← Back to All Research
Commission a Report
← Back to All Research
Commission a Report
Every major automation wave of the past two centuries produced the same reassurance for workers who used their bodies rather than their desks: your jobs are safe. The power loom destroyed hand-weaving but could not fit a pipe. The computer eliminated bank clerks but could not wire a panel. Software ate white-collar roles wholesale but could not diagnose a failing HVAC compressor in a residential loft conversion with a non-standard duct layout. The physical trades — plumbing, electrical, HVAC, construction — survived every wave not through political protection or regulatory moats but through a genuine technical barrier: dexterous manipulation in unstructured, novel environments was an unsolved problem in robotics.
That problem is now being solved. Not incrementally. Simultaneously, by multiple well-capitalised companies, on a five-to-ten year commercial deployment timeline. The two components that made physical trades genuinely automation-proof — hardware capable of dexterous manipulation in arbitrary spaces, and AI capable of reasoning about novel physical situations in real time — were both absent until approximately 2023. Both are now arriving together.
Tesla's Optimus, with over 1,000 units reported in factory deployment by end of 2025 and a stated target cost of $20,000–$30,000 per unit, is the most commercially credible path to volume production. Figure AI has raised $675M with a BMW factory partnership already active. Physical Intelligence has raised $400M focused specifically on the generalised manipulation policy model — the AI software layer that lets any robot learn new physical tasks from video demonstration. Boston Dynamics has released the electric Atlas commercially. The hardware-software convergence is occurring on an industry-wide basis, not as a single company's moonshot.
This report makes five arguments. First, the “trades are safe” assumption was correct until very recently and why. Second, the specific hardware landscape and the companies building it. Third, which trade categories face displacement and on what timeline. Fourth — and this is the dimension most investment research ignores — the social and political architecture that depended on the trades as the last mass-employment refuge for non-university-educated workers. Fifth, the investment implications: what to own, what to short, and why the bear case is more credible here than in knowledge-work automation.
The reassurance given to physical workers during previous automation waves was not dishonest. It was technically accurate. To understand why it is now expiring, it is necessary to understand precisely what the technical barrier was — and precisely what is now dissolving it.
The first industrial revolution eliminated domestic textile production but created factory work. The second eliminated factory piece-work but created assembly line roles. The computer eliminated typing pools, filing clerks, and bookkeepers but created data-entry operators, IT support staff, and software developers. Each wave destroyed one category of employment and created another, and the physical trades sat outside every wave's blast radius for the same fundamental reason: no machine could do what a plumber does.
The 1990s wave of manufacturing automation — CNC machining, industrial robots, programmable logic controllers — was sophisticated and consequential. But the robots operated within tightly controlled environments. A car factory robot arm performs the same motion, on the same part, in the same orientation, thousands of times a day. The moment anything falls outside that envelope — a component arrives at a slightly different angle, a fastener is stripped, the workspace is obstructed — the robot stops. It has no capacity to adapt.
A plumber arriving at a Victorian terraced house to replace a lead waste pipe under a kitchen sink does not know in advance what she will find. The pipe may run behind a load-bearing wall. There may be a junction not shown on any plan. The available access space may be eighteen inches at its widest. The material of the existing pipe may be lead, cast iron, MDPE, or something unidentifiable. The connection method for each is different. Tools must be selected, sized, and applied on the basis of real-time judgment in an environment that has never been exactly replicated anywhere on earth. This is what robotics researchers call an “unstructured environment problem,” and it was genuinely hard.
Robotics researchers who studied physical trades identified two distinct barriers, and both had to be crossed simultaneously for meaningful displacement to become possible.
The first is hardware dexterity: the ability to manipulate objects of varying shapes, sizes, and materials with the fine motor control that human hands provide. Human hands have twenty-seven degrees of freedom. They can apply precise force gradients — gripping a screwdriver firmly while adjusting torque feedback in real time, holding a pipe section steady while a second hand applies a wrench with calibrated force to avoid overtightening a compression fitting. Reproducing this in a robotic end-effector at a price point compatible with commercial deployment was not solved as recently as 2022. Robot hands either lacked degrees of freedom, lacked force sensing, lacked the compact actuators to place them in confined spaces, or were prohibitively expensive.
The second is AI situational reasoning: the capacity to perceive a novel physical environment, reason about what actions are required, select the correct approach from a large action space, handle unexpected obstacles or material states, and make safety judgments in real time. A robot that can physically grip a pipe is still useless if it cannot identify which pipe in a tangle beneath a sink is the correct one to replace, assess whether the existing joint will hold when the repair is connected, or recognise that the apparent access route runs through a section of flooring that shows signs of rot.
Boston Dynamics built Atlas starting in 2013. The bipedal, hydraulically actuated Atlas was genuinely impressive — it could traverse rough terrain, recover from being pushed, and perform backflips by 2019. But it was not a commercial product. Each unit cost an estimated $150,000–$200,000 to produce, required expert operators and maintenance technicians, ran on hydraulics that limited endurance and created reliability problems in dusty environments, and had no onboard AI capable of task reasoning. Atlas was a demonstration that bipedal locomotion was possible at human scale. It demonstrated nothing about the AI reasoning layer, because no such layer existed.
