The Industrial Automation & Robotics Stack: Where the Durable Profits Concentrate

The durable profit in this stack never reaches the machine that moves. It stops at a proprietary engineering estate — a controller, its protocol, and the tool that programs both — where the customer’s own past work becomes the switching cost, and at a handful of application-engineering chokepoints in sensing (the narrow points where making a technology work on a particular factory floor takes scarce human expertise, not better code) that no algorithm has yet dissolved. The market is paying for motion and being handed a commodity.

Every consensus has a piece of truth in it, and this one has a large piece. Factories do need automating. Labour is scarce, reshoring is real, and perception models can now do things in 2026 that no vision system could do in 2019. None of that is in dispute here. The dispute is narrower and more useful: demand for a thing and profit from making it are different questions, and in this industry they have been pointing in opposite directions for several years.

Three facts, all from the companies’ own disclosures within the last eighteen months. Teradyne’s Robotics segment — Universal Robots and Mobile Industrial Robots, the franchise that invented the collaborative-robot category (arms built to work alongside people without a safety cage) and has shipped more than 110,000 arms since 2008 — generated $308.3 million of revenue in 2025 and a pre-tax loss of $99.4 million, on segment assets of $733.5 million. Revenue fell 15.5% (Teradyne 10-K for FY2025, segment note and MD&A). In the same year, ABB agreed (8 October 2025) to sell its entire Robotics division to SoftBank for an enterprise value of $5.375 billion (the price of the whole business, debt included, rather than the price of its shares alone) — EU merger clearance obtained 12 March 2026; as of the 2 August 2026 evidence cutoff the transaction had not closed, remaining subject to China/US and customary conditions, having concluded that “there are limited business and technology synergies between the ABB Robotics business and the remainder of ABB’s businesses” (ABB ad hoc release, 8 October 2025). And Rockwell Automation raised prices by roughly three percentage points in a fiscal year when unit volume fell two points, lifting its Software & Control segment margin to 29.7% (Rockwell 10-K for FY2025, MD&A).

Three companies, three layers, one pattern. Where the customer can swap the box, the seller takes whatever price the market offers. Where the customer cannot swap the box without re-engineering a plant, the seller sets the price.

That second condition is what economists call a rent, and the word is worth pinning down because the rest of this brief runs on it. A rent is profit that keeps coming because something stops a competitor from copying the thing — not profit earned by being temporarily cheaper, faster, or first. Being in a growing market earns you revenue. Being impossible to replace earns you rent. Those are separate achievements, and this industry has been handing them to separate companies.

01How a factory is automated — and where the money changes hands

Strip the marketing away and an automated production cell is a small, stubborn hierarchy. At the bottom sit the things that move and the things that see: motors, gearboxes, linear actuators, grippers, cameras, encoders, proximity sensors. Above them sit drives — power electronics that translate a commanded velocity into current. Above the drives sits a controller, almost always a programmable logic controller (a PLC — an industrial computer that executes a deterministic control program on a fixed scan cycle) or its process-industry cousin, a distributed control systemA DCS. Where a PLC runs one machine or cell on a fast, rigid loop, a distributed control system co-ordinates a whole continuous process — a refinery, a chemical train, a paper machine — through controllers spread across the plant and tied to one operator console. Refineries and pharmaceutical plants run on these, and they are replaced roughly as often as the plants themselves.. Everything is wired together by an industrial network. Above the controller sits the supervisory layer: SCADA (supervisory control and data acquisition, the screens and historiansA historian is the plant’s memory: a database that records every sensor reading, valve position and alarm, timestamped, for years. Engineers query it to work out why a line stopped last Tuesday, and regulators read it to confirm a batch was made the way it was supposed to be. Because it is written in the vendor’s format against the vendor’s tag names, it is also one of the things a customer loses on a platform change. an operator watches) and MES (manufacturing execution systems, which schedule and track production orders). Somewhere off to the side sits the design estate of CAD, simulation, and increasingly a digital twinA running software model of a real machine or production line, fed by the same control logic and often the same live data. Engineers use it to test a change — new product, new sequence, new robot — before touching the physical line, which is the only way to make a change to a factory that cannot afford to stop. The twin is built inside the vendor’s toolchain, so it deepens the same attachment the engineering software already creates. of the line itself.

The robot arm, in this picture, is not the top of the hierarchy. It is a peripheral. It has its own controller, and that controller takes its orders from the plant PLC over the plant network. That wiring decides who is replaceable.

Follow a dollar of automation capex through that stack and the accounting is unflattering to the machine. In North America in 2025, buyers ordered 36,766 industrial robots worth $2.25 billion, a 6.6% rise in units and 10.1% in value and the strongest year since 2022, on data compiled by the Association for Advancing Automation and reported on 7 February 2026 (A3 2025 order statistics, as reported). That is an average of about $61,200 per robot at the OEMOriginal equipment manufacturer — the company that builds and badges the machine itself, as against the integrator who installs it or the end-user who runs it. FANUC, Yaskawa, ABB and Universal Robots are the robot OEMs in this brief. The OEM’s price is the factory-gate price, before anyone has made the thing work in a specific plant.’s door. Collaborative robots — 7,212 units for $241 million — averaged about $33,400. Now set the whole of it against a single segment of a single American company: Rockwell’s Software & Control segment billed $2,383 million in fiscal 2025 (Rockwell 10-K FY2025).

The comparison is unfair in one direction. The A3 number is order value for robot units alone; a delivered cell costs a multiple of the arm, and most of that multiple goes to integrators. That multiple is the point. The arm is the least valuable part of the thing it gives its name to. What surrounds it is either billed as labour, which earns no rent, or embedded in a controller a competitor cannot displace, which earns most of it.

Where the money actually changes hands

Four transactions matter, and they behave nothing alike.

1. 01
   Specification

   The end-user’s corporate engineering standard names the PLC brand and the network. It is written once and lives for decades. No money changes hands here; every later dollar is decided here.
2. 02
   Component purchase

   Arms, drives, motors, cameras. Competitive, quoted, re-quoted, and increasingly quoted against a Chinese alternative. The buyer sets the price.
3. 03
   Integration & commissioning

   Engineering hours, safety certification, ramp support. The largest single line on most projects and the one with the least operating leverageThe degree to which extra sales drop through to profit rather than being eaten by extra cost. A software licence has enormous operating leverage: the tenth copy costs nothing to make. Billable engineering has almost none, because doubling the work means hiring twice the engineers — which is why a labour business can grow for years without ever getting more profitable..
4. 04
   Twenty-year aftermath

   Spares, migrations, engineering-tool licences, network hardware, service contracts. Sold at the incumbent’s price because the alternative is re-writing the plant’s control logic.

Step four is the entire game. It never appears in a robot-installation statistic. The industry’s most-quoted number — annual installations — measures step two, the one node where nobody has pricing power.

02The protocol franchise: how a decades-long lock was built, and what threatens it

A plant cannot casually change PLC vendors, and the hardware is the least of the reason. Three things are welded together: the controller, the engineering software that programs it, and the network protocol that carries its I/OInput/output — the traffic between a controller and the physical world. Every sensor reading in and every command out is a piece of I/O, and a mid-sized plant has tens of thousands of them, each with a name, an address and a meaning recorded somewhere.. Change one and you touch the other two, and behind all three sits a decade or more of the customer’s own engineering.

