Chapter 9 — Manufacturing and Industry 4.0

Manufacturing is the AI deployment domain with the deepest history of integration, the strongest direct-productivity record, and the most-cautionary safety-critical failures. The deployment history runs from the numerical-control machines of the 1950s, through the statistical-process-control discipline of the 1980s, into the sensor-and-machine-learning wave of the 2000s, and now into the generative-design and humanoid-robotics frontier of 2024–2026. The pattern is more continuous than the AI literature’s headline cases suggest. Manufacturing has had high-functioning automation for decades; the contemporary AI wave is layering on top of an existing infrastructure rather than building from greenfield.

The contemporary Industry 4.0 framing originated in 2011 with the German government’s Industrie 4.0 strategy, articulated formally in the 2013 Acatech report (Kagermann et al., 2013). The framing’s central thesis is that the integration of cyber-physical systems, the Internet of Things, and machine learning constitutes a “fourth industrial revolution” — distinct from the steam-mechanisation of the 1780s, the electrification-and-mass-production of the 1870s, and the computer-and-automation of the 1970s. Klaus Schwab popularised the framing globally through the World Economic Forum (Schwab, 2016). The framing has been useful as organising vocabulary; its substantive content is more contested. Critics note that the productivity numbers underlying the “fourth industrial revolution” claim are mixed, with much of the manufacturing productivity gain over 2000–2024 attributable to global-supply-chain optimisation rather than to specifically AI-driven automation (Acemoglu and Restrepo, 2020). The chapter accepts the Industry 4.0 vocabulary as useful organising language while staying alert to where the framing over-claims.

This chapter develops the manufacturing AI landscape across thirteen sections. Section 9.1 covers the Industry 4.0 framework and its critique. Section 9.2 covers predictive maintenance — the deepest-deployed manufacturing AI category. Section 9.3 covers computer-vision-based quality assurance. Section 9.4 covers robotics from classical industrial robots to the contemporary humanoid push. Section 9.5 develops the Foxconn / Hon Hai case as the contract-manufacturing-at-scale exemplar. Section 9.6 covers Tesla’s vertical-integration approach to manufacturing. Section 9.7 develops the Boeing 737 MAX MCAS case in detail — the most-detailed contemporary cautionary case in safety-critical automation. Section 9.8 covers generative design. Section 9.9 covers digital twins. Section 9.10 covers the semiconductor industry, with attention to TSMC and the upstream-downstream cluster including Malaysian E&E. Section 9.11 covers Australian mining-and-resources automation. Section 9.12 develops the Malaysian E&E and broader Southeast Asian manufacturing context with attention to ViTrox Corporation. Section 9.13 sketches the 2026 frontier including humanoid robotics, labour displacement dynamics, and supply-chain resilience.

9.1 The Industry 4.0 framework — origin, evolution, and critique

The Industrie 4.0 programme was launched as a strategic initiative of the German federal government in 2011, with formal articulation in the 2013 Acatech report and substantial European Union policy alignment thereafter. The programme’s central proposition was that German manufacturing competitiveness — historically anchored in the high-quality engineering of the Mittelstand — required a transition to fully-instrumented “smart factories” that integrated cyber-physical systems, IoT-style sensor networks, and data-driven decision-making. The framing was deliberately chronological: the first industrial revolution centred on water and steam mechanisation; the second on electrification and the assembly line; the third on computer-numerical-control machines and early industrial automation; the fourth on cyber-physical integration with full-instrumentation data flows.

Klaus Schwab gave the framing global currency through his 2016 book The Fourth Industrial Revolution and through the WEF annual meetings that thematised the topic. The framing entered policy discourse globally; China’s “Made in China 2025” programme (2015) shares many elements; Japan’s “Society 5.0” (2016) is a parallel framing with broader social claims; the United States’ approach has been more decentralised but with similar component initiatives. Each major manufacturing economy now has Industry-4.0-equivalent vocabulary in its industrial policy.

The component technologies. Industry 4.0 is more aggregator than singular technology; the components include IoT sensors and edge computing, industrial machine learning, cyber-physical systems, cloud-and-edge integration, robotics and automation (including humanoid robots from 2024), digital twins, generative design, and additive manufacturing. The aggregator framing has been useful for organising disparate technologies under a common policy banner. It has been less useful for deciding which of the components is genuinely transformative and which is incremental. The 2024–2026 reality is that some components (predictive maintenance; computer-vision quality assurance) have produced demonstrable productivity gains at scale; others (digital twins at factory-level; generative design at production-volume scale) have produced more limited deployment than the framing implied.

The productivity-claim critique. A specific critique concerns whether the claimed productivity gains have actually materialised. Manufacturing productivity growth in OECD economies over 2000–2024 has been lower than in the 1990s, despite the broad adoption of the technologies the framing identifies as transformative. Acemoglu and Restrepo (2020) document this pattern across major economies. The proposed explanations vary: that the productivity gains are real but mismeasured; that the technologies have produced labour-displacement-without-productivity; that the productivity gains have been captured by specific firms rather than diffused broadly. Each explanation is partially true. The structural fact is that Industry 4.0 has not produced the order-of-magnitude productivity step-changes its more-enthusiastic articulations claimed; the deployment pattern is incremental productivity gain at firms that invest substantially, with limited spillover to firms that do not.

The implication for AI-deployment thinking. The Industry 4.0 framing’s most-useful insight is its emphasis that the interaction of multiple technologies — sensors, ML, robotics, cloud — is what produces deployment value, not any single technology in isolation. The Iansiti-Lakhani factory framework from Chapter 3 generalises the same insight: the AI factory’s value comes from the integration of the four components (data, ML, deployment, operations), not from any single component. The Industry 4.0 framing, applied to actual factories, is one specific instantiation of the broader factory thesis.

9.2 Predictive maintenance — the deployment success

Predictive maintenance is the AI manufacturing use case with the deepest deployment history and the clearest demonstrated value. The basic premise: instead of running equipment to failure or replacing components on fixed schedules, use sensor data and machine learning to predict when each specific component will fail and replace it just before. The economic value comes from three sources: avoided downtime (a failure during production is more costly than a planned maintenance window); avoided collateral damage (a failure that propagates damages adjacent components); and reduced spare-parts inventory.

The technology stack has matured over a 20+ year period. Early predictive maintenance used vibration analysis with rule-based thresholds; physics-based models followed in the 1990s and 2000s; machine learning entered the field in the late 2000s with classical methods (random forests, support vector machines on hand-engineered features); deep learning arrived in the late 2010s with the application of recurrent and convolutional architectures to sensor time series. The contemporary state of practice combines all of these.

Sensor proliferation. The deployment foundation is sensors. A modern industrial machine — a CNC machine tool, a turbine, a pump, a press — typically carries 20–200 sensors monitoring vibration, temperature, pressure, acoustic emissions, current draw, and process parameters. The cost of a sensor has fallen by 1–2 orders of magnitude since 2010, driven by smartphone-component supply chains. The cost of communicating sensor data has fallen similarly. The cost of storing and processing the data has fallen with the broader cloud-and-compute trends. The combined cost reduction has made comprehensive instrumentation economically viable for a much broader range of equipment than was the case in 2010.

The major deployments. GE Aviation operates one of the world’s largest predictive-maintenance deployments, with sensor data from approximately 30,000 commercial jet engines in service. Each engine produces 10–100 GB of data per flight; the cumulative dataset supports component-level failure prediction. GE’s “Predix” platform (founded 2015) was the company’s attempt to commercialise the underlying capability; the commercialisation story has been mixed (Predix was substantially scaled back in 2018–2019), but the core in-house deployment has continued and produced documented value. Similar deployments exist at Rolls-Royce (with the TotalCare programme), at Pratt & Whitney, and at Siemens Energy.

