From Fabs to Factories: What General-Purpose Robotic Manufacturing Can Learn from Semiconductors

Quick summary

Having spent some time working in photonics, I have found myself returning to the semiconductor industry as a useful model for thinking about the future structure of robotics. The most interesting lesson is not that semiconductors became a large industry. It is that, as the underlying technology matured, value spread into a dense ecosystem of specialized companies spanning equipment, metrology, design software, process interfaces, packaging, test, standards, and asset lifecycle management. Robotics already has fragments of an analogous stack, but the transition toward more general learned manipulation could create new layers whose job is to make probabilistic robotic production measurable, designable, and eventually interchangeable. The analogy is imperfect, particularly because learned robot policies introduce a form of behavioral uncertainty that conventional industrial automation and semiconductor process recipes do not have, but that imperfection may be precisely where some of the most interesting opportunities lie.

Why the semiconductor ecosystem is worth studying

This topic has been on my mind recently in part because I previously spent some time working in photonics. Photonics sits close enough to semiconductor fabrication that one quickly becomes aware of how much machinery surrounds the nominal act of “making a chip.”

A modern chip might be designed by a fabless company, implemented using software from an electronic-design-automation vendor, checked against a foundry-specific process design kit, manufactured using equipment supplied by firms such as ASML, Applied Materials and Lam Research, repeatedly inspected using process-control equipment from firms such as KLA, and ultimately packaged and tested by an outsourced semiconductor assembly and test provider. Standards bodies define interfaces among equipment, factories and software, and an entire separate market exists for the refurbishment and resale of older equipment.

Specialization has undoubtedly become one of the defining characteristics of the semiconductor industry.

The foundry/fabless split is likely the most familiar example. TSMC, founded in 1987, describes itself as the pioneer of the pure-play foundry model: it manufactures products designed by customers rather than competing with those customers by selling its own chips. In 2025, TSMC says it manufactured 12,682 products across 305 technologies for 534 customers, which illustrates the scale at which a common manufacturing platform can serve heterogeneous designers. [1]

To interface with designers, a process design kit, or PDK, packages foundry-specific information into forms that electronic design tools can consume. TSMC provides design rules, device models, and technology files for layout-versus-schematic verification, parasitic extraction, and automatic place-and-route. Its Open Innovation Platform (OIP) connects this foundry-specific information to third-party tools covering circuit design, simulation, physical implementation and sign-off. [2]

EDA itself became an independent industry because chip complexity made manual implementation increasingly untenable. Synopsys was founded in 1986 around logic synthesis; the company describes synthesis as a transition from computer-aided design toward increasingly automated electronic design automation. Cadence emerged in 1988 from the combination of earlier EDA companies, including ECAD, which commercialized design-rule checking. [3, 4]

Manufacturing equipment followed its own specialization trajectory. Applied Materials began in 1967 and subsequently became a major supplier of wafer-fabrication equipment; Lam Research, founded in 1980, specializes in process technologies including etching and deposition. ASML was established in 1984 to commercialize wafer-stepper technology and today supplies lithography systems central to patterning semiconductor layers. Its EUV systems are the result of decades of specialized optical, plasma, mechatronic and systems engineering. [5–7]

Alongside the tools that perform fabrication grew another category of tools whose purpose is to determine whether fabrication is working. KLA traces its process-control history to a mask-inspection instrument introduced in 1975, and today its businesses span defect inspection, metrology and yield-management systems. [8, 9] This distinction between performing a process and measuring a process becomes important for the robotics analogy.

Eventually, secondary equipment markets grew into a significant business in their own right – once enough expensive equipment exists across multiple technology generations, the residual value of older machines can often be repurposed to service applications with less stringent process requirements. Consider for instance SurplusGLOBAL, founded in 2000, which has supplied more than 40,000 pieces of legacy semiconductor equipment. [10]

In general, the technological maturation of semiconductor manufacturing has produced an expanding set of specialized firms, interfaces, measurements, and abstractions around various aspects of the ecosystem.