The middle item deserves a moment, because the argument in this brief turns on it. The engineering software — Siemens calls its own TIA PortalSiemens’ engineering environment for its SIMATIC controllers — the single application in which a customer’s engineers configure the hardware, write the control logic, build the operator screens and keep the project files. A plant that has standardised on it has years of its own work stored in Siemens’ format, which is the reason the tool matters more than the controller it programs., Rockwell calls its Studio 5000Rockwell’s equivalent: the design and programming environment for its Logix controllers, holding the control routines, the tag database and the project history for a plant built on Rockwell equipment. — is where a plant’s engineers write, test, version and archive the control programs that make the machinery do what it does. It is not a design aid the customer could take or leave; it is the filing cabinet that holds the plant’s accumulated engineering, in a format only its own vendor reads. Buying a controller is a purchase. Adopting the tool that programs it is a tenancy, and the rent is paid in the work that piles up inside.

The network layer gets counted, which makes it the most legible evidence here. In 2025, industrial EthernetThe same physical Ethernet cabling and chips used in offices, carrying protocols rewritten so that a message arrives within a guaranteed time rather than merely eventually. A machine that must fire an actuator every few milliseconds cannot tolerate ordinary network jitter, which is why industrial Ethernet is a family of specialised dialects rather than one open network. accounted for 79% of newly installed industrial network nodesA node is one connected device on the plant network — a drive, a sensor, a valve island, an I/O rack, a controller. Node counts are the industry’s census: they measure how much equipment now speaks a given protocol, and therefore how much of a plant would have to be re-commissioned if it ever spoke a different one., up from 76% the year before, with legacy fieldbusThe generation before industrial Ethernet: serial wiring standards such as PROFIBUS or DeviceNet that connected field devices to a controller from the 1990s on. Still running in millions of installations, still supported, and no longer where new plants are built. down to 14% and shrinking 11% year on year. Within industrial Ethernet, three protocols took three-quarters of all new nodes: PROFINETThe industrial-Ethernet protocol built around Siemens’ technology and governed by the PROFIBUS & PROFINET International organisation. It is the dominant dialect in Europe and the default in any plant standardised on Siemens controllers. at 30%, EtherNet/IPThe rival dialect governed by ODVA, the consortium in which Rockwell is a principal member. It is the North American default, and it is the network estate a plant built on Rockwell equipment runs on. The “IP” is Industrial Protocol, not internet protocol. at 25% and EtherCATA third dialect, originated by the German controls maker Beckhoff and prized where motion has to be synchronised very precisely — multi-axis machines, high-speed packaging, robotics. Fast, technically admired, and still governed by a body its originator convened. at 20%, on HMS Networks’ annual survey published 18 June 2026 (HMS Networks industrial network market shares, 2025 data).

Read that as a market-structure fact and it looks like healthy standards adoption. Read it as an ownership fact and it looks like something else. PROFINET is governed by PROFIBUS & PROFINET International, an organisation founded around Siemens’ technology. EtherNet/IP is governed by ODVA, whose principal members are Omron, HARTING, Schneider Electric, Rockwell Automation, Endress+Hauser and Honeywell (ODVA, EtherNet/IP key technology page), and which describes EtherNet/IP as “the most developed, proven and complete industrial Ethernet network solution available for manufacturing automation” (HMS Networks annual industrial network market shares 2026 (2025 new nodes; EtherNet/IP 25%)). EtherCAT originates with Beckhoff. Three-quarters of the world’s new industrial network nodes speak a language whose grammar is set by a consortium that an incumbent convened.

The estate is counted every year. PROFIBUS & PROFINET International reported on 17 April 2026 that the PROFINET installed base had reached 89.2 million nodes, with 10.4 million added during 2025, against 78.8 million a year earlier; PROFIBUS stood at 71 million total with only 1 million new nodes, a business in managed decline; IO-LinkA short-range standard for talking to the smallest devices — individual sensors, indicator lights, simple actuators — over the ordinary wire that already powers them. Its value is that a device can report its own identity and settings, so a replacement configures itself instead of needing an engineer. It sits one level below the plant network and feeds into it. reached 71 million with 9.7 million new devices and master ports; and PROFIsafe — the functional-safety profile — reached 31.5 million nodes with 2.8 million added (PI press release, 17 April 2026; prior-year comparatives from PROFINEWS 2024 annual node count, 8 April 2025). The EtherCAT Technology Group, for its part, describes EtherCAT as “originally developed by Beckhoff Automation,” standardised in IEC 61158 and 61784, and requires conformance testing through its own test tool and accredited test centres before a product may reach market (EtherCAT Technology Group, technology overview).

Functional safety is worth understanding properly, because it is the part of this stack that argues back hardest against every disruption story told about it. A machine that can crush, burn or amputate has to be built so that the failure of any single component still leaves it in a safe state, and proving that is not a matter of writing careful code: an engineer performs a formal hazard analysis, specifies the protective functions, and has the result certified by an independent body against a published standard. The certificate attaches to a named combination of hardware, software and wiring. Change any part of it and the analysis has to be redone, re-certified, and personally signed by somebody who is liable if it is wrong.

The PROFIsafe number is that switching cost made countable: 31.5 million safety-rated nodes, each one attached to a hazard analysis and a certification that a vendor change would invalidate. And every one of these “open” standards operates a conformance gate — a body that decides whether your device may carry the mark. A standard with a gatekeeper is a franchise with better public relations.

“Open standard” in industrial automation has never meant vendor-neutral in the way it does in enterprise IT. It has meant published, conformance-tested, and multi-vendor at the device end — which lowers the cost of building a sensor that talks to a Rockwell PLC, and thereby makes the Rockwell PLC more valuable, not less. The standard is the moat’s masonry, not its breach.

The genuine threat, and its actual clock speed

The OPC Foundation’s Field Level Communications initiative is extending OPC UAOpen Platform Communications Unified Architecture — a vendor-neutral standard for describing what a piece of industrial equipment is and what data it holds, so that any system can read it without a custom driver. It won the layer above the controller: the link from the plant floor up to scheduling, analytics and corporate IT. Extending it down to the controllers themselves is the initiative described here, and it is a different and much harder problem. — already the lingua franca above the controller — down to the field level, with the explicit aim of controller-to-controller and controller-to-device interoperability across brands. That is a serious, well-resourced attempt to break the franchise, and the major vendors are inside it.

The scope matches the intent: its steering committee comprises 23 member companies funding the work, and its technical groups cover safety communication over a black channelA design trick for safety-critical messages: rather than certify the whole network as safe, the safety information is wrapped in its own checks and sequence numbers and sent across an ordinary network that is treated as untrusted — a “black” channel nobody has to vouch for. If a message is corrupted, delayed or lost, the receiver detects it and the machine goes to a safe state. for both controller-to-controller and controller-to-device links, plus motion devices including servo drivesThe electronics that make a motor hold an exact position or speed, correcting hundreds or thousands of times a second against feedback from the motor itself. Every robot joint and every precision machine axis has one behind it. They are components, bought to a datasheet, and they are the “drives” referred to throughout this brief. and encoders (OPC Foundation, Field Level Communications initiative). If it succeeds it does not merely add a protocol; it dissolves the reason a plant standardises on one vendor.