The cost-savings calculus. Detailed deployment studies (Lee et al., 2017; Susto et al., 2015; Carvalho et al., 2019) document predictive-maintenance cost savings of 8–25% relative to preventive maintenance baselines, with corresponding downtime reductions of 30–50%. The variability reflects substantial heterogeneity in starting conditions: facilities with poor preventive maintenance discipline have larger absolute gains but smaller relative gains. The 25–50%-improvement figure that is sometimes cited in vendor materials reflects best-case deployments rather than typical results.

Failure modes and limitations. Three failure modes recur. First, sensor failures themselves: a sensor that drifts or fails produces misleading data that the ML system may not detect. Sensor-validation infrastructure is necessary and frequently under-invested. Second, concept drift: the ML model trained on historical operating conditions may degrade as the equipment ages or as operating conditions change. Continuous monitoring of model performance, with retraining triggers, is necessary. Third, the integration with maintenance operations: the predicted failure must translate into a maintenance action, which depends on the work-order system, the parts inventory, the technician scheduling, and the production scheduling. Many deployments stall at this integration step.

The factory pattern, applied. Predictive maintenance is one of the cleanest examples of the AI factory thesis from Chapter 3. The data flows from sensors through aggregation to ML through deployment to operations to outcomes; the outcomes feed back into the data through new failure events that update the model; the loop turns continuously. A facility that has built this loop captures durable value; a facility that has the technology but not the loop captures little. The variation across firms is wide; the variation reflects loop-completeness, not technology choice.

9.3 Computer vision in quality assurance

Computer-vision-based quality assurance is the second-most-mature manufacturing AI category. The technology applies to defect detection (identifying damaged units on a production line), surface inspection (detecting cosmetic flaws), dimension measurement (sub-pixel-accuracy measurement of dimensions), and assembly verification (confirming that components are correctly placed). The deployment scale is substantial: roughly 70% of high-volume manufacturing globally uses some form of automated optical inspection (AOI), with the AI-augmented share growing rapidly through 2020–2026.

The classical era. Before 2015, AOI systems were predominantly rule-based and used hand-engineered image-processing pipelines. Cognex (founded 1981, US-based) and Keyence (founded 1974, Japanese) became the dominant vendors of this generation. The systems worked well for well-defined inspection tasks but required substantial expert engineering for each new application; the per-task development time was measured in weeks-to-months.

The deep-learning wave. Convolutional neural networks fundamentally changed the AOI landscape. CNNs learn relevant features from data rather than requiring hand-engineering, which compresses development time from weeks to days and makes inspection of subtler defects feasible. The deployment trajectory through 2018–2024 has seen most major manufacturers transition substantial portions of their inspection operations from rule-based to ML-based systems.

ViTrox Corporation — Malaysian AOI vendor. ViTrox (KL-listed; founded 2000 in Penang) is among Southeast Asia’s most-significant manufacturing-AI companies. The company makes automated optical inspection systems for the semiconductor and electronics industries, with focus on backend semiconductor packaging inspection (where Malaysia has substantial industrial cluster strength). ViTrox’s product line includes Machine Vision Systems, 3D X-ray inspection systems, and embedded software products. The company’s competitive positioning is mid-market — between the high-end Cognex/Keyence positioning and the low-cost Chinese-manufactured equipment — with particular strength in the semiconductor backend market in Asia. ViTrox revenue grew from RM 200 million in 2015 to RM 700+ million in 2023; the company is among Bursa Malaysia’s most-valued technology firms with a market capitalisation around RM 10 billion. The case is significant for KL-based students directly: ViTrox is a Malaysian-domiciled, Malaysian-listed firm operating at the global frontier of manufacturing AI, with R&D and engineering teams based in Penang. The case demonstrates that frontier manufacturing AI is not exclusively a US-or-China phenomenon; the Malaysian E&E sector hosts genuine manufacturing-AI capability at international competitive scale.

Other deployment patterns. Semiconductor wafer inspection: KLA Corporation (US-listed; ~USD 11 billion 2024 revenue) provides the high-end wafer-inspection systems for leading-edge process nodes. Automotive paint inspection: paint defects on automotive bodies are inspected with combined human-and-machine-vision systems; deep-learning-based paint inspection has reduced human inspector requirements substantially over 2018–2024. Pharmaceutical packaging inspection: package integrity, label accuracy, fill-level verification — high-volume tasks that ML-based vision systems handle at speeds humans cannot match. Food and beverage: foreign-object detection, fill-level verification, packaging integrity. Textile and apparel: defect detection in fabric production; the Asian textile cluster (China, Vietnam, Bangladesh, Cambodia) is increasingly deploying ML-based inspection.

Deployment realism check. Computer-vision QA deployments have a high success rate at the unit-level (the inspection works) but a more variable success rate at the operations-level. Common operational failures include integration with the production-line speed; false-positive rates that overwhelm operators; concept drift as products evolve; and the tail-of-defects problem (rare defects not in the training data and not detected). The mature deployment requires not just the ML capability but also the operational infrastructure to handle the failure modes.

9.4 Robotics — from FANUC to humanoid

The robotics history in manufacturing splits into three eras with distinct technology stacks and deployment patterns.

The classical industrial robotics era (1970s–2010s). The classical era is dominated by four major vendors: FANUC (Japan; founded 1956 as a Fujitsu subsidiary), ABB (Swiss-Swedish; robotics division founded 1988 from earlier ASEA work), KUKA (German; founded 1898, robotics from 1973), and Yaskawa (Japan; founded 1915, robotics from 1977). Combined these four vendors account for over 60% of global industrial robot installations (~3.5 million globally as of 2023, per the International Federation of Robotics). The classical robots are typically position-controlled — they execute pre-programmed trajectories with high precision and speed but without responding to the environment. The application is concentrated in automotive (welding, painting, assembly), electronics, and high-volume manufacturing generally. The deployment pattern is capital-intensive: each robot installation costs USD 50,000–300,000 depending on payload and configuration. The classical robots are not “AI” in the contemporary sense; the intelligence is in the programmer, not in the robot.

The collaborative-robotics era (2008 onward). Universal Robots (Danish; founded 2005) introduced the first commercial collaborative robot (cobot) in 2008. Cobots differ from classical industrial robots in being designed to operate safely alongside humans without safety enclosures: lower payload, slower speeds, force-limited joints that stop on contact. The economic case is different from classical robots: cobots are cheaper (USD 25,000–60,000), faster to deploy, and more flexible. The market has matured; Universal Robots was acquired by Teradyne in 2015 for USD 285 million, and the broader cobot market has reached approximately USD 2 billion in annual sales by 2024. Other notable vendors include Doosan Robotics (Korean) and various Chinese entrants.

Autonomous mobile robots (AMRs). A parallel category covers mobile robots that navigate factory or warehouse environments autonomously. Geek+ (Chinese), Locus Robotics (US; covered in Chapter 8), KUKA AMR, and many others operate in this space. The technology combines SLAM, reinforcement-learning-augmented path planning, and integration with warehouse management systems. The deployment growth has been substantial through 2018–2024.

