Connected

By Donald Southerton

Contents

Introduction

I’m a product of the 1960s. My first car was a low mileage 1962 Ford Green Falcon with a 144 cubic inch engine – no frills. As a sixteen-year-old, I painted the hood flat black and installed chrome hood pins. One accessory, the AM mono radio that featured manual tuning knobs and pushbuttons, which now are seen as vintage.

Over the years, and several Ford Pintos later, I moved on to a muscle car – a formidable 1970 Boss 302, with a mechanical 4-barrel carburetor and dual point distributor vs. any fuel electronics.

Working in the car industry since 2003, I’ve been an observer in the pivot to EV, autonomous, hybrid, and ADAS, with some work in LIDAR, battery tech, and more recently AI and semiconductors. I’d layer on, too, frequent VIP visits to “state of the art” vehicle manufacturing plants and Tier suppliers; so, I’ve seen the role of the vehicles, smart manufacturing, and robotics evolve, not to mention the IT behind the scenes.

As I recall, one of the first mentions of “software” was in an OEM not loading the latest updates prior to shipping cars from Asia, but they waited until the shipments arrived and then updated the data in the ports. Today a standard practice – although most updates now Over-the-air (OTA) updates.

I also remember how the car industry pivoted from the industry’s Detroit Auto Show to the CES, both held early January. Tech, infosystems, seen early for automotive as growing in importance.

My roadmap for this manuscript was to add some of my personal insights along with layers of the connected technologies. When possible I added how the connected architecture was adopted across platforms along with compliance.

As an example: Hyundai Motor Group’s push into connected cars runs through a single in-house team: 42dot, the group’s software center where Hyundai has folded its electronics, infotainment, and autonomous driving.

Pleos Connect is where their technology first shows up in a next-generation infotainment system with a widescreen display and split-screen multitasking.

On top of that is Gleo, the voice assistant that handles navigation, climate, and vehicle functions.

Gleoguard is its safety guardrail module.

Taken together, Pleos Connect, Gleo, and Gleoguard are one stack. It’s a clear example of building its AI software in-house.

Chapter 1: Connected Vehicles

It’s been fascinating to see car tech evolve from the early days of infosystems tethered to my phone, to higher resolution dash camera and displays, Advanced Driver Assistance System (ADAS), and GPS map systems. With each new model year these packages grew. They tend to be outward systems, with an array of technologies accumulating “under the floorboards.”

A “connected car” is defined as any vehicle with a network connection linking it to external systems: the OEM cloud, mobile apps, other vehicles (V2V), and road infrastructure (V2I).

We begin by looking at the Architecture Basics, for which the core components are:

  • Telematics Control Unit (TCU) – the vehicle’s cellular modem/gateway to the outside world.
  • In-vehicle network Controller Area Network (CAN bus) is a specialized internal communications network. It allows microcontrollers and Electronic Control Units (ECUs) to talk to each other directly without a central computer. The vehicle’s Ethernet links the ECUs: engine, brakes, infotainment, and ADAS.
  • Over-the-air (OTA) update system – pushes software/firmware to ECUs remotely.
  • Backend cloud platform – OEM servers handling data, diagnostics, app connectivity.

Software Update Management System

The core requirements are:

  • A certified Software Update Management System (SUMS) process for the full update lifecycle: development, validation, deployment, rollback.
  • RX Software Identification Number (RXSWIN) – a traceable ID confirming which software version is installed.
  • Safe update delivery – updates must not compromise safety-critical systems, even if interrupted mid-update.

Cybersecurity Management System

What OEMs must demonstrate:

  • A certified Cybersecurity Management System (CSMS) covering the full vehicle lifecycle: design, production, and post-production (in-service monitoring).
  • Risk assessment process for identifying and mitigating cyber threats before type approval.
  • Incident monitoring and response – ability to detect and react to cyberattacks on vehicles already on the road.
  • Supply chain oversight – verifying supplier components meet the same cybersecurity bar.

Every one of the connection points above is also an attack surface. Connected vehicles face threats similar to any networked IT system, but with safety consequences.

Key risk categories

  • Remote exploitation – attacker gains control via TCU, infotainment, or app APIs.
  • Supply chain risk – vulnerable components/software from Tier 1 and Tier 2 suppliers.
  • OTA update hijacking – malicious or corrupted software pushed to fleet vehicles.
  • Data privacy – location, driving behavior, and biometric data exposure.