It is also slow, and its own progress reports say so. As of the March 2026 status update, the initiative had reached Release Candidate V1.00.04 of the UAFX specification series — still described as “another important milestone toward completing the UAFX 1.00 specification series” — with a multi-vendor controller-to-device demonstration scheduled to debut at the SPS trade fair in Nuremberg in November 2026 (OPC Foundation, Field Level Communications Corner, March 2026). Specifications that are still being completed in 2026 do not displace installed control logic before the 2030s. An investor should treat OPC UA FX as a genuine, dated, decade-scale risk to the controls franchise and as no part of the next five years’ earnings.

The second threat is one most commentary reads backwards: the controller itself is being virtualised. Siemens now sells SIMATIC S7-1500V, a software PLC deployed entirely through its Industrial Edge platform — “a hardware-independent controller” that lets a customer “adjust the number of instances used according to demand” (Siemens, SIMATIC S7-1500V product page). Commentators file this under “hardware becomes a commodity, incumbents disrupted.”

Look at what Siemens actually requires. The virtual controller “is completely compatible with TIA Portal, allowing you to build on projects that are already in the tool.” The box is being dissolved; the tool is not. Siemens is doing to its own hardware exactly what a rational rent-holder should do — commoditising the part it does not own the customer through, while re-anchoring the customer to the part it does. If control logic can be spun up as instances on standard compute, the vendor stops selling steel at a cyclical margin and starts selling licensed instances against an engineering estate. That is not disruption of the incumbent. That is the incumbent upgrading its own toll booth.

03The industrial-software layer: fat margins, and the bill that comes with them

Software is where the automation industry has spent its money for a decade, and where the investment case is most often argued from the wrong line of the income statement. A gross margin of 80% tells you what it costs to serve the next customer. It tells you nothing about what it cost to acquire the first one.

One piece of accounting has to be understood before the numbers below mean anything. When a company buys a software business, the price above the value of the identifiable assets is split into goodwill and acquired intangibles — the customer relationships, the technology, the brand it has just paid for. The intangibles are then written off against profit a little each year, which is what “amortisation of acquired intangibles” means: the purchase price arriving in the income statement in instalments, years after the cheque cleared. Almost every acquisitive industrial asks investors to look past that charge, and calls the result adjusted. Whether that is fair is the single question this section turns on, and the answer depends entirely on whether the company intends to keep buying.

Three disclosures, all opened this session.

Emerson. In fiscal 2025 Emerson’s Software and Control group, which holds Control Systems & Software with a now wholly owned AspenTech alongside Test & Measurement, essentially the former National Instruments, posted $5,691 million of sales and adjusted EBITAEarnings before interest, tax and amortisation. It is operating profit measured before the cost of debt, before tax, and before the write-off of intangibles bought in past acquisitions. “Adjusted” on the front means the company has also removed items it considers one-off. Each exclusion is defensible on its own; the question is always what is left out and who benefits from leaving it out. of $1,763 million, a 31.0% margin. On a GAAP basis the same group earned $827 million: a 14.5% margin. The difference is not a rounding convention. Amortisation of acquired intangibles in that group ran $893 million for the year — $468 million at Control Systems & Software and $425 million at Test & Measurement (Emerson 10-K FY2025, business-segment analysis). Test & Measurement, acquired at an equity value of $8.2 billion in October 2023, lost $68 million on a GAAP basis in fiscal 2025 on sales of $1,486 million.

Siemens. Siemens’ own segment table carries the number. Digital Industries, the world’s largest factory-automation-plus-industrial-software business, reported fiscal 2025 revenue of €17,788 million, down 4%, and profit of €2,643 million, down from €3,498 million; margin fell from 18.9% to 14.9%. Over the same year, Digital Industries’ net capital employedSiemens’ own measure of the money tied up in a division — its assets, including the goodwill from acquisitions, less the liabilities that fund it without cost. It is the denominator of a return: the capital the division has to justify. When it doubles while profit falls, the division has become a worse business by the only test that matters, whatever happened to revenue. rose from €10,476 million to €22,885 million (Siemens Earnings Release Q4 FY2025, segment overview), the consequence of the Altair and Dotmatics acquisitions. Profit divided by net capital employed — Siemens’ own two disclosed figures, on Siemens’ own definition — fell from 33.4% to 11.5%.

Set beside it the segment Siemens does not talk about at technology days. Smart Infrastructure earned €4,506 million on net capital employed of €6,709 million: 67.2%, up from 55.7%. Switchgear and building electrification, on the same page of the same disclosure, out-earning the industrial-software franchise six times over on the capital each consumes.

This is not a NOPATNet operating profit after tax — what the business earns from operations, taxed, but before any of it is paid out to lenders. It is the numerator of return on invested capital, and it is used precisely because it does not care how the business is financed: two identical factories should show the same NOPAT whether one is funded by debt and the other by equity.-over-invested-capital return, and it must not be laid beside the ROIC figures computed later in this brief for the US-listed names — different numerator (segment pre-tax profit vs. tax-effected EBIT), different denominator (Siemens’ net capital employed vs. debt-plus-equity-less-cash), different accounting standard. It is an internally consistent, disclosed, year-on-year comparison of one company against itself, and against its own sister segment. Within those limits it is devastating.

Dassault Systèmes. For fiscal 2025 the company reported a non-IFRS operating margin of 32.0% and an IFRS operating margin of 21.7% (Dassault Systèmes FY2025 results release, February 2026). Recurring revenue grew 6% on subscription growth of 11%; total revenue grew 4%. The company will begin reporting Annual Run RateARR — the annualised value of the subscription contracts a software company holds right now, as against the revenue it happened to book last year. For a business selling multi-year licences, ARR is the honest measure of the installed base, because reported revenue can rise or fall on the timing of renewals without anything real changing underneath. in 2026 — a disclosure improvement that is also, quietly, an admission that revenue no longer describes the business well.

What the software layer actually gets right

None of this makes the layer a bad business. It makes it a business whose quality has to be judged on the capital, and on that test it splits cleanly in two.

The design estate — CAD, PLM, simulation — is defended and capital-light when it is grown rather than bought. PTC’s fiscal 2025 shows what that looks like unencumbered: ARR of $2,478.5 million, total recurring revenue of $2,600.5 million on total revenue of $2,739.2 million (94.9% recurring), operating margin of 36%, free cash flow of $856.7 million, and capital expenditure of $11.0 million — four-tenths of one percent of revenue (PTC 10-K FY2025, results of operations). Autodesk, on a January fiscal year, reported remaining performance obligationsRPO — the total value of work a company has been contracted to deliver and has not yet recognised as revenue. It is contracted, not forecast, which makes it the most conservative visibility a software business can offer: a backlog with signatures on it. The share converting inside twelve months tells you how much of it is money for next year rather than for the decade. of $8.30 billion at 31 January 2026, of which 66% converts within twelve months (Autodesk 10-K FY2026).