The 2024–2026 humanoid push. A distinctive recent development is the push toward general-purpose humanoid robots in manufacturing settings. The deployment thesis: a humanoid robot can be trained to perform a much wider range of tasks than a fixed-purpose robot, with the long-run economics potentially better for low-volume or task-varied operations. Major entrants include:

• Tesla Optimus (announced 2021, prototype demonstrations 2022–2024). Tesla’s stated ambition is to deploy Optimus in Tesla’s own factories first, then commercialise. The 2024 demonstrations have been more measured than the early hype suggested.
• Figure AI (founded 2022). Figure has been the most-prominent VC-funded humanoid entrant, with USD 675 million Series B in 2024 led by Microsoft, OpenAI, NVIDIA, Jeff Bezos, and others, at USD 2.6 billion valuation. Figure has announced commercial deployments at BMW (2024) and other partners.
• 1X Technologies (Norwegian; founded 2014 as Halodi Robotics). Different positioning, emphasising consumer-and-domestic robotics alongside industrial. OpenAI is a strategic investor.
• Apptronik (US; founded 2016 from University of Texas Austin). Apptronik’s Apollo robot was unveiled in 2023 with deployment partnerships including Mercedes-Benz.
• Agility Robotics (US; founded 2015 from Oregon State University). Agility’s Digit robot has commercial deployments at Amazon and several other warehousing customers.
• Unitree, Fourier, UBTECH (Chinese). The Chinese humanoid sector has been particularly active in 2023–2025.

The 2026 state of humanoid manufacturing deployment is operational at narrow scale, experimental at broader scale. A small number of partnerships have produced documented deployments at single-digit-numbers-of-robots scale. Whether the trajectory steepens (broad deployment by 2030) or stalls is contested. The technical questions are whether the foundation-model-driven control approaches can produce reliable enough behaviour for manufacturing settings; the economic questions are whether the per-robot cost falls fast enough to compete with existing automation.

Boston Dynamics and the research-deployment gap. Boston Dynamics (founded 1992; Hyundai-owned since 2021) has been the most-prominent robotics-research firm for decades, producing technically remarkable platforms. The commercial-deployment trajectory has been more limited: Spot has achieved meaningful enterprise deployment in inspection-and-monitoring use cases; the bipedal Atlas has remained primarily a research platform until the 2024 transition to an electric Atlas with stated commercial intent. The case illustrates the gap between research demonstrations and commercial deployment: technically impressive capability does not automatically translate into deployable products without the integration, reliability, and economics that commercial users require.

9.5 Foxconn and the EMS scale story

Hon Hai Precision Industry, trading as Foxconn (Taiwanese; founded 1974 by Terry Gou; listed on Taiwan Stock Exchange), is the world’s largest electronic-manufacturing-services (EMS) contractor and Apple’s largest manufacturing partner. The company employs approximately 1 million workers globally (with peaks of 1.3+ million during high-demand periods), operates 30+ major factories across China, Taiwan, India, Mexico, Brazil, and other geographies. Foxconn’s revenue exceeded NTD 6.6 trillion (USD 220 billion) in 2024.

The Apple partnership. Foxconn’s relationship with Apple, dating to the 2007 iPhone launch, defines much of the contemporary EMS industry’s structure. iPhone and other Apple-product manufacturing accounts for approximately 50% of Foxconn revenue. The partnership has shaped Foxconn’s investment in process automation, quality control, and labour management; Apple’s quality requirements have been a forcing function for Foxconn’s manufacturing-AI deployment. The contemporary iPhone production process at Zhengzhou (the “iPhone City” facility, with peak workforce of 350,000) involves extensive automation in component placement, soldering, testing, and assembly.

The lights-out manufacturing aspiration. Terry Gou’s stated ambition since 2010 has been to move Foxconn toward “lights-out” (fully-automated, no-human) manufacturing. The company announced in 2014 a programme to deploy 1 million robots over five years; the actual deployment has been substantially below this aspiration, with approximately 60,000–80,000 robots in production by 2020 and growth to roughly 100,000+ by 2024. The aspiration has not been fully realised because (a) certain tasks remain harder to automate than the planning assumed; (b) the labour cost of Chinese manufacturing workers has not risen as fast as the cost of capital-intensive automation infrastructure; (c) the flexibility cost of full automation in a high-mix manufacturing environment is greater than fully-automated production economics suggest. The lights-out aspiration is now framed by Foxconn as a 2030+ goal rather than an imminent reality.

The labour and Chinese context. Foxconn’s labour practices have been the subject of substantial public scrutiny since the 2010 worker-suicide incidents at the Shenzhen Longhua facility. The company has invested in worker welfare improvements, but the broader pattern of Chinese manufacturing labour persists at Foxconn and across the EMS industry. The 2022 worker-protest incident at Zhengzhou (over COVID-related production-bonus and quarantine policies) and the November 2024 strike at the Cheng-Yang facility have continued this pattern. As Chinese labour costs rise and worker-relations risks accumulate, the case for automation strengthens.

The geopolitical context — diversification. Foxconn’s substantial production concentration in mainland China has become a strategic vulnerability since the 2018 onset of US-China trade-and-technology tensions. The company has been actively diversifying production geographically: India (Apple’s iPhone India production growing from 0% in 2017 to ~14% of total iPhones in 2024); Vietnam; Mexico (substantial AI server production); Brazil. The diversification reshapes the manufacturing-AI deployment pattern: each new facility represents an opportunity to deploy automation from the start rather than retrofitting an existing facility. The Indian facilities have higher automation density than the comparable Chinese facilities.

The 2024–2026 evolution. Foxconn’s contemporary AI investments span three categories. Operational AI — predictive maintenance, vision-based QA, robotics, process-control — is the largest category. AI server production — Foxconn manufactures NVIDIA H100/H200/B200 servers and similar AI infrastructure for Microsoft, Google, AWS, Meta, and others (estimated USD 30+ billion in 2024 revenue). Foxconn-as-AI-customer — the company is a substantial buyer of AI tooling. The combination positions Foxconn as both the world’s largest AI-infrastructure manufacturer and as one of the most-extensive deployers of manufacturing AI in its own operations.

9.6 Tesla’s vertical-integration approach

Tesla’s manufacturing approach contrasts with the Foxconn pattern in nearly every dimension. Where Foxconn is contract-manufacturing-at-scale-for-others, Tesla is in-house-manufacturing-of-own-product. Where Foxconn’s automation strategy is incremental, Tesla’s has been deliberately ambitious. Where Foxconn’s growth has been horizontal, Tesla’s has been vertical. The contrast is instructive for understanding the range of viable manufacturing strategies in the AI era.

The vertical-integration philosophy. Elon Musk has articulated Tesla’s manufacturing approach as “the machine that builds the machine” since approximately 2016, with the explicit ambition to make manufacturing engineering a core Tesla competency comparable to vehicle design. The premise: in conventional automotive manufacturing, vehicle designers design vehicles and manufacturing engineers figure out how to build them; Tesla’s approach is to design vehicles and the manufacturing process simultaneously. The premise is more radical than conventional in degree rather than kind; Toyota’s production system was built on similar premises in the 1950s.

The Giga press strategy. Tesla’s most-publicised manufacturing innovation is the “Giga press” — a single die-casting machine of unprecedented scale (8,000+ tons clamping force) that produces large vehicle-body components as single castings rather than as assemblies of stamped-and-welded smaller parts. The first Giga press was deployed at Fremont in 2020 for the Model Y; subsequent installations at Texas, Berlin, and Shanghai gigafactories produce the front and rear underbody components for current Tesla vehicles. The economic case: a single 80-piece subassembly is replaced by one casting; the 80 robots that previously did the subassembly are eliminated; the engineering complexity is shifted from the assembly process to the casting design. The approach has been adopted by other manufacturers (Volvo, Toyota, BYD have announced similar strategies through 2023–2024).