Compliance, The Global Compliance Landscape – R155 and R156

More connectivity also means more regulatory exposure.

Regulators responded with two linked UN regulations: R155 on how you manage cyber risk and R156 on how you manage software updates safely. Together, they are now a precondition for vehicle type approval in many major markets.

More specifically, R155 requires a certified, organization-wide Cybersecurity Management System (CSMS) covering design, production, and post-production in-service monitoring, plus risk assessment, incident response, and supplier verification.

And R156 governs how OTA and wired software updates are planned, deployed, and verified; it is the companion regulation to R155.

OEM Approaches

I am assuming the major vehicle manufacturers may be taking similar approaches, although maintaining their proprietary oversight.

What is universal is how they may be looking at the shared landscape, the above-mentioned links to external systems: the OEM cloud, mobile apps, other vehicles (V2V), and road infrastructure (V2I) as well as smart charging, V2G, and V2H.

Examples are Hyundai, Volvo, and their Connected Vehicle Architecture and Compliance.

The Hyundai Motor Group is converging its connected vehicle effort on one in-house software brand (Pleos), one consolidated telematics unit (Mobis MTCU), and in-house certification through Hyundai AutoEver. This vertical integration allows for greater oversight in the supply chain that R155 imposes on OEMs as well as maintaining strict internal governance.

Software Layer: Pleos

Hyundai launched Pleos as its mobility software brand at the Pleos 25 developer conference in Seoul in March 2025.

The first production deliverable, Pleos Connect, shipped in May 2026 on the new Grandeur in Korea, with the IONIQ 3 in Europe next in a phased global rollout. Target is roughly 20 million Hyundai, Kia, and Genesis vehicles by 2030.

Three design choices matter for risk. First, continuous OTA feature and performance delivery, which puts R156 process discipline on the critical path for the product roadmap, not just for recalls.

Second, an LLM-based agent (Gleo AI) that interprets driver intent rather than executing fixed commands.

Third, an open development platform giving third-party mobile developers access to automotive software endpoints, with plug-and-play support for external hardware connecting directly to the vehicle.

Hardware Layer: The TCU

Hyundai Mobis introduced a 5G antenna-integrated Multi-Function Telematics Control Unit (MTCU) to prospective customers at CES 2026. Consolidating modem, antenna, and gateway into a single unit reduces the number of exposed physical interfaces but concentrates risk in one node.

Compliance Layers

Hyundai AutoEver, the Group’s IT and the vehicle software arm, holds a Level 3 cybersecurity capability certification, aligned to the ISO/SAE 21434 processes that underpin a CSMS.

AutoEver earned ASIL-D functional safety certification for its mobilgene adaptive vehicle software platform in February 2025, the highest classification under ISO 26262.

Because the Group sells into the EU and Korea, it holds valid CSMS and SUMS certificates.

Hyundai launched AllDayEnergy in July 2026 as its global V2X brand – smart charging, V2G, and V2H.

This is the EV charging grid interconnection, an additional risk category from the V2V and V2I in the standard connected vehicle framework.

Volvo

Building on Volvo’s connected technology, first launched in 2016, their free over-the-air (OTA) updates enable hazard warnings using data shared by other vehicles.

As of August 2026, more than one million Volvo models have had active OTA capabilities, not to mention by using the European Data for Road Safety ecosystem, alerts can also be passed beyond Volvo models.

For example, Volvo software updates in 2026 introduced four safety features including a ‘Large Animal Alert’ that tell other drivers if a big animal such as a deer is detected on or near the road ahead.

Volvo’s ‘Vulnerable Road User Alert’ flags if pedestrians and cyclists are spotted ahead, with ‘Roadwork Alert’ and ‘Accident Ahead Alert’ also ensuring drivers are forewarned of hazards on their route.

OEMs

I mentioned that OEMs are converging on similar approaches with proprietary oversight, and that is broadly true at the architecture level. The real differentiator is ownership of the stack.

Hyundai owns AutoEver (IT), Pleos (software), and Mobis (hardware), thus, can certify in-house. Toyota uses its subsidiary Woven for software and smart mobility, while the Volkswagen Group has relied on its software company CARIAD.

Under R155, supplier cybersecurity verification is one of the harder obligations because it depends on the parties you control, but not those you have no oversight. Hyundai has structurally reduced that exposure.