The position in 2020 was therefore clear: the hardware was improving but not commercially viable; the AI reasoning layer for physical tasks was essentially absent. The standard large language models of that era were trained on text and had no grounding in physical reality. They could describe how to replace a waste pipe in fluent prose but could not perceive, plan, or act in the three-dimensional world.
Three concurrent developments broke the logjam. First, large-scale transformer models extended from text to multimodal inputs including video, enabling AI systems to reason about physical scenes from visual data. Second, reinforcement learning from simulation — training robot policies in simulated environments at scale before physical deployment — produced manipulation policies that generalised to real-world conditions in ways that purely physical training could not. Third, hardware cost curves followed the pattern of every other technology exposed to venture capital and manufacturing scale: prices fell sharply as investment surged.
By 2024, the question was no longer whether the technical barriers would fall. The question became: at what price point, on what timeline, and with whose hardware and software stack does physical trade displacement become economically rational for the businesses deploying it?
The humanoid robotics landscape has consolidated around a small number of well-capitalised companies, each with a differentiated position on the hardware-software stack. The competitive dynamics are more complex than the factory automation wave of the 1990s: software is as important as hardware, and the company that owns the generalised manipulation policy model may ultimately matter more than the company that assembles the chassis.
Tesla is the most commercially credible entrant by a significant margin, for reasons that have little to do with robot technology and everything to do with manufacturing infrastructure. Elon Musk announced in Q4 2025 that more than 1,000 Optimus units were deployed within Tesla's own factories, primarily performing battery cell sorting and parts manipulation at the Fremont, California and Gigafactory Texas facilities. These are structured factory environments — not the unstructured problem of a residential plumbing job — but the deployment represents the first time a humanoid robot has operated at commercially meaningful scale outside a laboratory.
The economics are what make Tesla's position distinctive. Musk has stated a target production cost of $20,000–$30,000 per unit at scale, with a target of one million units per year by 2030. Whether those specific figures are achieved precisely is less important than the directional reality: Tesla has the stamping dies, the battery supply chain, the electric motor expertise, the automation engineering capability, and the manufacturing floor experience to produce humanoid robots at a price point that no pure-play robotics start-up can match. A company that has never manufactured anything at volume is structurally unable to compete with a company that builds 1.7 million cars per year on the cost curve for physical hardware.
The addressable market at $20,000 per unit with ten million units is $200 billion — comparable to Tesla's current automotive revenue run rate. At one million units per year, the revenue contribution begins to rival the car business within this decade. Optimus is not a side project or a demonstration piece. It is a product line with a larger long-run addressable market than the vehicle business that funded its development.
Figure AI raised $675 million in early 2024 from a consortium that includes Microsoft, OpenAI, Nvidia, Jeff Bezos's Bezos Expeditions, and ARK Invest. The funding round valued Figure at approximately $2.6 billion pre-money. The Figure 02, announced in 2024, features dexterous hands with finger-level actuation capable of fine manipulation tasks. The BMW partnership for deployment in BMW's South Carolina manufacturing plant represents the first third-party commercial deployment of a Figure robot, performing material handling, part feeding, and quality inspection tasks.
Figure's architectural bet is significant: it partnered with OpenAI specifically to integrate large-scale language model reasoning into robot action planning, rather than building a proprietary AI stack. This means Figure's robots reason about tasks using the same model family that powers ChatGPT — an AI trained on human language, description of physical tasks, and (via multimodal extensions) visual understanding of physical scenes. The hypothesis is that a model with broad world knowledge will generalise to novel trade tasks better than a model trained exclusively on robot sensor data.
Physical Intelligence is the entrant most focused on the software layer, and arguably the most important company in the sector for long-run trade displacement. Founded by Sergey Levine (Berkeley), Chelsea Finn (Stanford), and others from the academic robotics community, it raised $400 million in late 2024 with participation from Google DeepMind. The company's research focus is the generalised manipulation policy: a single AI model that can learn new physical tasks from video demonstration and generalise that learning across different robot hardware and different task variants.
This is the critical piece. A robot that can only perform tasks it was specifically trained on is commercially limited. A robot running a generalised policy model can, in principle, learn to perform a new task — say, replacing a specific type of isolation valve in a specific boiler model — by watching a video of a human doing it. The learning cycle that previously required months of reinforcement training compresses to hours or days. Physical Intelligence's π0 model, demonstrated in 2024, showed a single policy running across multiple robot body types and performing multiple dexterous tasks — folding laundry, assembling objects, operating kitchen appliances — from the same underlying model weights.
If Physical Intelligence succeeds in building a genuinely generalised manipulation policy, the analogy to GPT-3's impact on language tasks is appropriate. GPT-3 did not just improve one language application; it generalised across all language applications simultaneously. A generalised physical policy does not just improve one robot task; it generalises across all physical manipulation tasks simultaneously. That is the magnitude of the leverage.