Manufacturing engineering as core competency. Tesla has invested heavily in in-house manufacturing engineering, including the acquisition of automation companies (Grohmann Engineering 2017; Compass Automation 2017) and the build-up of internal teams. The strategic theory is that manufacturing-engineering capability compounds — each generation of Tesla factory benefits from the lessons of prior generations. The empirical record is mixed; Tesla’s Texas and Berlin factories have had ramp-up issues that suggested manufacturing-engineering compounding has not fully delivered.

The 2018 production hell lessons. The Model 3 production ramp through 2017–2018 was the most-public manufacturing crisis in Tesla’s history. Musk publicly announced that the company would produce 5,000 Model 3 per week by end of 2017; the actual rate was approximately 800 per week. The crisis produced a series of operational decisions — including an explicit mid-stage de-automation of certain production processes — that have shaped the company’s automation strategy since. Musk publicly admitted that “humans are underrated” and that Tesla had over-automated. The lesson generalises: ambitious automation can be a deployment risk; the right level of automation depends on the maturity of both the technology and the production process; over-automation can be more costly than under-automation. The Tesla case is instructive because the over-automation was publicly acknowledged and reversed, in contrast to many other firms’ similar decisions that are not publicly disclosed.

AI in Tesla’s operations. Beyond manufacturing automation per se, Tesla deploys AI extensively in operational decisions: production scheduling, parts-inventory optimisation, quality monitoring, supply-chain forecasting. The company’s “Vision-only” approach to driver-assistance technology has structural connections to its manufacturing approach: vertical integration of capabilities, willingness to make controversial technical bets, ambitious timelines.

9.7 Boeing 737 MAX MCAS — the cautionary case in safety-critical automation

The Boeing 737 MAX case is the most-detailed contemporary example of safety-critical automation failure with documented public consequences. The case involves 346 fatalities (in two crashes), a 20-month worldwide grounding, USD 20+ billion in direct costs to Boeing, criminal prosecution, regulatory reform, and substantial reshaping of US-and-global aviation safety regulation. The case is structurally important for AI deployment thinking because the underlying failure modes — automation that operators cannot effectively override; sensor-derived inputs treated as authoritative; inadequate pilot training; certification-process gaps — generalise to many AI deployment contexts.

The MCAS context. The Boeing 737 MAX, the fourth-generation 737 with new CFM LEAP-1B engines, was Boeing’s response to the Airbus A320neo. The new engines were larger than the prior generation and had to be mounted further forward and higher on the wing than on the 737NG. The mounting change altered the aircraft’s aerodynamic behaviour, particularly at high angles of attack: the 737 MAX had a tendency to pitch up that was different from the 737NG. Boeing’s response was the Maneuvering Characteristics Augmentation System (MCAS) — an automated system that would push the aircraft’s nose down when it detected high-angle-of-attack conditions. The system was designed to make the 737 MAX behave like the 737NG from the pilot’s perspective, eliminating the need for additional pilot training. The simulator-training elimination was strategically central: airlines could deploy 737 MAX without retraining pilots, reducing transition costs.

The architectural decisions and their consequences. MCAS, as deployed in the 737 MAX flying configuration, had several architectural properties that proved consequential. First, MCAS used input from a single angle-of-attack (AoA) sensor at any given time — there was no cross-check between the two sensors on the aircraft. A single sensor failure could produce erroneous activation of MCAS. Second, MCAS could activate repeatedly without limit in response to a single AoA-sensor reading; if the pilot countered MCAS’s input via the trim wheel, MCAS could re-activate moments later. Third, the disable mechanism for MCAS was the same as for ordinary stabiliser-trim runaway, but the cognitive load of recognising and responding to MCAS-driven repeated nose-down inputs was substantially greater than the ordinary runaway-trim scenario. Fourth, pilots were not informed about MCAS in the flight-crew operating manual; the system was treated as a transparent control augmentation rather than as a system pilots needed to understand.

The two crashes. Lion Air Flight 610 (Boeing 737 MAX 8) crashed into the Java Sea on 29 October 2018, eleven minutes after takeoff from Jakarta, killing all 189 people on board. The Indonesian investigation (KNKT, 2019) determined that a faulty AoA sensor (replaced shortly before the flight) had triggered MCAS, which pushed the aircraft into a series of nose-down inputs the crew could not effectively counter; the aircraft entered an unrecoverable dive. Ethiopian Airlines Flight 302 (Boeing 737 MAX 8) crashed near Bishoftu, Ethiopia on 10 March 2019, six minutes after takeoff from Addis Ababa, killing all 157 people on board. The Ethiopian investigation reached similar conclusions: a faulty AoA reading triggered MCAS; the crew followed Boeing’s published recovery procedure but could not stabilise the aircraft.

The grounding and Congressional report. Aviation regulators worldwide grounded the 737 MAX between 11 and 18 March 2019. The grounding lasted 20 months, with the FAA approving return-to-service in November 2020 after Boeing implemented MCAS modifications (cross-check between AoA sensors; activation limits; pilot-disable improvements; pilot training requirements). The 2020 House Transportation and Infrastructure Committee report documented systematic failures: Boeing engineers’ concerns about MCAS were not communicated to senior management or to the FAA; the FAA’s Organization Designation Authorization (ODA) program — which delegates substantial certification authority to manufacturers — had produced inadequate scrutiny; cost and schedule pressures had compromised safety review.

The legal and regulatory aftermath. Boeing entered a deferred prosecution agreement (DPA) with the US Department of Justice in January 2021, paying USD 2.5 billion in penalties and restitution. The DPA was substantially modified in May 2024 after the January 2024 Alaska Airlines Flight 1282 incident (a door plug detached at altitude on a 737 MAX 9; no fatalities but exposed continuing manufacturing-quality issues at Boeing). The Justice Department determined Boeing had violated the DPA terms; Boeing entered a guilty plea agreement in July 2024 to felony charges. The settlement is in continued litigation as of 2026; the eventual disposition will further shape Boeing’s regulatory and corporate-governance posture.

Structural lessons. The 737 MAX case yields five lessons that generalise to safety-critical automation broadly.

Lesson 1 — sensor single-points-of-failure are unacceptable in safety-critical systems. MCAS’s reliance on a single AoA sensor without cross-check was a foundational design error. Redundancy and cross-validation are standard in safety-critical engineering; their absence in MCAS reflects either an oversight or a deliberate cost-driven omission. The lesson generalises directly to AI systems: an AI system whose outputs drive safety-critical action must use redundant inputs and cross-check against other systems.

Lesson 2 — automation that operators cannot override is dangerous. MCAS could re-activate repeatedly after pilot input, against the pilot’s intent. The pattern violated the principle that operators must remain the final authority in safety-critical contexts. Modern AI systems often face similar architectural questions; the principle that human operators can override the AI must be designed in, not assumed.

Lesson 3 — training-cost-saving is a brittle design driver. Boeing’s strategic rationale for MCAS was to eliminate the need for simulator training in the 737 MAX transition. The cost-saving became a forcing function that pushed against the safer architectural choice. When safety considerations conflict with cost-optimisation considerations, safety must dominate; when an architectural decision exists primarily to avoid a training cost, the architectural decision is suspect.

Lesson 4 — certification process can be captured. The FAA’s ODA programme delegated substantial certification authority to Boeing; the delegated review did not catch the MCAS issues that an independent regulator would arguably have caught. The structural lesson generalises beyond aviation: regulatory frameworks that depend on the regulated party’s good-faith disclosure are vulnerable to capture when the regulated party faces strong commercial pressures. The EU AI Act’s structure must be designed with this dynamic in mind.