What might the analogous robotics stack look like?

The analogy should not be interpreted too literally, but it still may be useful to ask what function each semiconductor layer performs in the industrial system, and then ask whether general-purpose robotic manufacturing eventually requires an equivalent function.

One plausible robotic counterpart to the foundry is a flexible robotic manufacturing provider. A product company could eventually specify a product and production requirements without owning all of the automation engineering needed to build it. The robotic manufacturer would expose its capabilities in a sufficiently standardized form that products could increasingly be designed against them. This resembles the economic role of a semiconductor foundry more than the physical operation of one.

That, in turn, suggests a possible analogue to a PDK. A robotic process or capability kit might describe reachable process classes, end-effectors, tolerances, material constraints, geometric accessibility and empirically measured reliability. Unlike a semiconductor rule deck, many of its outputs might be probabilistic. The important concept is an interface between what a designer wants manufactured and what a manufacturing system can reliably execute.

Above this interface sits a possible robotic version of EDA. Traditional industrial software already implements substantial parts of this idea. Siemens Process Simulate supports workcell layout, kinematic definitions, robotic path development, simulation and virtual commissioning; Dassault Systèmes’ DELMIA links product design, process planning, robotic simulation and offline programming. [11, 12] NVIDIA is building adjacent infrastructure for factory-scale digital twins in which multi-robot fleets can be simulated before deployment. [13]

Thus, the open question is not whether “EDA for robots” exists – much of it already does for conventional automation. The question is what changes when the manufacturing primitive moves from a classically programmed trajectory toward a learned policy expected to generalize across parts, embodiments, and tasks.

Several other correspondences follow naturally. Semiconductor process equipment suggests specialized suppliers of robots, end-effectors, fixtures and joining machinery rather than a single vertically integrated robot vendor. Advanced packaging suggests increasingly modular products with self-aligning, machine-friendly interfaces, perhaps supported by reusable carriers or other “physical interposers” that regularize how irregular objects meet robots. Semiconductor refurbishment suggests a future market for robot health characterization, remanufacturing and residual-value assessment. The infrastructure of a fab suggests that some industrial buildings may eventually be designed explicitly around autonomous machines: standardized docking, spatial references, charging, material interfaces and machine-accessible logistics.

Though these are little more than hypotheses or predictions for now, I have recently witnessed firsthand several of these patterns beginning to emerge in major robotics firms and startups alike. Their chance of materializing depends significantly on how large and heterogeneous general-purpose robotics manufacturing ultimately becomes.

Where the analogy breaks

The most consequential difference is uncertainty.

Semiconductor fabrication is by no means deterministic. Yield is statistical; materials vary; equipment drifts; defects occur; this is precisely why metrology and process control are valuable. But the manufacturing recipe itself is deliberately constrained. A lithography tool is not expected to infer from visual context what operation the process engineer probably intended.

A learned robot policy is different though. Its behavior arises from a trained model interacting with a physical environment. Small errors made by VLAs, world models, and WAMs alike can push the robot into states further from its training distribution, and this “exposure bias” often causes subsequent errors to compound and deteriorate policy rollout performance. Across the board, model developers have no choice but to evaluate policies in terms of empirical task-success distributions rather than specified solely from first principles. [14, 15]

A future robotic manufacturing process might therefore have to accommodate both physical process variation, as well as policy-behavior variation.

This makes a robotic PDK fundamentally different from a semiconductor PDK. A conventional rule might state that a geometric dimension cannot be smaller than some threshold. A robotic capability model may instead need to express something closer to

The second difference concerns incumbent infrastructure. Robotics is not starting from a blank slate: as mentioned, Siemens, Dassault Systèmes, NVIDIA, and related industrial-software companies already own important engineering workflows, including simulation and digital-twin infrastructure. That creates a plausible future in which probabilistic robotic design functionality is incorporated into existing platforms rather than becoming a wholly new standalone EDA industry.