1X Technologies, backed by OpenAI with $23.5 million in its OpenAI-led round (total funding approximately $100M by 2025), is developing NEO, a humanoid designed for commercial environments rather than pure factory applications. The company's earlier EVE platform was wheeled; NEO is bipedal, suggesting a product roadmap toward the kind of unstructured navigation required in residential and commercial buildings.
Apptronik, a spin-out from the Human Centered Robotics Lab at the University of Texas at Austin and with NASA heritage from work on Valkyrie, is developing the Apollo humanoid. The GXO Logistics partnership for warehouse deployment places Apptronik in the same logistics automation space as Agility Robotics but with a more humanoid form factor designed to transition to general physical tasks.
Boston Dynamics released the fully electric Atlas in April 2024 — a commercial pivot from the hydraulic research platform that had defined the company's public identity since 2013. Hyundai, which acquired Boston Dynamics from SoftBank in 2021, provides the manufacturing and distribution infrastructure that Atlas lacked under its previous owners. The electric Atlas is designed for factory deployment, with Boston Dynamics citing automotive and logistics as initial markets — but the platform's physical capability envelope exceeds factory requirements, and the Hyundai connection brings both capital and an industrial customer base.
Agility Robotics' Digit, backed by Amazon (which led the 2023 Series B), is bipedal but not fully humanoid — it lacks humanoid arms and uses a distinctive reverse-jointed leg configuration. Amazon has deployed Digit units in its fulfilment centre network for tote-moving tasks. Digit is less relevant to physical trade displacement than the humanoid platforms but demonstrates commercial-scale deployment of bipedal robots in real logistics environments — an important proof point for the industry.
| Company | Robot | Funding / Owner | Target Cost | Deployment Status (2025) | Primary Target Market |
|---|---|---|---|---|---|
| Tesla | Optimus Gen 2 | Public (TSLA); internal capital | $20,000–$30,000 | 1,000+ units, own factories | Manufacturing, then general labour |
| Figure AI | Figure 02 | $675M raised; OpenAI, Nvidia, Microsoft, Bezos | ~$30,000 est. | BMW partnership, factory pilot | Automotive manufacturing, logistics |
| Physical Intelligence (π) | π0 policy model | $400M raised; Google DeepMind | Software layer (hardware-agnostic) | Research / pilot stage | Generalised manipulation AI layer |
| 1X Technologies | NEO (bipedal) | ~$100M raised; OpenAI-led | Not disclosed | Pre-commercial | Commercial environments, services |
| Apptronik | Apollo | ~$160M raised; NASA heritage | ~$30,000–$40,000 est. | GXO Logistics pilot | Logistics, manufacturing |
| Boston Dynamics (Hyundai) | Atlas (electric, 2024) | Hyundai-owned; Hyundai manufacturing | Not disclosed (commercial licensing) | Commercial launch, automotive focus | Automotive manufacturing, logistics |
| Agility Robotics (Amazon) | Digit | Amazon-backed; ~$150M raised | Not disclosed | Amazon fulfilment centres, live | Logistics, warehouse tote-moving |
Not all physical trades face the same displacement timeline. The relevant variables are: the degree of environmental standardisation (factory vs. residential); the fineness of manipulation required; the extent of regulatory and safety inspection requirements; and how much of the task portfolio is amenable to decomposition into learnable sub-tasks. Each of the major trade categories has a different profile on these dimensions.
Plumbing is arguably the most challenging physical trade to automate, and therefore the most instructive case for assessing the realistic timeline. The core tasks — pipe fitting, soldering and crimping connections, fixture installation, drain clearing, leak diagnosis — involve a combination of challenges that stress every dimension of current robotic capability.
The environmental challenge is significant. No two residential properties have identical plumbing layouts. In the United Kingdom, where a large fraction of the housing stock dates from before 1970, plumbing systems may incorporate lead waste pipes, copper supply lines with varying compression and solder joint types, CPVC hot water runs, and proprietary push-fit connectors, sometimes within the same property. The access constraints — under sinks, inside wall cavities, beneath raised floors with joists at irregular spacing — represent exactly the kind of confined, non-standard space that current robot platforms struggle with. Bipedal robots operate well in open spaces. The ability to manoeuvre in a forty-centimetre cupboard under a kitchen sink, apply lateral force to a wrench in a constrained overhead position, and withdraw safely without damaging surrounding surfaces is a materially harder problem.
What is changing: dexterous manipulation advances in the Figure 02 and Optimus platforms directly address the fine-motor component of pipe connection. Spatial mapping via LiDAR and depth cameras is now reliable enough to generate three-dimensional models of confined spaces from a robot's onboard sensors. AI fault diagnosis from water pressure sensors and acoustic emission data — identifying the location and cause of leaks before physical intervention — is already commercially available in IoT plumbing monitoring systems and reduces the discovery-phase work a robot would need to perform.
The displacement sequence follows a logic: standardised tasks in new-build construction (identical floor plan, consistent pipe routing, standard fixture types) will be automated before retrofit work in old housing stock. A robot fitting bathroom fixtures in a newly built housing development where every unit is identical is a categorically simpler problem than diagnosing and replacing a section of Victorian lead waste in a non-standard terrace. New-build installation work is structurally closer to the factory environment where the current generation of humanoids is already deploying.
