-backed) opened pre-orders for its consumer-focused NEO humanoid in October 2025 at US$20,000 (or US$499/month), with first US deliveries planned for 2026.

The fragmentation of the OEM landscape is an important structural feature for supply chain investors. A fragmented market usually means no single OEM has the scale or incentive to vertically integrate every component, creating persistent demand for some specialist suppliers. Conversely, if the market consolidates around one or two vertically integrated players, the specialist component supplier opportunity narrows dramatically although may not disappear.

Further Remarks

Whether humanoid robotics can achieve the technical benchmarks required for mass deployment remains uncertain, but the underlying demand case and potential benefits are compelling if even some of those technical goals are met. The fundamental promise is one of expanding the effective labour supply and quality. Robotic workers can operate continuously without fatigue, scale predictably without recruitment constraints, maintain consistent output quality, and perform dangerous or physically demanding tasks without injury risk and in many applications, they would simply be more efficient and proficient than humans. This matters because the need is already acute. Potential labour shortages are structural and worsening across developed economies. Ageing populations in Japan, South Korea, Germany, and increasingly the United States are creating a gap between labour demand and supply that immigration alone cannot fill, and global labour costs continue to rise. Furthermore, the long-term view is that much like agricultural innovation throughout history, this is not simply a story of replacement. By absorbing repetitive, physically intensive, or hazardous work, robotics frees human workers to move into roles where humans hold a genuine comparative advantage and value, thereby creating more total wealth. Many of these roles probably do not exist yet and require the human imagination to unlock.

It is important to remember that automation adoption, along with nearly all technological adoption, has historically followed a predictable S-curve: slow initial uptake, then rapid acceleration once cost-per-task falls below the human wage equivalent.

Some catalysts that investors and keen observers should track include Tesla’s push toward mass production of its Optimus robot, expected around 2026–2027, is among the most visible, and success would validate demand, drive supply chain maturation, and set pricing benchmarks for the industry. But Optimus is one program among many with many of the aforementioned companies progressing on different timelines and with different technical approaches. The competitive landscape is broad, and investors should be cautious about equating the industry’s prospects with any single company’s execution until we start seeing consolidation. In other words, as with all industries at a nascent or reinventment stage, its bit of a crapshoot with a lot of luck involved and to maximise skill-based decision making, you must follow the real value propositions of companies and their execution in delivering to customers.

The cost curve is critical. The key inflection point arrives when unit costs fall far enough that a robot’s total cost of ownership (purchase price, maintenance, energy, and software) becomes cheaper over its operational life than the fully loaded cost of the human labour it replaces. In warehouse and manufacturing settings, where annual labour costs per worker often run between US$40,000 and US$70,000 including benefits and overheads, a robot that costs meaningfully less than a single year’s wages and operates for several years reaches a compelling payback threshold. Several factors are driving costs downward: battery cost curves continuing their decade-long decline, actuator and sensor commoditisation, supply chain standardisation, and software costs amortised across larger fleets.

AI capability continues to expand what these machines can do, but another bottleneck lies in integrating perception, planning, and fine motor control in real-world environments. Current systems still struggle with deformable objects, truly unstructured spaces, and graceful recovery from the unexpected. The gap between demo performance and production-grade reliability remains significant, and closing it is arguably the single most important variable in the deployment timeline. Furthermore, not all applications are equal. Deployment in a controlled warehouse with structured layout, predictable tasks, limited human interaction is seemingly the relatively near-term opportunity. Aged care, construction, and retail will require substantially more capable systems. The likely path is not a single moment of mass adoption but a rolling sequence of sector-by-sector deployment with different S-Curves and different players or products involved in each area.

Regulatory frameworks will be another big part as there may be unrest from unemployment effects if these robots lead to increased layoffs and a decrease in hiring. Beyond this, there will be further questions about the safety standards and liability clarity that commercial deployment at scale requires.

The bull case is that humanoid robots follow the trajectory of smartphones: expensive early devices giving way to exponential cost reduction and a mass market (although unlike smartphones, there may be more space for different types of humanoid robot companies and products to take different parts of the market). The bear case is that they remain in technical and/or commercial limbo: perpetually near but never quite arriving.