Most other OEMs will be challenged managing dozens of Tier 1 relationships.

Chapter 2: Smart Manufacturing

The car taught the industry how to build that architecture. The factory is now an application of the same lesson.

I see both a connected vehicle, and a modern plant are the same architecture, but on different scales. They are software defined systems from networked subsystems that sense, decide, act, and update.

Subsystem Parallels

  • Electronic Control Units (ECUs) and domain controllers map to Programmable Logic Controllers (PLCs), edge controllers, and cell controllers. Both moved from many small, distributed brains toward fewer consolidated ones.
  • Controller Area Network (CAN bus) and automotive Ethernet map to industrial Ethernet, time sensitive networking, and private 5G on the plant floor.
  • The telematics control unit maps to the plant gateway that links the floor to Manufacturing Execution System (MES), Enterprise Resource Planning (ERP), and the cloud.
  • Over-the-air updates push firmware to machines remotely, so the plant becomes updatable.
  • Vehicle health monitoring maps to predictive maintenance on presses, welders, and robots. Same telemetry, same anomaly detection, different asset.
  • ADAS sensor fusion maps to machine vision and inline quality inspection. Multiple imperfect sensors resolved into one confident decision.
  • The cockpit HMI maps to the operator interface and the augmented worker. Both are about presenting a complex system to a human in a way that does not overload them.
  • R155 cybersecurity management and R156 software update management map to IEC 62443 and factory operational technology (OT) security. IEC 62443 governs the cybersecurity of OT and manufacturing plant control systems.
  • Tier supplier oversight maps to supply chain data integration and traceability. As I mentioned, any weak connected supplier comes with risks.

Digital twin can be used first in design and in live operations. A digital twin in manufacturing creates a dynamic, virtual copy of a physical factory, product, or production line. It lets engineers run virtual simulations, predict machine failures, and test adjustments safely in software. Look at Chapter 5 for more on digital twins.

As with Connected Vehicles some OEMs, like Hyundai, are vertically integrated. This vertical integration is the structural advantage. The same group designs the vehicle, the battery, the robots, and the plant, so thinking transfers across the group in a way many OEMs cannot match.

Hyundai Metaplant America, Hyundai Motor Group Innovation Center Singapore and the Ulsan EV plant are the most visible showcases, purpose built around software defined production.

In addition, Hyundai E-Forest is their smart factory brand. It combines artificial intelligence, robots, and big data to build a software-defined factory ecosystem.

The ecosystem also carries the regulatory head start from connected vehicle compliance, which is transferable to industrial cybersecurity.

More so, for example Samsung and LG smart factory programs, as suppliers, have their own parallel systems supplying batteries, displays, and components into end user vehicles.

Just in Time

Just-in-time (JIT) manufacturing is a production approach where parts and materials arrive exactly when they are needed on the line, rather than sitting in inventory.

Toyota developed it in the 1950s-60s as a core pillar of the Toyota Production System, the foundation of what is now called “lean manufacturing.”

About ten years ago, I saw it in action at a Hyundai Motor Tier 1 – an off-site location just miles from the vehicle manufacturing plant.

The Mobis facility was final assembler for the dash module. More specifically, it was assembling the dash with components from the downline of Tier 2 and 3 suppliers. This process prepped the dash for JIT for a specific vehicle at the plant down the road.

Today, many OEMs have adopted the model either onsite or in supplier parks nearby.

Biopharma

Biopharma runs on the same underlying logic as we see in smart manufacturing, but looks nothing like an assembly line. A biologic drug, whether a vaccine, a monoclonal antibody, or a cell therapy, is grown, not assembled.

Living cells are engineered to produce the target molecule, then cultured inside a bioreactor under tightly controlled conditions of temperature, pH, oxygen, and nutrient feed.

Traditionally in biopharma this is what happened in batches: seed a reactor, grow it out over one to two weeks, harvest, purify, test, repeat. Each batch was its own isolated event, and a contamination or yield problem discovered late in the process meant the entire batch was lost, along with the weeks it took to grow.

The shift toward smart manufacturing changes that model in three ways.

The first is continuous bioprocessing. Instead of discrete batches, cell culture and purification run as a continuous flow, with product harvested and processed in real time rather than all at once at the end. This shortens production cycles, reduces the size of the physical footprint needed, and, most importantly, means a problem shows up in hours instead of weeks.