Nonetheless, independent companies may be more likely to remain valuable where independence itself matters. A robot OEM can characterize its own machines and a robot foundation-model firm can measure its own policies, but an independent process-control company can potentially compare many robot, policy, tooling and process combinations without having an incentive to make any particular one look favorable.

Standards also become particularly important in this heterogeneous world. Robotics already has early examples: the OPC UA Robotics Companion Specification provides a manufacturer-independent information model for robot condition and asset data; MassRobotics released an AMR interoperability standard in 2021; and VDA 5050 version 3.0, released in 2026, defines interfaces between mobile robots and fleet-control systems. [16–18] These are narrower than the abstractions envisioned here, but they illustrate the general progression: interoperability becomes economically important once customers operate sufficiently heterogeneous systems.

The order of ecosystem development

Perhaps the most presently relevant semiconductor lesson concerns when these abstracted industries appeared.

It was by no means “foundry first, ecosystem second”. Several specialized layers preceded the pure-play foundry by decades. Teradyne was founded in 1960 around electronic test; Applied Materials dates to 1967; Amkor’s semiconductor packaging business dates to 1968; and the aforementioned KLA, Lam Research, ASML, and Synopsys followed suit in 1975, 1980, 1984, and 1986, respectively. Only in 1987 did TSMC pioneer the dedicated foundry business model. [1, 3, 5–8, 19, 20]

In general, specialized infrastructure seems to become viable when the cost of a recurring industrial problem exceeds the cost of addressing it in a separable, independent manner.

Measurement appeared because expensive processes needed to be understood; outsourced backend manufacturing appeared because that function could be functionally separated from chip design; equipment suppliers appeared because specialized physical processes became too technically deep for every chipmaker to reinvent; EDA grew as design complexity crossed the threshold at which abstraction and automation became economically necessary; foundries strengthened the need for standardized interfaces between design and process.

In the robotics space, it may well be the case that lifecycle services and refurbishment are already economically plausible, because there is already a large installed base of robot hardware. Telemetry, interoperability, and metrology, too, could possibly appear relatively early – they depend less so on the commoditization of general-purpose robots and more so on the existence of heterogeneous deployed systems and analytics bottlenecks.

On the other hand, manufacturing capability descriptions and probabilistic design rules may take a bit longer – they likely require learned behaviors to become stable enough such that empirical characterization remains meaningful across deployments and software revisions. A true robotic manufacturing compiler may appear later still. A system that accepts product geometry and production constraints and synthesizes a manufacturing configuration needs mature representations of everything from products to policy capability to tooling, operations, and hardware costs. Without such fine-grained lower-level information, such a product risks becoming bespoke systems integration software.

What this could mean for the robotics industry

The semiconductor analogy suggests a way of identifying where economic specialization in general-purpose robotic manufacturing is likely to emerge.

Some layers are naturally adjacent to existing incumbents: factory layout, conventional robot simulation, offline programming, virtual commissioning and increasingly some forms of automated process planning fit naturally into platforms such as Siemens and Dassault Systèmes. NVIDIA has an equally plausible position in large-scale simulation and digital-twin infrastructure.

Other layers have stronger reasons to remain independent, and may serve as potential opportunities for new companies. Cross-OEM process measurement, standardized empirical capability characterization, neutral asset health and residual-value information, multi-vendor interoperability, specialized tooling, and robot-native physical components all potentially benefit from a heterogeneous ecosystem rather than a vertically integrated monolith.

The central technical question of the past several years in physical AI has been whether robots can acquire sufficiently general manipulation capabilities. If general-purpose robotics succeeds, the next industrial question may be how to build an economy around those capabilities: how to measure them, specify them, design products for them, compare them, integrate them, and trust them in production.