Lesson 5 — the post-incident management is itself a failure mode. Boeing’s response to the Lion Air crash, in late 2018 and into early 2019, was to defend MCAS rather than to acknowledge problems. The same defensive posture continued through the Ethiopian crash and into the early grounding period. The defensive posture extended the brand damage and the legal exposure substantially; an earlier acknowledgement and ground initiative would have produced a less-costly resolution. The pattern recurs in AI deployment.

The 737 MAX case has become a standard reference in safety-critical-automation engineering and in AI-deployment ethics literature. The integration of the case into AI-deployment thinking is not a strained analogy: AI systems that drive safety-critical decisions face the same architectural questions as MCAS faced, and the failure modes are recognisably similar. The Watson Health case (Chapter 7) and the Klarna case (Chapter 8) show the broad-system pattern; the 737 MAX case shows the safety-critical pattern. Together they constitute the primary cautionary record for contemporary AI deployment.

9.8 Generative design and digital product development

Generative design is the application of algorithmic optimisation to product design. Given a set of constraints (load requirements; manufacturing constraints; material availability; cost targets), a generative-design algorithm produces design candidates that optimise some objective function. The technology has matured over a 30-year period from academic optimisation methods through commercial CAD integration to the contemporary foundation-model-augmented generative design.

Topology optimisation — the foundation. The methodological foundation is topology optimisation — a class of optimisation methods (Bendsøe and Sigmund, 2003) that solve for the optimal distribution of material within a given design domain, subject to load and constraint conditions. The classical method is the SIMP (Solid Isotropic Material with Penalisation) approach. The methods produce designs that often look organic — strut-and-truss structures with material concentrated where loads concentrate — and that human designers would not naturally produce. The classical methods have been integrated into CAD software and used in production contexts since the 2000s.

The Autodesk era. Autodesk’s Generative Design product, integrated into Fusion 360 from 2018 onward, brought topology-optimisation-style functionality to mainstream CAD users. The product allows designers to specify constraints (loads, fixtures, materials, manufacturing methods) and have the software produce design alternatives. The deployment scale has been substantial, with use across automotive, aerospace, consumer products, and architecture.

Airbus A320 cabin partition — the canonical case. The most-cited generative-design deployment is Airbus’s redesign of the cabin partition in the A320 aircraft. The original partition was a stamped-aluminium component; the generative-design replacement (developed by Autodesk and Airbus’s APWorks subsidiary, deployed from 2016) is a 3D-printed component using Scalmalloy alloy. The redesigned partition is approximately 45% lighter than the original (around 30 kg saved per partition; 40 partitions per aircraft; 4,000+ A320 family aircraft in service representing several thousand tons of cumulative weight savings) with equivalent or better strength characteristics. The case demonstrates the integration of generative design with additive manufacturing — without 3D printing, the optimised geometry could not be manufactured.

Siemens NX and PLM integration. Siemens NX has integrated topology-optimisation and increasingly generative-design capabilities into its broader product lifecycle management (PLM) suite. The integration matters because real-world generative design is rarely standalone; it sits within a broader product-development workflow that includes simulation, manufacturing process specification, supply-chain alignment, and quality verification.

The 2024–2026 generative-AI extension. The contemporary wave is extending generative design into new modalities. Text-to-CAD interfaces have appeared in 2024 from multiple startups. Foundation-model-based design recommendation is in early commercial deployment. The integration with simulation is the deeper opportunity, though simulation accuracy remains a binding constraint for many applications.

Deployment realism check. Generative design in 2026 is operationally deployed for specific use cases (lightweight components in aerospace and automotive; complex consumer products) but not yet broadly embedded in routine product development. The constraints include simulation accuracy, manufacturing-method limitations, and the cultural-organisational shift required for design teams to incorporate generative outputs into established workflows.

9.9 Digital twins

A digital twin is a virtual representation of a physical system that updates in real time from sensor data, can simulate the system’s behaviour under different conditions, and supports decisions about the system. The concept has roots in NASA’s pre-flight simulation practice (Glaessgen and Stargel, 2012); the contemporary commercialisation began approximately 2010 and matured through the late 2010s.

The methodological structure. A digital twin combines four elements: (1) the physical system being modelled; (2) the sensor instrumentation that produces real-time data on the physical system’s state; (3) the virtual representation (a simulation model, often combining physics-based models with machine-learning-fitted components); (4) the analytics and decision layer that uses the virtual representation. The twin is “live” — the virtual representation continuously updates from sensor data — which distinguishes it from a static simulation.

GE Aviation’s twin programme. GE Aviation operates digital twins of its commercial jet engines in service, with each engine having a corresponding twin updated from operational data. The twin enables per-engine performance prediction, anomaly detection, and lifecycle management. The deployment scale is substantial; with 30,000+ engines in service, the twin programme is among the largest digital-twin deployments globally.

Siemens NX digital twin. Siemens has integrated digital-twin capability into its broader PLM and industrial-software portfolio. The 2017 acquisition of Mendix and subsequent integrations have positioned Siemens as a major digital-twin platform vendor, with deployment customers across automotive, aerospace, and discrete manufacturing.

Rolls-Royce engine twin. Rolls-Royce’s Trent engine series operates under power-by-the-hour contracts (TotalCare programme), where Rolls-Royce is paid per engine flying hour rather than per engine sold. The contract structure aligns Rolls-Royce’s incentive directly with engine reliability and efficiency, which makes per-engine digital twin investment economically sensible.

The deployment realism check. Digital twins have produced strong deployment results at the unit-level (single engines; single turbines; single major equipment) but more variable results at the factory-level and city-level. Factory-level twin deployments have been operationally useful but have not produced the order-of-magnitude transformations that the broader-Industry-4.0 framing claimed. The city-level digital twins (Singapore’s Virtual Singapore; various smart-city programmes) have produced limited operational deployment despite substantial public investment. The pattern suggests that twin scale-up is bounded by the simulation-accuracy challenge: as the system grows in complexity, the twin’s simulation accuracy degrades. Unit-level twins are the deployment sweet spot; broader-system twins remain operationally challenging.

9.10 Semiconductor manufacturing — TSMC and the cluster

The semiconductor industry is the industrial domain where manufacturing precision is most extreme and where the AI-and-manufacturing intersection is most consequential. Modern semiconductor manufacturing involves processes operating at nanometre scales, with single-defect tolerances measured in parts-per-billion, in clean-room environments at scales approaching billions of dollars per fab. The industry’s deployment of AI is correspondingly extensive.

TSMC — the world’s most-sophisticated manufacturer. Taiwan Semiconductor Manufacturing Company (founded 1987; listed on TWSE and NYSE) is the world’s largest dedicated semiconductor foundry, with approximately 60% of the global pure-play foundry market. TSMC’s leading-edge nodes (3nm in mass production from 2022; 2nm planned for 2025; advance research at 1.4nm and below) define the semiconductor industry’s frontier. The company’s revenue exceeded TWD 2.5 trillion (USD 80 billion) in 2024 with operating margin around 40% — among the highest in any large-scale manufacturing operation globally.

Process control and AI. Semiconductor manufacturing requires controlling thousands of process variables (temperature, pressure, gas flow, exposure time, alignment) within tolerance bands that human-controlled processes could not achieve. The AI deployment in semiconductor manufacturing is comprehensive: predictive maintenance for the equipment; statistical process control with ML-driven anomaly detection; defect classification; yield prediction; supply-chain optimisation across the multi-month production cycle. The integration is not always visible from outside; the industry’s competitive secrecy around manufacturing know-how means that public documentation is limited. What is visible is the productivity gap between leading-edge and trailing-edge manufacturers — at any given process node, leading manufacturers achieve substantially higher yields than trailing manufacturers.