The operations estate of IIoT platforms, connectivity middleware and generic analytics has been a graveyard, and its epitaph was written this year by one of its own believers. In the second quarter of fiscal 2026 PTC divested Kepware and ThingWorx, the industrial-connectivity and IIoT-platform businesses it had spent a decade building into the centrepiece of its factory-floor story (PTC 10-Q for the quarter ended 30 June 2026). Its ARR excluding those businesses grew 7% (9% in constant currency) in the June 2026 quarter; the reported total was flat. A company that sells software for a living decided the plant-floor connectivity layer was worth less inside its portfolio than outside it.

04Sizing the pools — context, not conclusion

Two numbers set the scale of the demand, and both get stretched in the telling.

Global industrial robot installations reached 542,000 units in 2024 with an operational stock of 4,664,000 units, up 9%, on the International Federation of Robotics’ World Robotics 2025 report published 25 September 2025. The IFR forecast 575,000 installations for 2025 — about 6% growth — and projected the annual figure to pass 700,000 by 2028, an average of roughly 10% a year (IFR, World Robotics 2025 press release). Asia took 74% of new deployments, Europe 16%, the Americas 9%.

The second number is the one to keep in front of you throughout. In 2024, for the first time, Chinese manufacturers sold more robots at home than foreign manufacturers did: domestic share rose to 57% from 47% in a single year. In metals and machinery — 54,600 units, a record — Chinese suppliers held 90%. In electrical and electronics, China installed 83,000 units, 64% of the world’s total for that industry, with domestic producers supplying 59%. In textiles, leather and apparel, every one of the 5,700 units installed came from a Chinese manufacturer (IFR press release, “China Tops World Record of 2 Million Factory Robots”, 25 September 2025).

Sources: IFR World Robotics 2025 (installations, Chinese domestic share, metals & machinery); A3 North American order statistics for calendar 2025 as reported 7 February 2026. Dated figures; the A3 series counts orders, not installations, and the two are not additive.

A ten-percent-a-year unit market whose largest national buyer watched its domestic suppliers gain ten points of share in twelve months is not a pricing environment. It is a share war with a state balance sheet on one side. Growth of that kind lifts revenue at every layer and margin at almost none of them.

For the software pools, the honest answer is that this brief has not verified a third-party industrial-software TAMTotal addressable market — a consultant’s estimate of every dollar that could theoretically be spent on a category. TAM numbers are the most quoted and least verifiable figures in any technology thesis, since nobody audits them and everybody selling into the market benefits from a larger one. to a standard that would let it carry a conclusion, and does not use one. What it uses instead are the disclosed revenue bases of the actual competitors: Siemens Digital Industries at €17,788 million, Dassault at €6.24 billion IFRS revenue, PTC at $2,739 million, Autodesk at $7,206 million, Rockwell’s Software & Control at $2,383 million, Emerson’s Software and Control at $5,691 million. Those are facts, and they bound the pool from below far more reliably than any forecast.

05Conclusion first: rent versus story

The durable profit in this stack concentrates in proprietary engineering estates — the controller, its tool and its protocol — and in application-engineering chokepoints in sensing. It has already been competed out of the robot arm, is being competed out of drives and motion, and has never existed in integration. The market is pricing an AI-and-humanoid narrative onto layers that cannot hold a rent, while the layer that can holds it quietly and is valued as a cyclical.

Howard Marks’s formulation applies almost too neatly: it is not what you buy, it is what you pay. Here the problem comes before price: the thing most investors think they are buying — exposure to robot units — is exposure to the one node in the chain where nobody has ever earned a durable return.

Five conclusions follow. Each contradicts a specific consensus view, rests on a company’s own filing, and is falsifiable and dated in Section 12.

1. 1The robot-hardware layer is not merely low-margin; it is capital-destroying, and the incumbents are exiting it. The consensus holds that robot OEMs are the primary way to own the automation wave. Teradyne’s Robotics segment lost $99.4m pre-tax on $308.3m of 2025 revenue with $733.5m of assets committed; ABB agreed to sell its entire Robotics division, which earned a 12.1% Operational EBITA margin (ABB’s own measure of operating profit, before interest, tax and the write-off of intangibles from past acquisitions) on $2.3bn of 2024 revenue against 19.0% for the group; Yaskawa’s consolidated operating margin was 8.7% with a 7.7% return on equity — profit measured against the shareholders’ own money in the business — in the year to February 2026. Three of the four Western-listed robot franchises are shrinking, loss-making or sold. Known Known
2. 2The industrial-software land grab has destroyed return on capital at the very companies the market rewards for it. Consensus says software acquisitions transform an industrial’s quality. Siemens Digital Industries’ profit divided by its own disclosed net capital employed — the money tied up in the division — fell from 33.4% to 11.5% in one year as it bought Altair and Dotmatics. Emerson’s computed ROIC (return on invested capital: the profit earned on every dollar committed to the business) is 8.3% against an estimated 9% cost of capital (what that money must earn to have been worth committing at all), with goodwill and intangibles — the accounting record of what was paid above the worth of the assets acquired — equal to 95% of invested capital. Fat software gross margins bought at 8× sales are not a moat; they are a liability with a good story. Known Known
3. 3Open standards have strengthened the controls franchise, not weakened it, and controller virtualisation will strengthen it again. Consensus expects OPC UA and software-defined control to commoditise the PLC. Three vendor-convened protocols hold 75% of new industrial network nodes; OPC UA FX was still at release-candidate stage — a near-final draft still out for testing, not a published standard — in March 2026 with its multi-vendor demo scheduled for November 2026; and Siemens’ own virtual PLC is sold as “completely compatible with TIA Portal” — the engineering software in which its customers’ control programs already live. The box is being commoditised by its owner, on purpose, to move the toll from steel to licence. Inference — mechanism well evidenced, outcome not yet observed
4. 4Machine vision is not one business, and only one of its two models earns. Consensus treats machine vision as a single defended niche. Keyence earned a 50.9% operating margin on 83.5% gross margin in the year to March 2026; Cognex earned an 18.4% EBIT margin and a computed 9.3% ROIC, roughly at its cost of capital, and names in its own 10-K the threat from “open-source tools available for free… including tools using AI.” The rent is in the direct-sales application-engineering model and product breadth, not in the vision algorithm — and the algorithm is exactly what AI is making free. Known Known
5. 5The humanoid programme is, on the primary evidence, one robot. Consensus reads the BMW–Figure deployment as proof of industrial humanoid viability. BMW’s own release of 25 June 2026 states that one Figure 02 robot inserted sheet-metal parts for the welding process over ten months, supporting production of over 30,000 X3 vehicles. Tesla’s 10-K states plainly: “We have yet to commercialize Bots and cannot predict how demand for Bots will develop.” Roughly four orders of magnitude separate the narrative from the disclosure. Known Known

06Value capture by layer — the return-on-capital evidence

Margin is where most of this industry’s analysis stops. It should not. A capital-light software business at 36% operating margin and a capital-heavy motion business at 12% can end at the same return, and in this stack they frequently end in the wrong order. Nothing here counts as a durable profit until it has cleared a cost of capital.