Tesla as the Catalyst Driving the Sector (For Now)

It goes without saying that from the mainstream perspective, it is Tesla and Elon Musk that currently capture the imagination of most, and it is Tesla’s Optimus that induces the most excitement and attention.

As mentioned before, Tesla is undergoing a transition at their Fremont factory, repurposing production lines from the Model X and S to produce Optimus Gen 3. While it will take a while for Optimus to become available for sale, with the current projection being 2027 (and there’s probably a good chance it will be later), the signal is clear: Tesla believes they have a viable commercial product in their hands. Tesla’s stated production targets are ambitious: 1 million units per year at Fremont (the theoretical maximum capacity of the repurposed lines), and some reports expect scaling to at least 4 million per year at Giga Texas in the coming years, with 4-10 million per year across multiple facilities by 2028-2030. Musk has said that roughly 80% of Tesla’s future value will come from Optimus and related AI businesses. Whether or not one accepts that figure, the capital being deployed is real for now.

Optimus’ Functionality and Timeline

Gen 3 represents a meaningful technical step forward. Current demonstrated capabilities include walking at up to roughly 8 km/h, lifting and sorting small objects, scripted pick-and-place tasks in factory settings, autonomous self-charging, and improved navigation powered by Tesla’s Full Self-Driving neural network architecture. Over 1,000 Gen 2 and Gen 3 units are reportedly operating inside Tesla’s own factories, primarily on tasks like battery cell sorting and material handling.

The most significant hardware advance is in the hands, which now feature 22 degrees of freedom, which is double Gen 2’s 11. Degrees of freedom (DoF), in robotics, refers to the number of independent axes along which a joint or mechanism can move; a human hand is typically described as having 27, meaning each finger and the wrist can bend, rotate, and extend in multiple directions independently. Gen 3 achieves its 22 DoF by relocating the actuators (the motors that drive movement) from the hand itself into the forearm, connected via a tendon-driven system that mimics the way human forearm muscles control the fingers through tendons. This allows for a lighter, more dexterous hand capable of manipulating small and delicate objects. Tesla claims Gen 3 can perform over 3,000 discrete tasks, though this figure is unverified in unstructured real-world environments.

However, there are substantial limitations and challenges for Optimus. True autonomy in unstructured settings, the kind needed for reliable shift work in a factory, let alone household use, is unproven. At Tesla’s October 2024 “We, Robot” event, the company did not disclose that Optimus robots interacting with attendees were being remotely controlled by human operators, a fact later confirmed by then-program lead Milan Kovac. Balance on uneven terrain, manipulation of deformable or irregularly shaped objects, and sustained performance under real-world conditions (dust, moisture, impacts, thermal stress) all remain hard engineering problems. Battery life under continuous workloads is another open question. Roboticist Rodney Brooks, co-founder of iRobot, has described the vision of humanoid robots as general-purpose assistants as “pure fantasy thinking,” citing fundamental coordination challenges. Bridging the gap between a polished demo and an 8-hour shift is where most robotics programs stall. Tesla’s robotics (and automobile) ambitions have usually followed a pattern familiar to anyone who has tracked the company: directionally correct, unreliable on timing. Musk projected production-ready robots by 2023 at the original 2021 AI Day. In March 2025, he spoke of producing “at least one legion” (roughly 5,000 units) by year’s end, scaling to 50,000-100,000 in 2026. Actual 2025 output was only in the hundreds. Reporting from The Information revealed that Tesla hit major technical problems with the robot’s hands, leaving completed bodies sitting idle in factories awaiting hand and forearm components. Combine all of this with an increasingly competitive landscape, and Tesla is entering perhaps the most pivotal era of its robotics program.

The Optimus Supply Chain

*To be clear, much of the following is from reports and sources outside of official Tesla channels.