Timeline estimate: New-build fixture installation and standard pipe replacement: 2029–2033. Retrofit and maintenance plumbing in existing housing stock: 2035–2042. Full trade displacement (including complex diagnostics): 2038–2045.
Electrical work presents a different profile. The manipulation challenge is, in some respects, more severe than plumbing: wire routing requires handling deformable objects (cables) that do not hold their shape during manipulation, and connector installation demands sub-millimetre precision in tight terminal blocks. Physical Intelligence's specific research focus on deformable object manipulation — the ability to handle flexible materials like cables, fabrics, and tubing — addresses this directly. It is not a coincidence that this is among the company's stated research priorities; flexible object handling is one of the last manipulation domains where human hands substantially outperform current robotic end-effectors.
The regulatory dimension adds a genuine constraint not present in plumbing at the same severity. In the UK, Part P of the Building Regulations requires that certain types of electrical work be certified by a competent person or notified to local authority building control. In the US, electrical work in occupied buildings requires licensed electricians and inspection by the Authority Having Jurisdiction. These regulations exist for legitimate safety reasons — incorrectly wired circuits kill people. Any robotic electrical system will face a certification and inspection regime that adds cost and timeline to deployment and will likely require human sign-off on completed work for longer than in other trades.
New construction wiring — first-fix cable runs in newly framed walls, second-fix socket and switch installation in empty properties — faces lower regulatory friction than maintenance and fault-finding work in occupied buildings. The new-build pathway is again the entry point.
Timeline estimate: New-build first and second fix: 2030–2035. Maintenance and fault-finding in existing properties: 2032–2038. Full certification-equivalent capability: 2036–2042.
Heating, ventilation, and air conditioning is, paradoxically, among the more amenable trade categories for early robotic displacement despite its apparent complexity. The reason is that a larger fraction of HVAC work involves components at a scale and standardisation level more tractable for current robot platforms. Duct sections are large, geometrically simple objects that can be grasped and positioned without sub-centimetre precision. Air handling units and heat pumps arrive as packaged modules with standardised connection interfaces. Installation work is partially following a modular assembly logic that is closer to factory assembly than to the bespoke, confined-space work of plumbing retrofit.
Fault diagnosis is already partially automated. IoT sensors embedded in modern HVAC systems provide refrigerant pressure readings, temperature differentials, motor current draws, and airflow rates to cloud-connected monitoring platforms. The diagnostic reasoning that previously required a skilled engineer to arrive on-site, connect gauges, and interpret readings is increasingly performed by AI from remote sensor data. The physical intervention required when a fault is identified — replacing a failed component, topping up refrigerant, clearing a blocked filter drain — is a smaller fraction of the billable work than it was a decade ago.
Refrigerant handling adds a regulatory dimension: F-gas regulations in the UK and EU (and EPA Section 608 certification in the US) require certified technicians to handle refrigerants. Like electrical safety, this creates a genuine regulatory barrier to fully autonomous robotic HVAC work. But the certification requirement applies to refrigerant handling specifically; non-refrigerant installation work (duct fitting, mechanical connections, electrical hook-up of an air handler) faces no equivalent restriction.
Timeline estimate: Duct installation and module positioning: 2028–2033. Commissioning and fault intervention: 2030–2035. Full end-to-end automation including refrigerant handling (pending regulatory adaptation): 2034–2040.
Construction is the trade category where automation is furthest advanced and where the evidence base for near-term displacement is strongest. Several systems are already commercial.
SAM100 (Semi-Automated Mason), developed by Construction Robotics, has been commercially deployed since 2015. SAM lays bricks at a rate of 300–500 per hour, compared to a skilled human mason's 300–500 per day under optimal conditions. SAM is not a humanoid; it is a purpose-built bricklaying machine that requires human set-up, mortar loading, and finishing work. But the productivity differential it demonstrates is illustrative of what purpose-built construction robotics can achieve when the task is sufficiently repetitive.
Hadrian X, developed by FBR (Fastbrick Robotics) in Australia, takes a different approach: a truck-mounted robotic arm that constructs a complete wall structure from a 3D CAD model, cutting and placing blocks in the optimal sequence determined by the software. FBR has moved to commercial deployment with select residential construction partners. Hadrian X builds a standard residential house frame in one to two days, compared to the multi-week timeline for human bricklaying crews.
Dusty Robotics has deployed FieldPrinter, an autonomous robot that prints construction layout markings on concrete slabs from architectural drawings with millimetre precision. This eliminates the manual layout work that precedes all above-slab construction work and generates a digital audit trail of as-built versus designed positions. Dusty is already in commercial use on large construction projects in the US.
Ceiling and drywall installation robots from companies including Hilti (Jaibot for ceiling drilling), Canvas (drywall finishing), and Tybot (rebar tying) address specific high-repetition tasks within the construction process. Each reduces human labour hours in its specific domain without replacing the full trade. The convergence of multiple purpose-built robotic systems covering different phases of the construction process is, collectively, displacing construction labour hours even before any general humanoid platform arrives on site.