The test used throughout is return on invested capital against cost of capital, and the two halves are worth separating. Return on invested capital asks what the business earns on every dollar tied up in it — the factories, the working capital, and, crucially, the money spent buying other companies. Cost of capital is the return that money could have earned somewhere else at similar risk; it is the rent the business owes its owners simply for holding their capital. A company earning less than its cost of capital is not a slow-growing business. It is one that consumes wealth while reporting a profit, and margin will never tell you which kind you are looking at.

A stock return tells you what the whole business earns. For an acquisitive industry the question is what the marginal capital earned — and the arithmetic has to be handled carefully, because a single year’s incremental return is meaningless when invested capital shrinks. Emerson’s FY2025 incremental ROIC on this method computes to roughly −577%, because NOPAT rose while invested capital fell; Rockwell’s to +80%, because both fell. Neither number means anything. The figure below therefore reports incremental return only for the names that actually added material invested capital in their latest year, where the ratio answers a real question: was the new capital worth deploying?

Segment level shows the layer effects that consolidated returns obscure.

Two segment-level facts in that chart carry more weight than any consolidated number.

First, Rockwell’s internal mix. Software & Control produced $708 million of the company’s $1,703 million of fiscal 2025 segment operating earnings — 41.6% of earnings on 28.6% of sales — at a 29.7% margin, against 18.0% for Intelligent Devices and 14.5% for Lifecycle Services (Rockwell 10-K FY2025, summary of results of operations). In the six months to 31 March 2026 that segment margin reached 33.1%, and 34.9% in the March quarter alone, against 30.1% a year earlier (Rockwell 10-Q for the quarter ended 31 March 2026).

Second, Rockwell’s pricing. In fiscal 2025 the company recorded price realisationThe part of a company’s sales growth that came from charging more for the same thing, separated from the part that came from selling more of it. Issuers disclose it because it is the cleanest evidence of pricing power there is: anyone can grow revenue in a boom, but only a seller customers cannot easily leave can raise price into a falling market. of roughly three percentage points of total company sales while volume subtracted about two, delivering 1% total growth (Rockwell 10-K FY2025, MD&A). Emerson disclosed the same shape: Intelligent Devices underlying sales up 2% “on higher price”; Software and Control underlying up 5% on “2.5 percent higher volume and 2.5 percent higher price” (Emerson 10-K FY2025). That is what setting your own price into a soft market looks like on paper. No robot OEM in this coverage disclosed anything comparable.

07Who sets the price — the durability leg

A moat answers one question: when a credible substitute arrives, does the return survive? A market position is not the same thing. Four substitutes are arriving in this stack at once. They hit different layers, and the differences are the investment case.

Force one: Chinese champions

This force is the furthest along and the least appreciated, and it has moved strictly bottom-up. It took the arm first. In 2024 Chinese makers passed foreign ones in their own market for the first time, at 57% against 47% a year before, and reached 90% in metals and machinery (IFR China release, 25 September 2025). It is now taking the drive: Yaskawa told investors in April 2026 that it had accelerated restructuring in China “including the consolidation of substrate production and the downsizing of Drives production capacity” and that “large-scale low-margin projects in Korea and China are expected to phase out,” with the aim of restoring the mix toward “standalone unit sales that generate normal profitability” (Yaskawa FY2025 results-briefing Q&A, 13 April 2026).

A Japanese motion leader is telling the market that its Chinese and Korean volume is dilutive, that it is shrinking capacity in China, and that its route back to margin runs through selling fewer, better-mixed units. Price-takers talk like this. Price-setters do not.

The primary evidence does not yet show Chinese entry displacing mid-to-large controls at Western multinationals. Emerson’s fiscal 2025 disclosure records China down 3% in Intelligent Devices and down 6% in Software and Control (Emerson 10-K FY2025) — a weak market, not a lost one. Siemens reported orders and revenue up in all regions in the fourth quarter “with strong growth contributions from China” (Siemens Earnings Release Q4 FY2025). Rockwell’s Asia-Pacific sales fell 5% in fiscal 2025 — on a base of $1,024m, 12.3% of the company (Rockwell 10-K FY2025) — which is a small exposure to a real risk.

The sequencing matters more than the level. Chinese champions have monetised the layers where the specification is a datasheet, and they are now approaching the layer where the specification is a customer’s twenty-year-old engineering standard. That last step is a trust-and-migration problem rather than a price one. It is why this brief expects the erosion of the controls layer to run a decade behind the erosion of the motion layer, and to begin at the low end.

Force two: AI and the hyperscaler / model vendor

Two words carry the weight in this section. A hyperscaler is one of the handful of companies that operate computing at planetary scale — the cloud and model providers whose buying decisions set the price of compute for everyone else. Physical AI is the industry’s name for pointing those same trained models at the physical world: a machine that perceives its surroundings and decides what to do next from a learned model, rather than executing a sequence an engineer wrote in advance. The distinction that matters commercially is that a learned model is copied at no cost and a factory is not, so anything the model absorbs stops being scarce.

NVIDIA’s own revenue disclosure settles its role. In the fiscal year to 25 January 2026, NVIDIA reported $215,938 million of revenue, of which Data Center was $193,737 million and Automotive — the line that contains its self-driving platforms and related robotics/edge offerings — was $2,349 million, or 1.09% of the total (this is not a robotics-only line; NVIDIA separately disclosed Physical AI revenue of north of $6 billion in fiscal 2026) (NVIDIA 10-K FY2026, revenue by specialised market). Automotive grew 39%. It is still barely a percent.

NVIDIA is not trying to become a robot company, and its actions say so more clearly than its keynotes. On 31 May 2026 it announced the Isaac GR00T reference humanoid — “the first open humanoid robot reference design built on NVIDIA Jetson Thor and the NVIDIA Isaac GR00T open development platform” — aimed at academic and research institutions, with the hardware to be shipped by Unitree in late 2026 (NVIDIA newsroom, 31 May 2026). An open reference design manufactured by a Chinese OEM is not an attempt to capture the robot. It is an attempt to make the robot free so that the compute inside it is not.

At GTC on 16 March 2026 NVIDIA named ABB Robotics, FANUC, KUKA, Universal Robots, YASKAWA, Agility, Boston Dynamics, Figure, NEURA Robotics, AGIBOT and 1X among the companies building on its platform — Cosmos world modelsModels trained to predict how a scene changes when something acts on it — what happens if this arm pushes that box. They let a robot be trained largely in simulation, where a million attempts cost compute rather than crashed machinery, and only then be moved onto real hardware., Isaac Sim and Isaac Lab, the Newton physics engine, Isaac GR00T N1.7 in early access with commercial licensing, and Jetson Thor (NVIDIA newsroom, 16 March 2026). Jensen Huang’s framing was “Physical AI has arrived — every industrial company will become a robotics company.” Read the roster rather than the slogan: every major robot maker in this coverage is now a customer of the same brain. A component every competitor buys from the same supplier stops being a differentiator and becomes a cost line. The autonomy stack, which was supposed to be where robot makers finally earned a software margin, is being bought in.