ASML lithography systems. ASML (Dutch; founded 1984) makes the lithography equipment that defines the leading edge of semiconductor manufacturing. ASML’s EUV (Extreme Ultraviolet) systems, operational from 2018 at 7nm and below process nodes, are the most-complex production machines ever built — each EUV machine costs approximately USD 200 million, contains approximately 100,000 components, and requires a several-day installation. ASML’s effective monopoly on EUV lithography makes it strategically central to the semiconductor industry. ASML’s own use of AI is substantial: machine learning is integrated into the lithography systems’ control software, the metrology systems that characterise process outcomes, and the field-service operations that maintain deployed equipment.

The cluster — fab and backend. Semiconductor manufacturing splits into front-end (fabrication; the wafer-level process) and back-end (assembly, test, and packaging; the die-level process). The front-end is concentrated geographically: TSMC and UMC in Taiwan; Samsung and SK Hynix in South Korea; Intel in the US (with substantial foundry investment underway); SMIC in China. The back-end is more geographically distributed, with a substantial share in Southeast Asia — particularly Malaysia (Penang and Kulim) and the Philippines.

The geopolitical context. US-China tensions have produced substantial reshaping of the semiconductor industry’s geographic structure. The 2022 CHIPS and Science Act in the US (USD 280 billion in semiconductor manufacturing and R&D incentives) and parallel programmes in the EU, Japan, India, and elsewhere have driven substantial new investment in geographies outside Taiwan and South Korea. TSMC’s Arizona facility (USD 65 billion total committed investment by 2024) and Japan facility (USD 8.6 billion) are direct responses to this dynamic. The 2024 export-control restrictions on advanced semiconductor equipment to China have accelerated the bifurcation. The implications for manufacturing-AI deployment are mixed: new facilities being built in geographies without existing semiconductor expertise face a steeper deployment learning curve; existing-cluster facilities benefit from accumulated process know-how; the overall industry deployment cost rises.

9.11 Australian mining and resources automation

The Australian mining-and-resources sector is the deepest deployment of operational autonomy in heavy industry globally. The combination of the country’s resource-export concentration (iron ore, coal, gold, lithium, copper, rare earths), the geographic remoteness of major mining operations, and the high labour cost in Australian operations has produced strong economic incentives for automation that the major mining companies have pursued aggressively.

Rio Tinto’s Mine of the Future. Rio Tinto launched the Mine of the Future programme in 2008, with the strategic aim of progressively automating Pilbara iron-ore operations. The programme’s deployment scale is substantial: by 2024, Rio Tinto operated the world’s largest fleet of autonomous haulage trucks (over 250 Komatsu 980E and 930E autonomous trucks across multiple mines), the world’s first fully-autonomous heavy-haul rail (the AutoHaul system, operational from 2019, moves over 200 trains carrying iron ore from inland mines to coastal ports across the Pilbara), autonomous drill systems, and remote-operations centres in Perth that manage Pilbara operations 1,500+ km away. The cumulative investment in the programme exceeds AUD 5 billion. Documented productivity gains are substantial — autonomous trucks operate approximately 20% more hours per day than human-operated trucks (no shift changes; no break requirements), with comparable productivity per operating hour, producing approximately 25% effective productivity gain.

BHP automation. BHP, Rio Tinto’s main competitor in Australian iron ore, has pursued similar though less-publicised automation. BHP’s Spence copper mine in Chile has been a significant test bed for autonomous mining technology; the Pilbara iron-ore operations have substantial autonomous-truck deployment though not at the Rio Tinto scale; the South Flank iron-ore mine (commissioned 2021) was designed for autonomous operation from inception. BHP’s Olympic Dam copper-uranium operation in South Australia has been the subject of substantial digital-transformation investment over 2018–2024.

Fortescue Metals Group automation. Fortescue (founded 2003; Andrew Forrest as founder), Australia’s third major iron-ore producer, operates entirely automated truck fleets at multiple Pilbara mines. Fortescue’s automation deployment has been particularly aggressive given the company’s later entry into the iron-ore market — the company has been able to deploy automation from inception rather than retrofitting existing operations. Fortescue’s diversification into green hydrogen (via the Fortescue Future Industries subsidiary) extends the automation thesis into adjacent energy infrastructure.

The Pilbara remote-operations centres. The Rio Tinto, BHP, and Fortescue Pilbara operations are largely controlled from operations centres in Perth, with mine-site personnel reduced substantially. The remote-operations approach was made feasible by combining (a) high-bandwidth communications to remote mining sites; (b) operational technology (SCADA systems; truck-positioning systems; sensor networks); (c) AI-driven decision support for the operations-centre staff; (d) cultural and procedural changes that allowed operations-centre staff to act on remote authority.

The methodological transfer. The Australian mining-automation experience has informed similar programmes in adjacent industries: the offshore oil-and-gas sector (where remote operations have been standard for decades but autonomous-equipment deployment has accelerated through 2018–2024); the Australian agricultural sector (where autonomous tractors and harvesters have been deployed at scale); construction equipment (where autonomous earth-moving equipment is now operational at large infrastructure projects). The methodology generalises across heavy-industry contexts.

Workforce implications. The labour implications of Australian mining automation are substantial. Estimates by the Australian Resources and Energy Group (AMMA) suggest that mining automation has reduced direct mining-employment by 15–25% across automated operations relative to non-automated baselines, with corresponding increases in higher-skilled remote-operations and maintenance roles. The transition has produced political and union responses; the Mining and Energy Union has been negotiating workforce-transition arrangements with major operators. The pattern is a microcosm of broader AI-driven labour transitions: net employment effects are smaller than headline displacement numbers suggest, but the displacement is concentrated in specific job categories and geographies, producing concentrated distributional effects.

9.12 Malaysian E&E and Southeast Asian manufacturing

Southeast Asia hosts substantial manufacturing capacity, much of it AI-relevant for both the deployment of AI in the manufacturing operations themselves and for the production of AI infrastructure (servers, GPUs, networking equipment). The regional context differs from the US, EU, and Chinese patterns that dominate the AI-and-manufacturing literature.

The Penang cluster. Penang’s electronics-and-electrical (E&E) cluster originated with the establishment of the Bayan Lepas Free Industrial Zone in 1972 and the early decisions of Intel, Hewlett-Packard, and several Japanese firms to locate semiconductor backend operations there. The cluster has grown over five decades into one of Southeast Asia’s most-significant semiconductor and electronics manufacturing concentrations. By 2024, Penang hosted approximately 350 multinational E&E companies and over 3,000 supporting domestic firms, generating around 32% of Malaysia’s manufacturing output and employing over 250,000 workers across the broader cluster (Penang state government estimates; InvestPenang annual reports). The cluster’s particular strength is in semiconductor backend operations (assembly, test, packaging), with growing capacity in advanced packaging (chiplet integration, advanced substrate technologies) that aligns with the contemporary semiconductor industry’s shift away from pure node-shrink toward heterogeneous integration. The 2022–2026 capacity expansion has been substantial: Intel announced USD 7 billion of new investment in 2021 (its largest single overseas investment, with construction underway through 2024–2026); Infineon announced EUR 5 billion of expansion at Kulim through 2030; Texas Instruments announced new Malacca capacity; AMD, Bosch, and several others have made comparable announcements.