Framing — the erection of structural timber or steel frames — is the most resistant construction task due to dimensional variability in materials and the complex coordination required between multiple workers. A human framing crew compensates continuously for lumber that is not perfectly straight, fastener patterns that must adapt to knots or splits, and structural decisions that require immediate judgment. This is the hardest construction task for current robotics, but it is not immune — it is merely last in the displacement sequence.
The global construction industry is estimated at $10–13 trillion annually. It is the largest single employing sector in most developed economies by physical labour intensity. Labour costs typically represent 30–50% of construction project value. The cost reduction available from robotic construction is therefore not a marginal efficiency gain; it is a structural repricing of the largest physical sector in the economy.
Timeline estimates: Bricklaying and masonry (purpose-built): already commercial. Layout, ceiling, drywall, rebar: 2025–2028. Structural framing: 2030–2037. Full humanoid-capable general construction work: 2033–2040.
Painting is further along the automation path than most people recognise. PaintJet, a robotic painting system acquired by Sherwin-Williams in 2021, automates exterior commercial painting. Interior residential painting remains harder due to masking, trim work, and navigating furnished spaces — but the surface-following locomotion problem for a painting robot is substantially easier than the manipulation problem for plumbing or electrical work. Floor laying (screed, tile, and wood flooring) has robotic systems at the commercial deployment stage. Window installation involves large, fragile panels requiring two-person coordination — amenable to a humanoid pair-working approach. Roofing involves the most dangerous working conditions of any trade, creating both a cost case for automation and a regulatory one.
| Trade | Key Automation Challenge | Hardest Sub-Task | Structured Work Timeline | Full Displacement Timeline |
|---|---|---|---|---|
| Plumbing | Confined access, deformable connections, non-standard layouts | Retrofit in old housing stock | 2029–2033 | 2038–2045 |
| Electrical | Deformable wire handling, regulatory certification | Fault-finding in occupied buildings | 2030–2035 | 2036–2042 |
| HVAC | Refrigerant regulations, diagnostic reasoning | Refrigerant handling certification | 2028–2033 | 2034–2040 |
| Construction (bricklaying) | Scale, material handling | Irregular materials | Already commercial (SAM, Hadrian X) | 2027–2032 |
| Construction (framing) | Dimensional variability, load-bearing judgment | Structural decision-making | 2030–2034 | 2033–2040 |
| Painting (commercial) | Surface following, masking | Interior residential trim work | Already commercial (exterior) | 2028–2033 |
| Roofing | Pitched surface navigation, safety | Flashing and complex joins | 2031–2036 | 2035–2042 |
The preceding sections describe a technological and investment story. This section describes something different: the human architecture that was built on the assumption that the trades would remain available. Most investment research does not engage with this dimension because it is not directly legible in a discounted cash flow model. It is, however, the dimension most likely to produce non-linear political and social consequences that do eventually feed back into investment environments in ways that are hard to predict and impossible to price.
The physical trades in developed economies represent a specific social contract that emerged in the mid-twentieth century and has been maintained through the subsequent waves of automation. The contract is as follows: a young person who does not pursue university education — whether by choice, circumstance, or aptitude — can enter an apprenticeship programme in a physical trade, develop a skilled qualification over three to five years, and achieve an income, social status, and degree of economic autonomy broadly comparable to a graduate entering a mid-tier professional career.
In the United Kingdom, a qualified plumber or electrician earns £35,000–£60,000 per year, with self-employed tradespeople in London and the South East regularly earning £70,000–£90,000 once overhead is accounted for. These figures are not outliers; they represent the normal earning range for a competent, experienced tradesperson. Construction site managers and HVAC engineers with ten years' experience frequently earn above £60,000 in full-time employment. These are solidly middle-income positions — not elite, but stable, respected, and providing real economic security.
The social respect dimension matters as much as the income. Physical skill that produces visible results — a system that works, a structure that stands, a fault that is diagnosed and fixed — carries a form of dignity that is qualitatively different from the kind available in service work. A plumber who has completed a difficult repair in a confined space has achieved something concrete. The autonomy of self-employment, the social utility of maintaining the infrastructure of daily life, the physical fitness that comes with active work — these are not small things in the architecture of a life. They are part of why the trades were not merely tolerated as an alternative to university but positively chosen by a significant proportion of the population who could have pursued other paths.
The employment figures are significant. In the United Kingdom, approximately 3 million people are employed in construction-related trades. The plumbing and heating sector alone employs approximately 120,000 people, with a further 60,000 self-employed plumbers. The electrical contracting sector employs approximately 250,000. HVAC engineering employs a further 100,000–150,000. These figures do not include ancillary employment in materials supply, tool distribution, and trade training — sectors that are economically dependent on the trades being active.