The second is real-time sensor monitoring paired with a digital twin. Modern bioreactors are instrumented with sensors tracking dozens of variables continuously; a digital twin, a live software model of the reactor, ingests that sensor data and predicts how the batch is trending before a human operator would notice a deviation.

This is the same sense-decide-act-update loop that runs a connected factory floor or a connected vehicle, just applied to a biological system instead of a mechanical one. The stakes are different too. A defective part can be scrapped and remade in minutes. A failed batch of biologic drug represents weeks of cultivation and, at scale, millions of dollars.

The third is that the data itself becomes part of the regulatory record. In automotive or electronics manufacturing, sensor data is primarily an efficiency and quality tool. In biopharma, regulators like the FDA and EMA require detailed process documentation to approve and maintain a manufacturing license, so the same real-time data that catches a yield problem also satisfies compliance requirements.

Process, not just product, is what gets regulated and audited, which is the same logic underlying UN R155 and R156 for connected vehicle cybersecurity: certify the process that produces something, not only the thing itself.

South Korea’s Incheon Free Economic Zone (IFEZ; see Chapter 10) bio hub has put these principles into practice on industrial scale, and is anchored by companies like Samsung Biologics and Celltrion.

In turn, IFEZ became one of the world’s largest contract biomanufacturing hubs. Smart manufacturing is built into their plants with continuous bioprocessing and digital twin monitoring. The result – the hub competes directly against established Western manufacturers on speed, consistency, and cost.

It is a clear example of applying smart manufacturing thinking to a sector, biologics, that most people don’t immediately associate with factories or automation at all.

Logistics

If smart manufacturing is the plant becoming a networked system, smart logistics is the supply chain. Same architecture. The plant already knows the state of every machine inside it.

Smart logistics asks the state of every part still in transit toward it.

  • Visibility. Track and trace at item, pallet, and container level through sensors, RFID, and GPS. Continuous state, not checkpoint reporting.
  • Prediction. Demand forecasting and arrival estimates that adjust to weather, port congestion, and disruption rather than explaining delays after they happen.
  • Automation in the node. Automated storage and retrieval, robotic picking, autonomous yard equipment, and increasingly lights out warehouses.
  • Automation in the link. Truck platooning, autonomous line haul on fixed corridors, drone and robot last mile, and automated marine terminals.
  • Orchestration. Control tower systems sitting above carriers, third party logistics providers, and plants, making routing decisions across all of them. This is where the value actually lands.
  • Digital twin of the network, used for scenario planning before disruption rather than firefighting during it.
  • Autonomous replenishment. Systems that reorder and reroute without a human in the loop, which is the point where logistics stops being a reporting function and becomes a decision function.

The Challenges

Inside a plant you own every sensor. Across a supply chain you are asking competitors, carriers, and customers to share data.

There is no CAN bus for logistics. Formats, identifiers, and event definitions may differ by carrier, mode, and country.

Vertical integration removes the trust problem that stalls most Western visibility projects. In turn, when the logistics arm, the plant, and the OEM sit inside one ownership chain, data sharing is an internal decision.

Korea’s port and terminal automation work, and its position as a transshipment hub, give it operating scale.

Domestic e-commerce fulfillment has pushed last mile speed expectations further in Korea than in almost any comparable market, which pulls the rest of the chain forward.

The same chaebol structure that accelerates integration also concentrates risk. One disrupted affiliate propagates across the group faster than it would across arm’s length suppliers.

Hyundai Glovis, the Group’s global smart supply chain management and a third-party logistics provider, may be the clearest case for vertical integration. They span marine transport, vehicle distribution, and parts logistics for the group.

Chapter 3: Semiconductors

To paraphrase, Michael Malone, one of America’s best-known technology writers, who has predicted game-changing technology that fuels Silicon Valley’s growth.

“I think it is going to be battery power,” he said, in a 2011 talk. Malone felt batteries may follow the path of other technologies and Moore’s Law. Today, I’d argue “it is semiconductors.”

Moore’s Law, framed by Intel co-founder Gordon Moore, predicted that the number of components on a chip would double about every two years, making computers faster and cheaper. It has shaped the electronics sector, modern computing, and the semiconductor industry.