That is a rational and, for the rest of the stack, a hostile strategy. It is unambiguously bad for anyone whose margin depends on a proprietary perception stack — which is why Cognex’s own 10-K now names as competitors “internal engineering departments of current or prospective customers, as well as open-source tools available for free from various companies, including tools using AI” (Cognex 10-K FY2025, risk factors). It is close to neutral for the controls layer, because a foundation model does not schedule a safety-rated interlockA hard rule wired into the machine: the press cannot close while the guard is open, the robot cannot move while the gate is unlocked. Interlocks are the mechanism by which a factory is made safe, and they are required to work identically every single time — which is a different engineering problem from making a machine behave well on average. on a deterministic scan cycleA PLC does not react to events as they arrive. It runs a fixed loop — read every input, execute the whole program, write every output — over and over, typically in a few milliseconds, and takes exactly the same time round every loop. That predictability is the point: an engineer can prove what the machine will do and when. A model that answers “usually within 40 milliseconds” cannot be certified to hold a guard shut., and the liability structure of a certified safety function has no obvious home in a probabilistic policy.

Force three: open interoperability standards

Covered in Section 2 and priced here: real, dated, and slow. OPC UA FX at release-candidate stage in March 2026 with a multi-vendor demonstration scheduled for November 2026 is a 2030s risk to the installed estate, not a 2020s one. The correct treatment is to hold the controls verdicts with high confidence over five years and low confidence over fifteen, and to say so.

Force four: robot-hardware commoditisation

This one is not a forecast. It has happened, and the incumbents have already responded to it with the most honest signal available to a board: they sold. ABB’s divestment release states that the Robotics division had 2024 revenues of $2.3 billion, about 7% of group revenue, and an Operational EBITA margin of 12.1%, against a group Operational EBITA margin of 19.0% and a return on capital employed of 25.3% for full-year 2025 (ABB release, 8 October 2025; ABB Q4 and FY2025 results, 29 January 2026). The board chose a $5.375 billion sale over its own previously announced spin-off, banking an expected pre-tax book gain of about $2.4 billion.

Both quotations from ABB’s ad hoc release of 8 October 2025. The transaction is subject to regulatory approval in the EU, China and the US and was expected to close in mid-to-late 2026; as of this brief’s cutoff, completion is not verified.

08The per-name reads

Every name below is graded on its own latest filing or disclosure, opened this session. The grade answers one question only: does this company’s competitive position let it earn a return above its cost of capital through a cycle, and does the evidence say that position is structural rather than temporary? It is not a view on price, and nothing here is a recommendation.

Layer 1 — Enabling silicon and edge compute

Layer 2 — Actuation, motion and the robot OEMs

This is the layer the mandate anticipated would hold little durable profit, and the evidence is more emphatic than expected: on the disclosed 2025–26 evidence, no company in this coverage earns a durable excess return from selling robot arms or general-purpose motion hardware.

KUKA is characterised rather than graded: it has been majority-owned by Midea Group since 2017 and is no longer separately listed with disclosure sufficient for a return-on-capital read. Its relevance here is structural — the second-largest European robot franchise passed into Chinese ownership nearly a decade before the current wave, and the IFR’s 2024 share data suggests the strategic logic has been borne out at the arm level. This brief does not verify KUKA’s current financials and no conclusion rests on them.

Layer 3 — Sensing and machine vision

Two companies sell industrial vision. One earns 51% operating margins; the other earns roughly its cost of capital. The difference is not technology.

Layer 4 — Controls and industrial protocols

The layer this brief identifies as the durable one. The grades below are not uniformly advantaged: holding the rent and earning a return on the capital you have spent are two different achievements, and only Rockwell currently does both.

Layer 5 — Industrial software

AspenTech is no longer a public company and is therefore characterised, not graded. Emerson announced on 27 January 2025 the acquisition of the AspenTech shares it did not own at $265.00 per share — $7.2bn for the minority, valuing the whole at a $17.0bn fully diluted market capitalisation and a $16.8bn enterprise value (Emerson 8-K exhibit 99.1, 27 January 2025) — and completed the purchase on 12 March 2025. AspenTech now reports into Control Systems & Software leadership and its results are consolidated into that segment for all periods presented (Emerson 10-K FY2025). Its economics survive in this brief only as the amortisation line in Figure 5.

Layers 6 and 7 — System integration and the end-user verticals

Neither layer contains a graded name here, and that absence is itself the finding.

System integration is where the largest share of a delivered project’s value is created and where the least is captured. It is a labour business. Engineering hours, safety certification, commissioning, ramp support. It is fragmented and overwhelmingly private, which is why no listed pure-play appears in this coverage. Its structural weakness is visible in Figure 3 — an integrator’s certifications are assets held hostage to somebody else’s platform, and its customers hold the specification. Rockwell captures a slice of it through Lifecycle Services, which is instructive: that segment earned a 14.5% operating margin in FY2025 against 29.7% for Software & Control, and absorbed $262m of the year’s purchase-accounting charges (Rockwell 10-K FY2025). Services attached to a controls franchise are a good business; services as a standalone are a wage.

The end-user verticals set the demand and pay the rent. Their behaviour matters here in three specific ways. Automotive, historically the anchor customer, is weakening. China’s automotive robot installations fell 12% to 57,200 units in 2024, and Yaskawa attributes its robotics profit decline to sluggish automotive capital investment in Japan, the Americas and Europe. Electronics is now the largest customer industry in China at 83,000 units, 64% of the world’s total for that industry, and it buys on price. Logistics and warehouse automation is the growth vertical — it is Cognex’s largest end market and Teradyne’s stated Robotics focus — and it is also the vertical where end-users have the most in-house engineering capability, which is why Cognex names customers’ own engineering departments as competitors. Food, beverage and pharma is the quiet bright spot for the controls layer. It is regulated, validation-heavy, and therefore the most switching-cost-intensive vertical in the stack, because a change of control platform triggers re-validation of every recipe.

08bIncentive mapping — what each participant is actually paid to do

Strategy documents describe intentions. Compensation plans describe behaviour. The SEC’s pay-versus-performance rule requires every issuer to publish an unranked list of the financial measures it considers most important in linking executive pay to performance, which makes this the one place where a stack’s incentives can be read directly rather than inferred.

Of the fourteen US-listed names in this coverage, four disclose a return-on-capital measure among their most important metrics, and two of those four use a definition that cannot see an acquisition’s price. The rest are paid on earnings, cash flow, revenue and relative share-price performance — every one of which a debt-funded acquisition improves on day one, whatever was paid for it. That is much of Figure 8.

The rest of the map — participants who do not file a proxy

The non-US and private actors have incentives too, and they are legible from behaviour.