Despite Tesla’s aggressive vertical integration philosophy, Optimus relies heavily on a network of third-party suppliers. Musk acknowledged on the Q4 2025 earnings call that the robot uses “a completely new supply chain” with “really nothing from the existing supply chain” of Tesla’s vehicle business. He further noted that Tesla had “tried desperately” to use existing motors, actuators and sensors but “nothing worked for a human robot hand at any price,” forcing the company to “design everything from physics first principles.” This means Tesla cannot leverage its established automotive supplier relationships and must build new partnerships largely from scratch, a process Musk warned would produce a “longer and slower manufacturing S-curve” than products sharing components with existing lines.

One of the most striking features of the reported bill of materials for Optimus is the dominance of actuators. Linear and rotary actuators together account for an estimated 40-56% of total component cost, depending on the analysis. Morgan Stanley’s work places the figure at the higher end, while other Chinese brokerage research suggests a range closer to 44–45%. Sanhua Intelligent Control is widely reported to be the primary supplier of actuator assemblies for Optimus. In October 2025, Chinese media reported a 5 billion yuan (approximately US$685 million) order for linear actuators which is enough, by industry estimates, for roughly 180,000 units. Neither party confirmed it: Sanhua stated it had no material information to disclose, while Tesla China said there was no official information to share externally. A further 1.2 billion yuan order was reported in December 2025, estimated to cover around 43,000 additional units. The supplier landscape, however, has become less clear-cut over time. Initial Chinese brokerage research said Sanhua was the exclusive linear actuator supplier and Tuopu Group was the exclusive rotary actuator supplier. But the December reporting attributed supply of both linear actuators and all 14 rotary joints per unit to Sanhua, suggesting the changing dynamics or unreliable reporting. Tuopu, which began mass-producing self-developed rotary actuators in 2024 and has committed 5 billion yuan to a dedicated robotics production base in Ningbo, continues to be identified as a core Tier 1 supplier. Whether a genuine dual-supply arrangement is emerging or Sanhua is consolidating its position or there are other suppliers involved remains an open question. Both companies are established Tesla automotive suppliers who pivoted aggressively into robotics, and both saw significant share price moves on the back of Optimus-related reporting.

Beyond actuators, the supply chain includes TSMC and Samsung for AI chip fabrication (Tesla designs the chips in-house), CATL for high-density battery cells, Green Harmonics for harmonic reducers (a component critical for precise joint movement, where Green Harmonics holds over 60% of the domestic Chinese market, breaking the previous monopoly of Japan’s Harmonic Drive Systems) and Keli Sensing for six-dimensional torque sensors benchmarked against industry leader ATI. Keli remains in small-batch verification rather than full production supply. Tesla designs the dexterous hands in-house, and the FSD/AI chip at roughly a quarter of BoM cost (approximately 50,000 yuan per unit, or around 26.5% of total cost according to Chinese brokerage estimates) is Tesla-designed but externally fabricated.

The supply chain is overwhelmingly Chinese and Taiwanese. With the exception of Tesla’s in-house AI chip design and its US assembly operations, nearly every major Optimus component supplier is based in China or Taiwan. This creates meaningful geopolitical exposure and is a risk that has already materialised. In April 2025, China’s Commerce Ministry imposed export controls on seven rare earth elements and magnets, in retaliation for US tariffs. Musk confirmed on the Q1 2025 earnings call that Optimus production was being impacted by what he called the “magnet issue,” noting that China was requiring assurances that the rare earth magnets would not be used for military purposes. Tesla was working with Chinese authorities to secure an export licence, but the licensing process was described as opaque and potentially taking weeks to months. Public trade data compiled by supply chain intelligence firm Sayari indicates that Tesla appears to have sourced all of its neodymium-iron-boron (NdFeB) magnets from Chinese suppliers and there are currently no comparable non-Chinese suppliers for several key components at the required scale and cost. Supply chain diversification is a multi-year endeavour which means that for investors evaluating Optimus-related suppliers, there is both opportunity and risk due to geography.

The supplier relationships with Sanhua and Tuopu give those companies significant revenue visibility and near-term pricing leverage. However, they also carry the long-term risk that Tesla eventually brings actuator manufacturing in-house as volumes justify the capital investment or to find other alternative suppliers. This is a pattern well-established in Tesla’s automotive business with items like battery cells and computing chips. Sanhua’s own financial disclosures suggest awareness of this dynamic: the company has been working to reduce its Tesla revenue concentration from 35% to 28%, diversifying its customer base to avoid single-client dependence.