That said, shrinking transistors every two years to make cheaper, faster chips has largely ended. Instead, the industry now relies on alternative methods to boost performance.

Frankly, I’ve had a long-distance relationship with semiconductors including several decades living in New York State’s Hudson Valley. At the time in the 1980s it was IBM’s Hudson Valley semiconductor and mainframe golden age defined by peak employment, unmatched local economic dominance, and major technological output at its zenith.

Today, after more than 20 years supporting the major Korea Groups as they expanded globally, “chips,” again too, play a big part in my day.

Everything in Chapter 1 runs on chips. The vehicle’s TCU, the domain controllers, the ADAS sensor fusion, the PLCs on the plant floor, the RFID readers tracking a pallet across an ocean.

A semiconductor almost always is silicon, a material that conducts electricity only when needed like a switch. That switch is a transistor. These transistors make up a modern processor. A modern phone chip has 15 to 20 billion of them. Smaller transistors mean faster and more power-efficient chips. Each shrink in technology gets exponentially harder and more expensive.

The Three Layers of the Industry

I see three layers. Design, Manufacturing and the Equipment.

For Design, companies like Nvidia, Qualcomm, and Apple draw the blueprint. They own no factories, and this business model is called “fabless.”

In turn a manufacturer is a “fab.” They print the design onto silicon wafers, layer by layer, in a cleanroom that costs billions.

The Equipment are suppliers who build the tools that make the chips, including EUV lithography machines. An EUV lithography machine is, at its core, a printer, but instead of depositing ink on paper, it prints microscopic circuits onto silicon.

Logic and Memory

Within this ecosystem, Logic does the thinking, most commonly through CPUs, GPUs, and phone processors. Memory stores the data usually in DRAM and NAND flash.

The Foundry

A foundry is a chip factory for hire. For example, Nvidia designs a GPU and a foundry builds it.

  • Pure-play foundry. Manufacturing only, no products of its own.
  • IDM (integrated device manufacturer). Designs and builds its own chips. They may also sell their foundry services.

The Back End

OSAT (Outsourced Semiconductor Assembly and Test) A company that takes finished wafers from a “fab” and processes them.

More specifically, what actually happens is the “fab” hands over a wafer of finished circuits. In this back-end stage, the wafer may be cut apart into hundreds or thousands of separate chips. Each chip is often attached to a substrate, connected by fine wires or solder bumps, encased in a package, labeled with part numbers, and undergoes final testing.

Where the Chips Go Next

Once a chip is packaged and tested, it leaves the semiconductor industry and enters the electronics industry. Four steps follow.

Binning and Final Test

Chips are sorted by how they performed.

Packing and Shipping

Chips are loaded onto trays, sealed against moisture and static, and shipped either to a distributor or directly to the customer under contract.

Board Assembly

A contract manufacturer places the chip onto a printed circuit board alongside hundreds of other components and solders it in a reflow oven. This is where a chip becomes part of a working system. Some chips may also first go to a module maker, which combines several parts into a subassembly such as a memory module or a camera module.

Final Product

The board goes into a phone, a car, a server, an appliance. Then the whole assembly is tested again, because a chip that passed at the factory can still fail once it is soldered.

Before we move on…

In the 1967 film “The Graduate,” a family friend, Mr. McGuire, pulls young Benjamin aside at his graduation party to offer single-word wisdom about the future: “Plastics. There’s a great future in plastics. Think about it.” In 2026, I’d reframe it to “Semiconductors.”

A modern vehicle carries well over a thousand chips. A software defined plant and logistics ecosystems demand more. Every layer we just walked through, design, fab, equipment, back end, is upstream of every connected system.

Chapter 4: Robotics

A robot is a connected vehicle that never leaves the building. Same architecture. Sensors, controllers, a network, a gateway to the cloud, software pushed over the air. The difference is that it may have arms or wheels, and not driven on a road.

The vehicle chapter described a machine that became a networked system. The chip chapter described what that system is made of. Robotics is when the same architecture is given a body and put to work.

A robot does not simply run a model. It must perceive a changing environment, reason about contact and motion, act, and recover when actions fail. Development spans from real-world demonstrations, synthetic data, physics simulation, policy training, structured evaluation, deployment, and the collection of failure.

Physical AI infrastructure is becoming a system of interdependent layers rather than a synonym for computing capacity.

These categories cover most of what is deployed today.