• The Japanese motion and sensing houses. FANUC, Yaskawa and Keyence run on domestic governance with low leverage and large net cash. The consequence is patience and an aversion to the acquisition treadmill — which is why Keyence has never needed a return metric to earn 51% and why Yaskawa can spend a year restructuring Chinese capacity rather than defending share. It also means none of them will buy their way into the controls layer.
• The European electrification incumbents. ABB and Schneider are both being pulled by their best-performing end market — data-centre and grid electrification — and both are allocating accordingly. ABB sold Robotics; Schneider’s Energy Management grew 10.3% organically at a 21.8% margin against Industrial Automation’s 3.0% at 14.2%. The consequence for the stack is that the two most capable European automation franchises are quietly de-prioritising automation.
• Siemens. Uniquely positioned and uniquely conflicted. Its incentive is to defend the PROFINET-plus-TIA-Portal estate, which argues for organic reinvestment; its recent behaviour has been to buy software at prices that doubled the capital employed against a shrinking profit. The consequence is that the best asset in the stack is being managed for scale rather than for return.
• NVIDIA. Paid on revenue and non-GAAP operating income, with an Automotive end-market line (self-driving platforms and related robotics/edge) of only 1.09% of revenue — not a pure robotics P&L; Physical AI was separately disclosed as north of $6bn in FY2026, still a single-digit share of the group. The consequence is unambiguous: NVIDIA’s incentive is to maximise the number of robots, not the margin on any of them, which is why it publishes open models and an open reference design manufactured by a third party. Everyone downstream of the compute should read that as a commitment to commoditise them.
• The humanoid developers. Private, venture-funded, and paid in valuation marks rather than in unit economics. The consequence is a structural bias toward demonstrations that raise the next round over shipments that would disclose a cost. BMW’s release — one robot, one task, ten months — is what the disclosed reality looks like when a customer rather than a developer writes the description.
• The system integrators. Paid by the project, certified by the platform vendor, and therefore structurally aligned with whichever controls incumbent their engineers are trained on. This is the most under-appreciated reinforcement of the protocol franchise: a plant’s integrator has a personal, career-length interest in the specification not changing.
• The end-user verticals. Paid on plant uptime and unit cost, not on supplier diversity. The consequence is the single sentence that explains the whole stack: a plant manager will pay a 20% premium on a controller to avoid a 2% risk of an unplanned line stoppage, and will not pay a 2% premium on a robot arm for anything at all.

Recency check — what has changed since the last annual reports

Three developments after the fiscal-year data above deserve to be on the record, and one of them cuts against a grade in this brief.

First, and against this brief: Cognex has re-accelerated sharply. First-quarter 2026 revenue was $268.4 million, up 24% year on year, with gross margin at 71% against 67% a year earlier, on broad-based end-market strength and a favourable currency effect (Cognex 10-Q for the quarter ended 5 April 2026). That is a materially better quarter than the “exposed” grade anticipates. The grade is held for now because a single quarter of cyclical recovery with a currency tailwind does not answer a structural risk that the company itself names in its own risk factors — but the falsifier in Figure 16 is written precisely to catch this, and a reader who weights the quarter more heavily than the risk factor would reasonably reach neutral instead.

Second, Novanta’s automation exposure grew but the mix question remains open. First-quarter 2026 revenue was $257.7 million against $233.4 million, with Automation Enabling Technologies at $131.2 million and, within it, Robotics and Automation at $84.2 million against $78.6 million (Novanta 10-Q for the quarter ended 3 April 2026). Growth is real; the 2025 cash-flow collapse to $64 million and the 5.9% return on capital are what the grade rests on, and neither is resolved by one quarter of revenue.

Third, Honeywell’s reporting structure changed twice. Its 10-Q for the quarter ended 30 June 2026 confirms that prior to the Aerospace spin-off the company managed through four reportable segments — Aerospace Technologies, Building Automation, Process Automation and Technology, and Industrial Automation — and that Honeywell Aerospace’s historical results will be reflected as discontinued operations in Honeywell Technologies’ statements going forward (Honeywell 10-Q for the quarter ended 30 June 2026). The FY2025 segment figures used in Section 8 are on the earlier four-segment basis and are not directly comparable with what Honeywell Technologies will report for 2026.

Fourth, and the item that matters most for how this brief should be read: the automation capex cycle has turned up hard, which strengthens the bear’s cyclicality argument and must be conceded in full. ABB’s second quarter of 2026, reported 16 July 2026, showed record orders of $12,042 million, up 30% (28% comparable), revenues up 14% to $9,475 million, an Operational EBITA margin of 20.2% and a return on capital employed of 28.4% (ABB Q2 2026 press release, 16 July 2026). Siemens, on 13 May 2026, reported second-quarter orders up 18% comparable to €24,114 million with a book-to-billOrders taken divided by revenue billed in the same period. Above one means the order book is growing faster than the company can ship, which is the earliest reliable sign a capital-goods cycle has turned up; below one means the opposite, usually several quarters before it reaches reported revenue. of 1.22, and raised its Digital Industries full-year guidance to comparable revenue growth of 7–10% and a profit margin of 17–19%, from 5–10% and 15–19% (Siemens Earnings Release Q2 FY2026, 13 May 2026). Rockwell’s first-half fiscal 2026 sales rose 11.9%.

Two readings are available and the honest brief holds both. The cyclical reading is that much of the margin strength cited throughout Sections 6 to 8 is operating leverage on a recovering order book, and will reverse. The structural reading is that Rockwell’s Software & Control margin advantage over its own device segment — 29.7% against 18.0%, widening to 34.9% against 20.9% in the March quarter — is a mix fact that a cycle does not create and a downturn will not erase, and that Siemens raising the floor of its Digital Industries margin guidance by two points is a statement about structure, not about one quarter. This brief holds the structural reading, and the falsifiers in Figure 16 are written so that a reader who prefers the cyclical one can watch the same numbers and reach the opposite conclusion on schedule.

One more piece of capital allocation belongs here, because it is the same board acting twice. Having agreed to sell Robotics, ABB disclosed in the same July 2026 release an offer to acquire Rotork plc, described as “an important step to expand the ABB Automation portfolio — adding actuators and building on ABB’s broad market reach”, one of three recent acquisitions that together would add about 3.5% to 2025 revenues (ABB Q2 2026 press release, 16 July 2026). ABB sold robot arms and is buying valve actuators. That is not a contradiction; it is the thesis restated in capital. A robot arm is re-quoted every project. A valve actuator is written into a process plant’s engineering standard and re-ordered as a spare for thirty years. Both are motion. Only one is specified in.

09The scorecard: durable rent versus priced-in story

Two axes, both real, neither a valuation call. The vertical is moat durability on this brief’s single criterion — does the excess return survive the arrival of a credible substitute. The horizontal is how much AI-and-humanoid narrative the name carries in public discussion and in its own investor communication, which is a proxy for how much of the story is already an expectation rather than a disclosure.

Neutral, and therefore in neither column: FANUC (21.4% margin but no divisional profit disclosure), Siemens (advantaged franchise, deteriorating return on deployed capital), Honeywell Technologies (post-spin economics not yet observable), Schneider Electric (advantaged in electrification, not in automation), Parker Hannifin (18.7% ROIC earned in aerospace), AMETEK (13.8% ROIC — clears the hurdle, is not a chokepoint), Teledyne (7.5% ROIC, largely a defence business), Dassault Systèmes (genuine PLM switching cost, decelerating, return on capital not computed). Grades are structural assessments of competitive economics, not investment advice.

10The bear case, argued at its strongest — then answered

The opposing worldview is not the strawman that “robots will be big.” The serious bear case against everything above runs like this, and it deserves to be written to persuade.

Each of those is a real argument. Here is what the evidence does to them.