In the United States, the Bureau of Labor Statistics reports approximately 7 million people employed in construction trades. Plumbers, pipefitters, and steamfitters number approximately 500,000. Electricians number approximately 700,000. HVAC technicians number approximately 375,000. Construction labourers and helpers number over 1.5 million. In total, the physical trades represent one of the largest single employment categories in the US economy, concentrated among workers without four-year college degrees.
The demographic profile of this workforce is important. Construction and trades employment is disproportionately male — approximately 89% in US construction according to BLS data. It is disproportionately working-class by origin, without university qualifications, and concentrated in regions where the manufacturing base that previously provided comparable working-class employment has already been substantially eroded. In the United Kingdom, this describes exactly the geography of the post-industrial Midlands and North; in the United States, it describes the Rust Belt, Appalachia, and parts of the South.
Each previous automation wave displaced one category of working-class employment and, eventually, created or expanded another. Factory automation in the 1970s and 1980s destroyed manufacturing employment in developed economies at scale. The displaced workers, or more precisely their children, moved into service employment — retail, hospitality, logistics — and into the expanding public sector. The trades were an important part of this refuge: construction boomed in the 1990s and 2000s, providing employment for workers exiting manufacturing.
The software and digital economy wave of the 1990s and 2000s created a new professional class but largely left the trades untouched. A factory worker who could not code could still become a plumber or an electrician. The trades were not a fallback that people stumbled into reluctantly; they were a respected alternative path with real economic outcomes. The existence of that path was not incidental to social stability. It was load-bearing.
The 2010s automation wave — e-commerce logistics, automated warehousing, self-checkout — began to compress the service sector refuge. Retail employment peaked and began declining. Warehouse work became more automated. But during this period, the construction and trades sector expanded, partly filling the gap created by service sector compression. The UK construction sector grew through most of the 2010s. US construction employment recovered strongly post-2010 and reached multi-decade highs by 2019.
The sequence is therefore: manufacturing jobs (1970s–1990s) → service jobs (1990s–2010s) → trades jobs (2000s–present). Each category was the refuge of the previous displacement. The question this report is raising is not rhetorical: what comes after trades?
The honest answer, which no politician in any major developed economy has yet stated plainly, is that there is no next refuge in the same sense. The displacement of knowledge work (covered in The Headless Enterprise) and the displacement of physical work (this report) are occurring simultaneously rather than sequentially. There is no category of mass employment, accessible without university qualifications, paying a family-sustaining wage, that is clearly immune from the current automation wave.
The domains of human work that remain genuinely resistant to automation share a common characteristic: they require human social presence as the primary input, not a workaround. Care work — nursing, home care, childminding — requires human emotional responsiveness and physical proximity in a way that reflects a social preference, not a technical limitation. People want to be cared for by humans. But care work pays poorly by comparison with trades, is predominantly female, is undervalued by markets, and is fiscally dependent on public sector budgets that are already under pressure from aging populations. It is not an equivalent refuge.
The political economy consequence of closing the last large-scale working-class employment refuge is not difficult to trace. The Brexit referendum of 2016 and the Trump elections of 2016 and 2024 were driven by a coalition that political scientists have characterised in various ways: left-behind communities, non-college-educated white voters, post-industrial regions, economic losers from globalisation. Regardless of the framing, the underlying economic reality is consistent: a significant fraction of the working-class population of developed democracies experienced real wage stagnation or decline through the automation and globalisation waves of the preceding three decades, and their political expression of that experience was to vote for disruptive candidates and causes.
The trades, during this period, were an economic stabiliser. They provided a category of working-class employment that was genuinely not threatened by globalisation (you cannot outsource residential plumbing to a lower-wage country), paid well, and maintained the dignity and social standing of physical skill. Tradespeople were not among the economically distressed. They voted in various directions and for various reasons; many were economically comfortable by the standards of their communities.
Full trade automation, on the current trajectory, removes that stabiliser. The displacement will not be instantaneous — the timelines in Section III extend to 2040 and beyond for full displacement — but it will be visible and directional well before it is complete. An apprentice plumber starting a four-year training programme in 2027 will enter a market that is visibly contracting by 2033. The signal is not ambiguous: training pipelines will shrink, wages will compress as robot deployment grows, and the social contract that made trades the respectable alternative path will dissolve.
“Universal Basic Income becomes not a fringe proposal but a structural necessity when the last mass-employment category for non-university-educated workers is automated. The question is no longer whether we will need it. The question is whether we will have the political and fiscal capacity to implement it before the displacement creates conditions in which rational policy-making becomes very difficult.”
The political implications are non-linear in the precise way that economists use that term: small changes in conditions can produce large changes in outcomes. A population that has experienced fifty years of repeated reassurance that their jobs are safe, and has watched each reassurance expire in sequence, and is now told that the final category — the thing that required physical skill, local presence, and could not be offshored — is also being automated, will not respond with quiet acceptance. The political consequences of that moment, which is now a decade or less away in early-displaced sub-sectors, are not predictable from the current political baseline. They will be shaped by whether government policy responds with plausible economic alternatives before displacement is visible at scale, or after.