Supply Chain Hunting

The classic picks-and-shovels argument is that when a new technology wave emerges, it is often more profitable to invest in the companies supplying critical components than in the OEMs building the end product. The logic is that OEMs face intense competition, margin pressure, and execution risk, while component suppliers with proprietary technology can sell to multiple OEMs and benefit from the growth of the entire ecosystem regardless of which individual OEM wins. In humanoid robotics, this logic may be relevant. Currently, the OEM landscape is fragmented, competitive, and unprofitable. None of the humanoid robotics companies are generating meaningful revenue yet, and most are burning cash at venture-backed rates. Picking the OEM winner is nearly impossible. But every humanoid robot, regardless of manufacturer, requires the same categories of components: actuators, batteries, sensors, harmonic reducers, AI chips, and structural elements.

The Sensor Opportunity

Every humanoid robot requires extensive sensing: force/torque sensors at each joint for compliant motion control, tactile sensors in the hands for manipulation feedback, inertial measuring units (IMUs) for balance, encoders for joint position, and vision systems for navigation. The sensor stack is where the robot’s physical intelligence resides and without it, the robot is a blind, clumsy machine regardless of how sophisticated its AI model is. To understand why sensors matter so profoundly, consider that they are what close the loop between the robot’s AI brain and the physical world. Without them, even the most sophisticated neural network is essentially issuing commands into a void. Something as seemingly simple as picking up a coffee cup requires continuous real-time feedback at every stage of execution. Force/torque sensors in the wrist and fingers detect the moment of contact and modulate grip pressure as too little and the cup slips, too much and it shatters. The robot has no inherent sense of feel the way humans do; that entire channel of physical awareness must be synthetically recreated through sensor data. Without force feedback, the robot is guessing, and guessing doesn’t work when you’re handling eggs, assisting an elderly person, or inserting a component on an assembly line.

Balance presents another dimension entirely. Humans maintain upright posture through an extraordinarily complex feedback loop involving the inner ear, signals from muscles and joints, and visual cues and all processed subconsciously at remarkable speed. A bipedal humanoid must replicate this with IMUs measuring orientation and angular velocity, joint encoders tracking limb positions, and force sensors in the feet detecting ground contact and weight distribution. The robot’s balance controller is constantly making micro-adjustments based on this sensor fusion, hundreds of times per second. Remove any one input and the system degrades rapidly. For instance, a robot without foot pressure sensing cannot reliably detect uneven terrain or anticipate a stumble.

Then there is the navigation and spatial awareness layer. Vision systems and depth sensors allow the robot to map its environment, identify objects, avoid obstacles, and understand context. But vision alone is not sufficient for manipulation as you can see a door handle, but you cannot feel whether it is locked until you apply torque and sense resistance. The sensor types are complementary rather than redundant: vision tells the robot what and where; force and tactile sensing tell it how much and how carefully.

What makes this especially critical for the humanoid form is that these robots are designed to hopefully be able to operate in unstructured human environments like homes, warehouses, hospitals, where nothing is precisely positioned and conditions change constantly. An industrial robot arm on a factory floor can get away with less sensing because its environment is controlled and predictable. A humanoid walking through a cluttered room, picking up varied objects, and interacting safely with people needs rich, multi-modal sensory input just to function at a basic level. The AI provides reasoning and planning, but the sensors provide embodiment, the ability to exist in and interact with the physical world rather than merely issuing abstract motor commands. Every improvement in sensor quality, speed, and coverage translates directly into more capable, safer, and more adaptable robots.

The market numbers (there is a large range and many are speculative), for what they are worth, reflect this perceived importance. Force and torque sensors held approximately 28 per cent of the robotic sensor market in 2024, with strain-gauge technology accounting for 34 per cent of total robotic sensor technology by sensing method. The broader 6-axis force torque sensor market was valued at roughly US$225 million in 2023 and is projected to reach approximately US$2.3 billion by 2030, implying a CAGR of around 40.5 per cent. Within the humanoid-specific segment, the torque sensor market stood at US$515 million in 2024 and is forecast to reach US$7.87 billion by 2032 at a CAGR of roughly 48 per cent, while humanoid robotic sensors more broadly are projected to grow at a 38.5 per cent CAGR through 2030. If these projections are even directionally correct, the sensor market within humanoid robotics will grow from hundreds of millions to billions of dollars within a decade. The question for investors is which sensor companies will capture this value, at what margins, and with what competitive durability.