On the abstraction moving. This is the strongest point and it is half right. The abstraction is moving; it is moving in a direction that helps the incumbent. Siemens’ own product page for the virtual PLC states the controller “is completely compatible with TIA Portal, allowing you to build on projects that are already in the tool” (Siemens, SIMATIC S7-1500V). Virtualisation that leaves the engineering tool untouched upgrades the licensing model and leaves the lock exactly where it was. The bear case requires the tool to be displaced too, and no tool displacement is visible in any disclosure this session. The correct concession: if a third-party engineering environment ever achieves credible bidirectional round-tripping of TIA Portal or Studio 5000 projects at production scale, the thesis is in serious trouble. That is the falsifier, and it is stated as such in Section 12.

On models dissolving migration cost. Also half right, and the half that is wrong is the expensive half. Translating ladder logic is the easy part of a migration. The hard parts are re-validating safety instrumented functionsThe protective functions that shut a process down before it hurts somebody — trip the pump on high pressure, close the valve on flame failure. IEC 61508 and 61511 are the international standards that govern how they are designed, proved and documented, and compliance is demonstrated to a certifying body rather than asserted by the vendor. to IEC 61508/61511, re-qualifying regulated recipes under GxPThe family of “good practice” regulations governing regulated manufacturing — pharmaceuticals, medical devices, food. Under them a plant must prove, in documents an inspector will read, that its equipment does what it is supposed to do. Change the control system and that proof has to be produced again, batch by batch, before the line may legally run., re-certifying the integrator, and accepting personal liability for a line that can kill someone. A model that emits equivalent logic does not emit a safety certificate. This is why the food, beverage and pharma verticals carry the deepest lock and why the erosion, when it comes, will start in unregulated general industry and low-end discrete — exactly where Chinese small PLCs have already landed.

On the Chinese sequence. The bear is right that the sequence has not stopped, and this brief does not claim it will. It claims the sequence is ordered, and the ordering is observable: arms first (57% domestic share in 2024, from 47%), drives next (Yaskawa downsizing Chinese drive capacity in 2026), controls last. The interval between layers has been roughly five to eight years. That does not make the controls layer safe; it makes it the layer with the most time, and time is what compounding needs. An investor should hold the controls verdicts with high confidence over five years and treat a fifteen-year view as genuinely uncertain.

On cyclicality. This is the point the bull case must answer most carefully, and the honest answer is that it is partly conceded. Rockwell’s backlog did fall, to $2,878m from $3,091m. But cyclical recovery does not explain the specific shape of the fiscal 2025 disclosure: price up roughly three points while volume fell two, in the trough year, not the recovery. A price-taker in a destocking year takes price cuts. Rockwell took price. And the margin differential across its own three segments — 29.7% against 18.0% and 14.5%, on shared sales and supply-chain organisations — is a structural fact about product mix, not a cyclical one.

On humanoids. Here the bear case has no primary disclosure behind it at all. BMW’s own release of 25 June 2026 describes the Spartanburg programme precisely: one Figure 02 robot, inserting sheet-metal parts for the welding process, over ten months, in support of more than 30,000 X3 vehicles, with Figure 03 now beginning logistics sequencing work (BMW Group press release, 25 June 2026). No cost, no unit economics, no commercial terms are disclosed. Tesla’s 10-K for 2025 states that Optimus is “in development,” that Bots are “in a nascent industry that has yet to develop commercially,” and, flatly, “We have yet to commercialize Bots and cannot predict how demand for Bots will develop” (Tesla 10-K for FY2025, business and risk factors). Figure remains private and its deployment scale is not verifiable from primary disclosure; this brief characterises it and grades nothing on it.

The steelman survives in one place and one place only: the tool-displacement argument. Everything else in the bear case is either already priced into the exposed grades or is a decade-scale risk correctly held at low confidence. That is where an investor in this stack should spend their attention, and it is not where the market is spending its.

11Second-order effects — where this goes next

The consequences that matter arrive when the thesis is right and everybody eventually notices.

First: cheaper robots strengthen the controls layer. Every dollar competed out of the arm is a dollar the end-user can spend on a project, and every project that gets built adds nodes to a network estate and lines to a control programme. A commoditised arm makes automation projects clear their hurdle rateThe minimum return a company requires before it will approve a capital project. A factory manager with a fixed hurdle approves more automation projects when the equipment gets cheaper — which is why falling component prices can raise the number of projects even as they destroy the margin of whoever makes the component. more often. The controls incumbent is a toll-taker on project count, not on component value. Chinese robot price deflation is therefore, on net, good for Rockwell and Siemens and bad only for the companies whose revenue is the arm.

Second: NVIDIA is deliberately commoditising the layer beneath itself, and the collateral damage lands on machine vision before it lands on robots. The Isaac GR00T reference humanoid, open and shipped by Unitree, is a template for making the body free. But the same open stack — Isaac Sim, Isaac Lab, GR00T foundation models, ROS middlewareThe Robot Operating System — free, open-source plumbing that lets the separate pieces of a robot (cameras, arms, planners, navigation) talk to one another without a vendor’s proprietary framework in between. It is the standard nearly every robotics research group and startup builds on, and it is the reason the software layer of a robot is far less defensible than the controls layer of a factory. — makes perception free first, because perception is software and bodies are not. Cognex’s own 10-K already names free AI and open-source tools as competitors. Expect the margin compression to show up in general-purpose vision products before it shows up in arms, and expect Keyence’s advantage to widen relative to Cognex’s precisely because Keyence sells an organisation rather than an algorithm.

Third: the industrial-software M&A cycle is approaching its own reckoning, and the evidence is already in. Emerson at 8.3% ROIC, Siemens DI at 11.5% profit-to-net-capital-employed, PTC selling Kepware and ThingWorx, Zebra at 8.4% after Elo, Novanta at 5.9%. That is not one company’s misjudgement; it is a pattern of five. When a strategy that every peer pursued produces sub-cost-of-capital returns at every peer, the strategy was the error. The second-order consequence is a coming preference for organic software growth and for divestiture — which will look like retreat and will be an improvement. Watch for more Kepwares.

Fourth: the winner of the humanoid race, if there is one, will not capture the humanoid’s value. If a general-purpose machine ever reaches cost parity, the body will be built where bodies are built — NVIDIA’s own reference design ships from Unitree — and the value will accrue to whoever owns the compute, the fleet-management software, and the safety certification. None of those is a robot company. This is the same distribution of spoils that the arm produced, one abstraction level up, and there is no evidence in any disclosure opened here that the humanoid developers have solved for it.

Fifth: the separations are the tell. ABB sold Robotics. Honeywell separated Aerospace on 29 June 2026 and left a pure-play automation company behind. Emerson bought AspenTech outright and folded it into Control Systems & Software. PTC sold its plant-floor connectivity businesses. Four boards, four different conclusions about which of their own assets deserve their capital — and all four moved capital toward the control-and-software estate and away from generic hardware and generic platforms. Boards act on information investors do not have. When four of them act the same way inside eighteen months, the burden of proof shifts.

12Falsifiers and dated predictions

A thesis that cannot be killed is not a thesis. Below are the specific, observable events that would break each load-bearing verdict, followed by dated, testable predictions over a three-to-five-year horizon.