The investment landscape created by physical AI and humanoid robotics has both long and short sides. The long side is concentrated and analytically tractable. The short side is diffuse and requires careful positioning because the displacement timeline extends over fifteen years, creating long periods where trades businesses remain fully functional and the short thesis has significant time decay.
Tesla's position in humanoid robotics is the most compelling investment case in the sector for a straightforward reason: manufacturing scale. Every other humanoid robotics company is a start-up attempting to build a novel hardware product from scratch. Tesla is a company that has already built the supply chain, tooling, manufacturing processes, and quality management infrastructure required to produce complex electromechanical products at volume. Optimus shares actuators, battery technology, power electronics, and manufacturing processes with Tesla's automotive line. The marginal cost of adding humanoid robot production to Tesla's existing infrastructure is categorically lower than the total cost any pure-play robotics start-up faces.
The addressable market at $20,000 per unit is transformatively large. Ten million units — a plausible medium-term installed base if cost targets are met — represents $200 billion in revenue before any recurring software, service, or AI subscription layer. If Optimus achieves one million units per year by 2032, the revenue contribution rivals the automotive business that built Tesla's market capitalisation. The current Tesla valuation, which remains primarily a function of automotive revenue and the energy business, does not fully price the Optimus upside under a scenario where cost targets are achieved and deployment scales into commercial markets including, eventually, physical trades.
The risk to this thesis is execution: Tesla has a consistent history of missing production timelines, and Musk's specific numerical claims about Optimus production targets should be treated as aspirational rather than committed guidance. The 1,000-unit internal deployment figure is real; the one-million-unit-per-year-by-2030 target is almost certainly optimistic. But the direction is correct, and the manufacturing advantage is structural rather than contingent on any specific timeline.
Nvidia's position in humanoid robotics is the same position it holds in AI generally: the company that provides the computing substrate on which every other company's products are built. The Isaac platform — Nvidia's robotics simulation and training environment — is used by Figure, Agility, and other humanoid companies to train robot policies in simulation before physical deployment. Jetson, Nvidia's edge computing module, is the inference platform embedded in robot systems that require onboard AI processing rather than cloud inference. CUDA underpins every training run for every manipulation policy model.
Nvidia does not need to pick a winner in the humanoid hardware competition. It supplies the training and inference layer to all of them. The growth of humanoid robotics as an industry adds an entirely new demand vector for Nvidia compute — one that is structurally different from data centre AI inference because it extends Nvidia's revenue into the physical manufacturing economy rather than remaining concentrated in cloud providers. The scale of the physical economy is an order of magnitude larger than the cloud AI market that has driven Nvidia's growth since 2023.
Physical Intelligence is not yet publicly traded, and its current valuation reflects the private funding round rather than a liquid market. But the strategic importance of what the company is attempting — a generalised manipulation policy that serves as the AI software layer across all robot hardware — is directly analogous to what OpenAI built for language tasks. If Physical Intelligence succeeds, the company that owns the generalised physical manipulation policy owns the most valuable chokepoint in the entire humanoid robotics value chain. Every robot hardware company becomes a customer. Every trade-automation application runs on the policy layer.
This is a high-risk, high-upside private investment thesis. The company may be acquired before any liquidity event by Alphabet, Amazon, Apple, or a large industrial conglomerate seeking to control the software layer. An acquisition at $5–15 billion would be consistent with comparable AI software acquisitions and would represent a multiple of the 2024 funding round valuation.
All three companies are private, pre-revenue at scale, and operating in a sector where the competitive advantage of Tesla's manufacturing infrastructure is a structural headwind to stand-alone hardware success. The most likely outcome for each is acquisition rather than independent public markets success. Potential acquirers include Amazon (which already owns Agility Robotics and has the logistics deployment infrastructure to absorb another humanoid platform), Alphabet (which has Google DeepMind participation in Physical Intelligence and a persistent interest in robotics since the abortive 2013–2017 robotics investment programme), Apple (which has the hardware manufacturing relationships and consumer electronics experience to apply to a humanoid form factor), and large industrial conglomerates including Siemens, ABB, and Rockwell Automation.
Figure's Microsoft and OpenAI investor base makes an acquisition by either company possible, though Microsoft's primary interest in AI is software and cloud, and OpenAI's robotics interest is expressed through 1X rather than Figure. The investor overlap creates both alignment and potential conflict in any acquisition scenario.
The short thesis in trades services is real but requires careful positioning on the timeline. ServiceMaster, Rollins (Orkin), and large HVAC/plumbing service chains are businesses whose core cost of service delivery is skilled human labour. Their competitive advantage — a managed network of trained technicians — is the thing that robotic deployment will erode. As humanoid robots reach commercial viability in structured trade tasks (2028–2033 on the optimistic timeline), the first adopters will be large-scale services companies with the capital and operational infrastructure to deploy robots at fleet scale. This compresses labour costs for early adopters and creates margin pressure for those who cannot or do not deploy.
The short thesis is not that these companies fail; it is that their labour-based competitive moat erodes and their margins compress as the cost of robotic service delivery falls below the cost of human service delivery in structured task categories. The timeline for this compression to appear in public company earnings is 2030–2035. A short position initiated today has significant time decay.