A Historical Examination

Before committing capital to any supply chain thesis in humanoid robotics, it’s worth examining one of the more recent and arguably most relevant historical parallels: the smartphone revolution’s component ecosystem. The smartphone wave created enormous aggregate value, but that value was distributed unevenly, and often counterintuitively, across the supply chain. Understanding why certain suppliers captured outsized long-term returns while others were marginalised despite genuine technological contributions offers a framework for evaluating opportunities in robotics.

ARM Holdings captured value across the entire smartphone ecosystem through its architecture licensing model, collecting royalties on virtually every application processor shipped regardless of which OEM won market share in any given quarter. Its position was defensible not merely because the instruction set architecture was technically excellent, but because an entire software ecosystem (operating systems, compilers, developer tools, application libraries) had been built around it over decades. Displacing ARM didn’t require designing a better chip; it required rebuilding an ecosystem, which is a categorically different and vastly more difficult undertaking. ARM’s moat was architectural and systemic, not merely technical. Its licensing model meant that each incremental smartphone sold generated royalty revenue at near-zero marginal cost, and its customer base spanned Apple, Samsung, Qualcomm, MediaTek, and dozens of smaller licensees. No single customer’s trajectory entirely determined ARM’s fortunes.

TSMC became the indispensable manufacturing platform powering every major fabless chip designer. Its defensibility derived from the compounding nature of leading-edge semiconductor fabrication: each successive process node demanded billions in capital expenditure and years of yield optimisation that couldn’t be shortcut by competitors willing to simply spend more. The knowledge embedded in TSMC’s manufacturing processes was cumulative and deeply tacit. Its economics scaled powerfully with volume because fab utilisation is the primary driver of semiconductor manufacturing profitability; more wafers processed across a fixed capital base meant expanding margins. And TSMC served Apple, Nvidia, AMD, Qualcomm, Broadcom, and a long tail of smaller customers.

Qualcomm dominated mobile baseband processors through a combination of deep RF engineering expertise and an enormous patent portfolio embedded in 3G and 4G cellular standards. Replicating Qualcomm’s position required not just matching its silicon design capabilities but navigating a load of essential patents that made independent entry legally prohibitive. Its Snapdragon platform scaled with global smartphone volumes, particularly as emerging markets drove unit growth in the mid-2010s. Qualcomm’s customer diversification was real but imperfect as it maintained meaningful revenue across Chinese OEMs like Xiaomi, Oppo, and Vivo, but its relationships with Apple and Samsung were periodically contentious and subject to legal disputes that introduced genuine earnings volatility.

There were definitely other factors at play, like counter positioning in business model in TSMC’s case, but the general framework is that winners possessed three traits: technology that couldn’t be easily replicated or brought in-house, a business model that scaled with volume rather than being diluted by it, and customer diversification that avoided existential dependency on a single OEM.

On the other hand, many component suppliers that were genuinely critical during the early smartphone era ultimately failed to deliver outsized long-term investment returns and not because their technology was unimportant, but because the structural characteristics of their competitive positions made them vulnerable to the very market growth that was supposed to be their tailwind. Audience developed noise suppression technology designed into Apple and Samsung flagships, but its function sat within a processing pipeline controlled by the chipmaker. When Qualcomm and MediaTek integrated good enough noise suppression directly into their chips, the need for a standalone solution disappeared. InvenSense pioneered MEMS gyroscopes with design wins at Apple, Samsung, and Nintendo, but motion sensing sat within a framework increasingly managed by the same chip designers who bought InvenSense’s components, meaning its customer was also its most dangerous competitor. Peregrine Semiconductor produced high-performance RF switches designed into handsets across multiple OEMs, but its function was one element within a module whose integration was driven by larger players like Skyworks and Qorvo who could absorb it into broader solutions. In an ironic twist that demonstrates the recurring nature of these aforementioned patterns, Skyworks and Qorvo, the very companies that absorbed discrete players like Peregrine, announced a merger in 2025 after coming under pressure from the same integration logic they once exploited, as Qualcomm bundled RF into its modem platform and Broadcom captured premium sockets through superior filter technology.