ViTrox Corporation. Already introduced in Section 9.3. As the deepest Malaysian-domiciled manufacturing-AI deep dive case, ViTrox’s operational and strategic profile bears closer treatment for the Malaysian cluster context. The company was founded in 2000 by Chu Jenn Weng and Steven Siaw Kok Tong, two engineers from the Universiti Sains Malaysia, with initial seed funding from the Penang state government’s investment arm. The early years focused on board-level inspection systems (3D solder paste inspection; automated optical inspection of populated PCBs) for the regional electronics manufacturing services sector. The 2010–2015 period saw the company expand into x-ray inspection systems for ball-grid-array (BGA) and complex-package devices; the 2015–2020 period saw expansion into IC inspection for semiconductor backend operations. The 2020s have seen the company extend into advanced packaging inspection (the technology that aligns with the broader semiconductor industry’s shift toward chiplet integration). ViTrox listed on Bursa Malaysia in 2009 and has been consistently profitable since 2011; the company maintains R&D centres in Penang and Suzhou, with sales offices across major regional electronics-manufacturing markets. The case is a useful reminder that frontier manufacturing AI is not exclusively a US-or-China phenomenon: a Malaysian-founded, Malaysian-listed, Malaysian-headquartered firm with Malaysian engineering leadership operates at the global frontier of automated optical inspection technology.

Inari Amertron. Inari (KL-listed; founded 2006 from the consolidation of an earlier semiconductor backend operation) is Malaysia’s largest pure-play semiconductor backend services provider, with over 16,000 employees and revenue around RM 1.8 billion in 2023. The company’s primary customers are Broadcom (RF filters and other components for mobile applications) and several other tier-1 semiconductor firms. The deployment of manufacturing AI at Inari is substantial: the company operates extensive AOI infrastructure (including ViTrox-supplied equipment), automated testing operations, predictive maintenance for the assembly lines, and quality-control analytics that integrate sensor data across the production process. Inari’s own AI capability is more operational than R&D — the company is a deployer rather than a producer of manufacturing AI — but the deployment depth makes it one of the more-instrumented backend semiconductor operations in the region.

Pentamaster. Pentamaster (KL-listed; founded 1991) is a Malaysian factory-automation provider, supplying robotic systems, vision systems, and automated test equipment to semiconductor and electronics customers across Asia. The company’s revenue around RM 600 million in 2023 places it in the upper tier of Malaysian-domiciled industrial-automation firms. Pentamaster’s products span the full spectrum from low-cost automation (suitable for SE Asian manufacturing economics) to high-end semiconductor test equipment (competitive with US and Japanese vendors at the mid-market level). The company’s product-and-service mix is roughly evenly split between Malaysian customers and exports to other Asian markets.

Malaysian automotive — Proton and Perodua. Malaysia’s automotive sector consists primarily of two domestic manufacturers (Proton and Perodua) plus the Malaysian operations of major international assemblers (Toyota, Honda, Volkswagen, BMW, Mercedes-Benz). Proton (founded 1983; Geely as 49.9% strategic shareholder since 2017) and Perodua (founded 1993; Daihatsu as technology partner) together produce around 600,000 vehicles annually, primarily for the Malaysian and ASEAN markets. The AI deployment at Proton and Perodua is generally less advanced than at the leading global automakers; the deployment is concentrated in operational areas (predictive maintenance for production equipment; quality control vision systems; supply-chain forecasting) rather than in advanced applications (digital twins; generative design; autonomous-driving R&D). The Geely partnership at Proton has accelerated some areas of technology transfer; the broader Malaysian automotive sector remains structurally a step behind the leading global frontier.

Vietnam. Vietnam’s manufacturing sector has grown rapidly through 2018–2024, particularly in electronics assembly (Samsung’s substantial Vietnamese operations; LG; an expanding base of Chinese-supply-chain-relocation operations), in textiles, and in increasingly higher-value industrial categories. Vietnamese manufacturing-AI deployment is generally comparable to or slightly behind Malaysian deployment; the country’s strong technical-education base (particularly in mathematics and physics) supports a growing local engineering workforce. The 2023–2024 wave of US foundry investment in Vietnam (Intel’s continued expansion; Synopsys engineering centre; multiple semiconductor backend investments) has been substantial.

Thailand. Thailand’s manufacturing sector includes substantial automotive (Bangkok cluster as a major regional automotive-export hub), electronics (centred around Bangkok and Eastern Economic Corridor), and food processing. Thai manufacturing-AI deployment has been somewhat slower than Malaysian or Vietnamese deployment, partly reflecting the country’s industrial structure (more automotive-and-food-processing; less electronics-and-semiconductor) and partly reflecting policy emphasis. The Eastern Economic Corridor initiative (announced 2017, ongoing) is the country’s specific framework for upgrading manufacturing capability, including AI-and-automation components.

The regional supply-chain context. The Southeast Asian manufacturing context cannot be considered in isolation from the broader regional supply chain. The China+1 strategy adopted by many global manufacturers (diversifying production from China to ASEAN locations as a hedge against US-China tensions) has been a substantial driver of regional capacity growth through 2018–2026. Vietnam, Malaysia, and Thailand have all been beneficiaries; Indonesia and the Philippines somewhat less so. The 2024–2026 wave of generative-AI-driven manufacturing investment (AI server production; advanced packaging; substrate manufacturing) has further extended the regional opportunity. The Malaysian government’s National Semiconductor Strategy (announced 2024) is the country’s specific policy framework for capturing this wave, with explicit targeting of the integrated-circuit-design segment alongside the existing backend manufacturing strength.

9.13 The 2026 frontier and structural questions

The manufacturing-and-Industry-4.0 landscape in 2026 sits at an inflection point. The first three decades of computer-and-automation integration have produced documented productivity gains and cautionary failures. The contemporary AI wave is layering additional capability on top of the existing infrastructure. Five trajectories define the path ahead.

Trajectory 1 — humanoid robots in manufacturing. Section 9.4 covered the contemporary humanoid push. The 2026 question is whether the trajectory steepens or stalls. The technical capability has improved substantially through 2023–2025 (foundation-model-based control; affordable hardware platforms; better grasping and manipulation). The economic case at narrow scale is becoming demonstrable (the Figure-BMW partnership, the Apptronik-Mercedes partnership, and several others have produced operational deployments). The broader-scale economic case — humanoid robots replacing significant fractions of manufacturing labour — is more contested. The next 24 months will be informative. If humanoid deployments scale beyond the current narrow demonstrations into routine factory operations across multiple firms and industries, the trajectory is steeper than scepticism implied; if the deployments remain narrow despite further capability improvement, the trajectory is more constrained than enthusiasm implied.

Trajectory 2 — the labour displacement question. The labour implications of manufacturing AI are central to the political-and-economic conversation about the 2024–2030 trajectory. The empirical record is mixed: Acemoglu and Restrepo (2020) document that automation has produced labour displacement without proportional productivity gains in some manufacturing contexts; Autor, Salomons, and Seegmiller (2022) find that automation effects are heterogeneous across worker categories and geographies; the Australian mining case (Section 9.11) shows displacement concentrated in specific job categories. The policy responses vary substantially across jurisdictions. The 2024 EU AI Act has worker-information-and-consultation requirements that affect AI deployment in manufacturing settings; the US response has been more market-driven; the Chinese response has been state-directed. The pattern through 2026–2030 will partly depend on whether policy frameworks evolve to cushion labour-market transitions or whether displacement is allowed to compound.

Trajectory 3 — the energy intensity of manufacturing AI. A specific concern that has emerged through 2024–2026 is the energy intensity of manufacturing AI. Foundation-model inference and training are substantial energy consumers; manufacturing-AI deployments add to this load. The interaction with manufacturing’s own energy intensity (steel; aluminium; cement; chemicals; semiconductor fabrication) produces an aggregate energy-and-carbon profile that must be managed. The 2024 IEA AI and Energy analysis projected that data-centre electricity consumption could double by 2030, with manufacturing-AI deployment as one driver. The implications for industrial-policy-and-climate-policy alignment are substantial. The Malaysian National Energy Transition Roadmap (2023) and equivalent Australian, Singaporean, and other regional frameworks are responding to this dynamic.