Public construction companies with high labour intensity ratios face the same dynamic: early adopters of robotic construction gain cost advantages that put pressure on competitors. The bricklaying and masonry category is already at this inflection point with SAM and Hadrian X in commercial deployment. Construction companies that adopt robotic systems earliest will have lower cost bases than competitors, enabling them to underbid on projects or take higher margins, creating a competitive dynamic that compounds over time.
The specific companies most exposed are general contractors with high direct labour ratios in the sub-contractor categories most susceptible to early automation: masonry, drywall, painting, and floor laying. Investors should monitor labour cost as a percentage of revenue as the leading indicator of exposure.
Trade apprenticeship businesses, staffing agencies specialised in construction and trades, and vocational training providers face a structural problem that precedes the actual displacement of employed tradespeople. As displacement becomes visible — as the apprenticeship pipeline begins to perceive declining career value — enrolments in trade training will compress. A training business whose revenue depends on a steady supply of people choosing plumbing or electrical apprenticeships will experience demand erosion well before the actual employment market for qualified tradespeople contracts. The signal in enrolment data will precede the signal in employment data by five to ten years.
The bear case for physical trade automation is more credible than the bear case for knowledge-work automation, and investors should hold it seriously.
The core bear argument is that dexterous manipulation in truly unstructured environments — old homes with non-standard layouts, properties with structural surprises, retrofits in occupied buildings with unpredictable constraints — is harder than factory environments by an order of magnitude, and the gap between factory-grade capability and residential-grade capability may not close on the timelines presented here. Current humanoid deployments are in structured factory settings where the environment is mapped, the task is repeatable, and the robot can be removed and reprogrammed when it encounters an exception. A residential plumbing job is the opposite of all of these things.
Regulatory resistance is a genuine constraint. Building regulations in the UK require Part P notification for electrical work; gas work under the Gas Safety Regulations requires Gas Safe registration. Analogous regulations in EU member states and US jurisdictions create legal barriers to autonomous robotic work in occupied buildings that are not simply technical problems to be solved — they are political decisions about acceptable risk that will be contested by trade unions, professional bodies, and insurers with significant influence over the regulatory process. In countries with strong trade union movements — France, Germany, and the Nordic countries in particular — the political resistance to robotic displacement of construction trades will be organised, sustained, and potentially effective at extending timelines significantly beyond the estimates above.
Public acceptance is an underappreciated constraint. Having a humanoid robot in your home performing work unsupervised is a step change from having a robot vacuum cleaner. Insurance liability for autonomous robotic work — if a robot causes a gas leak or an electrical fire, who is liable and under what framework? — is an unresolved legal question that will slow deployment in residential settings independent of technical capability.
The honest summary of the bear case is: add five to ten years to every timeline in Section III for unstructured residential environments, and add additional uncertainty for jurisdictions with strong regulatory and union resistance. The technology arrives; the deployment is gated by human, legal, and political factors that are slower-moving and less predictable than hardware cost curves.
There is a specific quality to the reassurance that was given to physical workers through every automation wave — a quality that distinguished it from the reassurances given to factory workers, to bank tellers, to travel agents, to any of the other categories that were displaced. The reassurance given to tradespeople was demonstrably technical rather than merely rhetorical. Robots genuinely could not do what a plumber did. The barrier was real. The exemption was earned.
What makes the current moment different from every previous moment is not the rhetoric but the technical evidence. Boston Dynamics' electric Atlas performing automotive assembly tasks in a Hyundai factory. Tesla's Optimus sorting battery cells on a Fremont production line. Figure 02's dexterous hands threading components in a BMW facility. Physical Intelligence's π0 model folding laundry and assembling objects from the same underlying policy weights. These are not demonstrations of what might one day be possible. They are commercial deployments of systems that are already working in structured environments, on a clear trajectory toward the unstructured environments where trades work occurs.
The timeline from “works in a factory” to “works in your home” is not short. It is a decade or more, and it is subject to all the regulatory and political friction described in the bear case. But it is a trajectory that is now visible, directional, and driven by capital and competitive dynamics that are not going to reverse.
For investors, the primary questions are: who owns the manufacturing scale (Tesla), who owns the compute layer (Nvidia), and who owns the generalised manipulation policy (Physical Intelligence, currently private). For everyone else — for the societies that built economic dignity on the assumption that physical skill would always be economically valued — the primary question is different, harder, and more urgent: what is the alternative, and when does it need to be ready?
The last exemption from automation was not a privilege. It was a technical fact. The technical fact is changing. The social architecture built on it will change more slowly, more painfully, and with less predictable political consequences than any cost curve or deployment forecast can capture.
Disclaimer: This report is produced by PRZC Research for informational and analytical purposes only. It does not constitute investment advice, financial advice, or any solicitation to buy or sell securities. All views are those of the analyst and are based on publicly available information. PRZC Research makes no representations as to the accuracy or completeness of information contained herein. Past performance is not indicative of future results. Readers should conduct their own due diligence before making any investment decisions.