As mentioned before, the obvious framework is that winners possessed three traits: technology that couldn’t be easily replicated or brought in-house, a business model that scaled with volume rather than being diluted by it, and customer diversification that avoided existential dependency on a single OEM. These criteria are sound as a first filter but insufficient as a complete explanation. Several additional conditions distinguished companies that captured compounding long-term value from those that captured transient technological rents.

Position at an architectural chokepoint versus within a subsystem. The winners defined entire layers of the technology stack; ARM owned compute architecture, TSMC controlled the manufacturing platform, Qualcomm dominated connectivity. The losers occupied functional slots within subsystems ultimately defined and controlled by others. The critical distinction is between defining a layer of the stack and populating a slot within someone else’s layer. Companies in the latter position are inherently vulnerable to integration by the layer owner, who has both the incentive and the architectural vantage point to absorb adjacent functionality over time.

Ecosystem lock-in versus pure technical superiority. ARM’s advantage wasn’t simply that its architecture was hard to replicate technically; it was that switching costs were borne by the entire downstream software ecosystem, creating a coordination problem no single actor could resolve. The losers had advantages rooted in pure technical performance like better signal-to-noise ratios, more accurate motion sensing, lower insertion loss, which exist on a continuous spectrum where good enough alternatives can emerge. Ecosystem lock-in creates a discontinuous barrier where partial replication captures zero value.

Compounding versus static defensibility. TSMC’s moat widened with each process node. ARM’s ecosystem became more entrenched with each line of software compiled for its instruction set. Qualcomm’s patent portfolio expanded with each new cellular standard. The losers had defensibility that was static at best and eroding at worst as their leads were real at a point in time but didn’t compound. Being ahead technically is valuable; being ahead in a way that makes you more ahead next year is transformative.

Pricing power durability under volume growth. ARM’s royalty model captured a percentage of chip value regardless of ASP trends. TSMC’s pricing power increased at advanced nodes as fewer foundries could follow. The losers faced annual price-down negotiations of five to fifteen per cent that eroded margins even as volumes grew. The relevant question is not “does revenue grow with industry volume?” but “does the supplier’s economic share of each unit remain stable or expand as the market matures?”

Standard-setting influence. The winners defined or heavily influenced the standards around which the industry organised. Standard-setting creates a self-reinforcing dynamic where the industry’s technical direction inherently benefits the standard-setter. Suppliers that merely conform to standards set by others lack this structural tailwind and are perpetually adapting to specifications whose evolution may not serve their interests.

Being a critical component supplier in a new technology wave is necessary but not sufficient for outsized investment returns. The smartphone parallel suggests investors should evaluate any picks-and-shovels opportunity against a more demanding set of criteria than mere technological importance: whether the company sits at an architectural chokepoint or within someone else’s subsystem; whether its defensibility compounds over time or erodes as integration advances; whether its pricing power survives OEM negotiating leverage as volumes scale; whether it benefits from ecosystem lock-in that creates discontinuous switching costs; and whether it influences the standards around which the industry organises. Companies that satisfy most of these conditions tend to capture a disproportionate share of ecosystem value. Companies that satisfy only the first-order criterion of technological relevance tend to be acquired, compressed, or integrated away before the market’s full value is realised. In other words, to put it simply, what is the moat? In order to earn outsized returns, a company has to provide a valuable product but then it also has to have a moat to maintain its superiority in delivering better products, as well-resourced and well-motivated players will move to take their margin if they can replicate or replace them.

*Disclaimer: This information is for general informational purposes only and does not constitute financial, investment, or professional advice. The author may hold positions in the assets or companies discussed.

Bonnefin Research is a free publication focused on better investment thinking. Subscribe and share to support the work, and join the discussion in the comments as we continue building this community.