That is where I find the semiconductor analogy most useful. The mature semiconductor industry is an existence proof that the companies surrounding a core manufacturing technology can become as technically deep and strategically important as the companies producing the final product. Robotics may eventually produce its own equivalents, but the exact categories will likely be shaped by the scaling laws of physical AI: increasingly capable machines will also be increasingly general, learned, and probabilistic. [21]

I am particularly interested in speaking with people working at the boundary between robot learning, manufacturing engineering, uncertainty quantification for ML, industrial software, and factory automation. If you would be interested in chatting about research, startups, or the future of robotics, feel free to reach out at ishaanklv@gmail.com.

References

1. Taiwan Semiconductor Manufacturing Company, "About TSMC," TSMC, accessed August 2026. tsmc.com/english/aboutTSMC
2. Taiwan Semiconductor Manufacturing Company, "Technology Leadership," TSMC 2012 Annual Report, 2012. investor.tsmc.com/static/annualReports/2012
3. Synopsys, "Board of Directors," Synopsys, accessed August 2026. synopsys.com/company/corporate-governance-ethics/board-of-directors
4. Cadence Design Systems, "A Brief History of Cadence: The Solomon-Costello Era," Cadence Breakfast Bytes Blog. community.cadence.com/breakfast-bytes
5. Applied Materials, "About Applied Materials," Applied Materials, accessed August 2026. appliedmaterials.com/us/en/about
6. Lam Research, "A History of Innovation," Lam Research, accessed August 2026. lamresearch.com/company/history
7. ASML, "ASML's Founding Story: Our Roots in the Semiconductor Industry," ASML, 2024. asml.com/en/company/stories/2024/asml-founding-story
8. KLA Corporation, "Annual Report on Form 10-K for the Fiscal Year Ended June 30, 2026," U.S. Securities and Exchange Commission, 2026. ir.kla.com/sec-filings
9. KLA Corporation, "Our Company," KLA, accessed August 2026. kla.com/company
10. SurplusGLOBAL, "Company Profile: Overview," SurplusGLOBAL, accessed August 2026. surplusglobal.com/company-profile/overview
11. Dassault Systèmes, "DELMIA," Dassault Systèmes, accessed August 2026. 3ds.com/products/delmia
12. Siemens, "Process Simulate," Siemens Tecnomatix, accessed August 2026. siemens.com/tecnomatix/process-simulate-software
13. NVIDIA, "Industrial Facility Digital Twins," NVIDIA, accessed August 2026. nvidia.com/en-us/use-cases/industrial-facility-digital-twins
14. Physical Intelligence, "π0.7: A Steerable Model with Emergent Capabilities," Physical Intelligence Blog, 2026. pi.website/blog/pi07
15. Generalist AI, "GEN-1: Scaling Embodied Foundation Models to Mastery," Generalist AI Blog, 2026. generalistai.com/blog/gen-1
16. OPC Foundation, "OPC 40010-1: OPC UA for Robotics — Part 1: Vertical Integration," version 1.02. reference.opcfoundation.org/specs/OPC-40010-1
17. MassRobotics, "Autonomous Mobile Robot Standards Published by MassRobotics," MassRobotics, May 2021. massrobotics.org/autonomous-mobile-robot-standards
18. Verband der Automobilindustrie, "Version 3.0 of VDA 5050 Released," VDA, April 2026. vda.de/en/press/press-releases/2026
19. Teradyne, "Annual Report on Form 10-K for the Fiscal Year Ended December 31, 2025," U.S. Securities and Exchange Commission, 2026. sec.gov/Archives/edgar/data/97210
20. Amkor Technology, "Company History," Amkor Technology, accessed August 2026. amkor.com/company-history
21. Open X-Embodiment Collaboration, "Open X-Embodiment: Robotic Learning Datasets and RT-X Models," arXiv preprint arXiv:2310.08864, 2023. arxiv.org/abs/2310.08864