Trajectory 4 — supply-chain resilience post-COVID, post-Ukraine, post-Taiwan-tension. The 2020–2025 period has been the most-disruptive period for global manufacturing supply chains since at least the 1970s oil shocks. COVID-driven disruptions in 2020–2022, the Russia-Ukraine war and consequent commodity disruptions from 2022, and the rising Taiwan-Strait tensions have all forced major manufacturers to reconsider supply-chain concentration and resilience. The “China-plus-one” strategy already mentioned, the broader “friend-shoring” push articulated by US Treasury Secretary Janet Yellen in 2022 and continued under subsequent administrations, the EU’s strategic-autonomy framework — all push toward more-distributed manufacturing footprints. The implications for AI deployment are double-edged: distributed manufacturing requires more AI-driven coordination across geographies; the each-facility-scale of operations may be smaller, reducing the per-facility AI-investment economics. The net effect on AI deployment intensity per unit of output is unclear; it is plausibly higher (because coordination requires deployment) or lower (because per-facility scale is smaller).

Trajectory 5 — the generative-AI extension into product development. The contemporary generative-AI wave is most visible in consumer-and-services applications. The extension into manufacturing product development — generative design (Section 9.8); foundation-model-augmented engineering tools; AI-supported simulation; AI-driven supply-chain decisions — is at an earlier stage. The trajectory through 2026–2030 will see substantial maturation: generative design will move from research-curiosity to routine engineering practice in specific contexts (aerospace; automotive lightweighting); foundation-model-augmented PLM systems will deploy at scale; AI-driven simulation will compress R&D cycles for many product categories. The competitive implication: firms that integrate the generative-AI capability into engineering practice early will compound advantages relative to firms that lag. This is the manufacturing analog of the AI-factory thesis from Chapter 3.

The manufacturing-and-Industry-4.0 landscape has been the longest-standing AI deployment domain. The 2024–2026 period represents a meaningful inflection — generative AI extending the operational deployment, humanoid robotics opening new modalities, the regulatory framework catching up to the practice. The structural lessons from Boeing 737 MAX, Tesla, Foxconn, Rio Tinto, ViTrox, and the broader cases provide the most-detailed contemporary case material for the manufacturing-deployment discipline. The integration of those lessons with the analytical frameworks of Parts I and IV is what produces graduate-level competence in the field.

References for this chapter

Industry 4.0 framework

• Kagermann, H., Wahlster, W., and Helbig, J. (eds.) (2013). Recommendations for implementing the strategic initiative INDUSTRIE 4.0. Acatech.
• Schwab, K. (2016). The Fourth Industrial Revolution. World Economic Forum / Crown.
• Acemoglu, D. and Restrepo, P. (2020). Robots and jobs: Evidence from US labor markets. Journal of Political Economy 128(6): 2188–2244.
• Autor, D., Salomons, A., and Seegmiller, B. (2022). New frontiers: The origins and content of new work, 1940–2018. NBER Working Paper.

Predictive maintenance

• Lee, J., Bagheri, B., and Kao, H.-A. (2017). Recent advances and trends of cyber-physical systems and big data analytics in industrial informatics. IEEE Transactions on Industrial Informatics.
• Susto, G. A., Schirru, A., Pampuri, S., et al. (2015). Machine learning for predictive maintenance: A multiple classifier approach. IEEE Transactions on Industrial Informatics 11(3): 812–820.
• Carvalho, T. P., Soares, F. A. A. M. N., Vita, R., et al. (2019). A systematic literature review of machine learning methods applied to predictive maintenance. Computers & Industrial Engineering 137.
• Iansiti, M. and Lakhani, K. R. (2020). Competing in the Age of AI. Harvard Business Review Press.

Computer vision and AOI

• LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11): 2278–2324.
• Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. NIPS.
• He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep residual learning for image recognition. CVPR.
• ViTrox Corporation Berhad (2009–2024). Annual reports. Bursa Malaysia.

Robotics

• International Federation of Robotics (2024). World Robotics Report 2024.
• Universal Robots (2008–2024). Product documentation and corporate disclosures.
• Boston Consulting Group (2024). Humanoid robots in manufacturing: Status and outlook.
• Figure AI (2024). Series B announcements and BMW partnership disclosures.

Foxconn

• Hon Hai Precision Industry Co. Ltd. (2024). Annual report and investor materials.
• Reuters and Bloomberg (2010–2024). Coverage of Foxconn labour and operations.
• Apple Inc. (2024). Supplier Responsibility Progress Report.

Tesla

• Tesla Inc. (2017–2024). Annual reports (Form 10-K) and shareholder communications.
• Musk, E. (2018). Public statements on Model 3 production and de-automation.
• Reed, A. (2018). The Model 3 production hell: A retrospective. Bloomberg.

Boeing 737 MAX

• US House Transportation and Infrastructure Committee (2020). The Boeing 737 MAX: Costs, Consequences, and Lessons from Its Design, Development, and Certification. Final Report, September 2020.
• KNKT (Indonesian National Transportation Safety Committee) (2019). Aircraft Accident Investigation Report: PT Lion Airlines Boeing 737-8 (MAX); PK-LQP; Tanjung Karawang, West Java; Republic of Indonesia; 29 October 2018.
• Ethiopian Civil Aviation Authority (2019, 2022). Aircraft Accident Investigation Bureau interim and final reports on Ethiopian Airlines Flight 302.
• US Department of Justice (2021). Deferred Prosecution Agreement with The Boeing Company.
• US Department of Justice (2024). Plea agreement filings related to The Boeing Company.

Generative design and digital twins

• Bendsøe, M. P. and Sigmund, O. (2003). Topology Optimization: Theory, Methods, and Applications. Springer.
• Glaessgen, E. H. and Stargel, D. S. (2012). The digital twin paradigm for future NASA and U.S. Air Force vehicles. AIAA SDM Conference.
• Autodesk (2018–2024). Generative Design product documentation.
• Airbus / APWorks (2016, 2017). A320 cabin partition redesign technical documentation.

Semiconductor industry

• Taiwan Semiconductor Manufacturing Company (2024). Annual report.
• ASML Holding NV (2024). Annual report.
• Boston Consulting Group (2024). Semiconductor industry productivity benchmark.
• US Department of Commerce (2022, 2024). CHIPS and Science Act implementation reports.

Australian mining

• Rio Tinto (2008–2024). Mine of the Future programme reports and annual reports.
• BHP (2024). Annual report and operations reviews.
• Fortescue Metals Group (2024). Annual report.
• Australian Resources and Energy Group (AMMA) (2024). Workforce transformation in resources sector.

Malaysian and Southeast Asian context

• InvestPenang (2024). Penang state economic profile and E&E sector report.
• Inari Amertron Berhad (2024). Annual report.
• Pentamaster Corporation Berhad (2024). Annual report.
• Malaysia Investment Development Authority (MIDA) (2024). National Semiconductor Strategy.
• Bain & Company (2024). Southeast Asia manufacturing outlook.

The 2026 frontier

• International Energy Agency (2024). AI and Energy outlook.
• World Economic Forum (2024). Future of Manufacturing report.
• Yellen, J. (2022, 2024). Public statements on friend-shoring and supply-chain resilience.
• European Commission (2024). Strategic Autonomy framework documents.