  • The articulated arm, typically six axes, bolted to a floor or a pedestal and working inside a fenced cell. Variants can include Selective Compliance Assembly Robot Arm (SCARA) for high-speed pick up and place, delta robots for light and fast sorting, and Cartesian robots, multi-axis linear motion platforms mounted overhead or around a workspace.
  • The collaborative robot (Cobot), built to work in shared space with people. Lighter payload, slower, rounded surfaces, torque sensing in every joint so it stops when it meets resistance.
  • Automated guided vehicle (AGV). It follows a fixed route defined by infrastructure, magnetic tape, wire in the floor, reflectors, or floor markers.
  • Autonomous mobile robot (AMR). It navigates by onboard sensing, builds its own map and plans its own path, plus routes around an obstacle rather than stopping in front of it.
  • The general-purpose humanoid form, a bipedal or a wheeled torso with two arms.

As I mentioned, a robot does not simply run a model. It must perceive a changing environment. This is more evident in field robotics.

Field robotics has been moving quickly from specialized agricultural autonomous applications into outdoor work, including energy, construction, mining, environmental monitoring, landscaping, and infrastructure inspection.

These applications are in less predictable environments. In other words, navigating rough terrain as well as working in changing weather, which require acting on data in real time.

The Stack

The robotics industry splits the same three ways the chip industry does, and for the same reason. The work is too specialized.

  • Design and brand. These are robot makers. They build the kinematics, the controller, the motion software, and the customer relationship.
  • Integration. The equivalent of the fab. The integrator turns a general-purpose machine into a specific job.
  • Components. The equipment layer – precision reducers, servo motors, encoders, force and torque sensors, lidar, and 3D vision.

The stack runs in four layers.

  • Sensors and actuators. The body. Servo motors and gearboxes, encoders reporting position thousands of times a second, force and torque sensing at the wrist, safety rated laser scanners, and vision.
  • The controller. The brain of one machine. Kinematics, trajectory planning, and a hard real time loop. This is the domain controller parallel, and it is consolidating the same way vehicle control units did, from many small brains toward fewer large ones.
  • The fleet manager. The layer above the individual machine. Task allocation, traffic control on shared aisles, charging queues, and handoffs into the manufacturing execution system or the warehouse management system.
  • The cloud. Telemetry, predictive maintenance, over the air firmware, offline programming and simulation, and the digital twin of the cell. This is the telematics link, without the cellular modem.

Safety and Standards

More capability means more regulatory exposure. Same pattern as the vehicle chapter, and the standards landscape in the same way by regulating the process rather than the product.

The framework has ISO pieces.

  • ISO 12100. The base. Risk assessment and risk reduction for machinery. Everything else assumes you did this first.
  • ISO 10218. The core robot safety standard, in two parts. Part one covers the robot itself, part two covers the integrated system, meaning the cell as installed. The 2025 revision matters because it pulled most of the collaborative operation guidance into the standard proper and expanded the treatment of validation and cybersecurity. In the United States the same content is adopted as ANSI/A3 R15.06.
  • ISO/TS 15066. The technical specification for collaborative operation. It defines four modes, and they are the working vocabulary of the field. Safety rated monitored stop, the robot halts when a person enters. Hand guiding, the operator moves the arm directly. Speed and separation monitoring, the robot slows as the person gets closer and stops if the gap closes. Power and force limiting, the robot is engineered so that contact stays below defined biomechanical thresholds, which vary by body region.
  • ISO 3691-4 and ISO 13849. The first governs driverless industrial trucks, which is where AGVs and AMRs land. The second is functional safety for machinery control systems, the performance level and dual channel architecture that lets you claim a stop function actually stops.

The structural point is the one I made about R155. The obligation sits on the process, not on the box. A robot maker can certify the arm, but the risk assessment, the validation, and the conformity of the installed cell belong to the integrator and the end user.

What gets approved is the application. That is a different burden from buying a certified component and it catches buyers out regularly.

Then the R156 problem arrives. Once a robot accepts remote firmware and configuration updates, validation stops being a one-time event. An update that changes a motion profile, a speed limit, or a safety parameter can invalidate the assessment that cleared the cell for production. Change management becomes continuous, which is exactly the discipline R156 forces on vehicle OEMs. Most plants are not yet running their robots with that discipline.

The compliance issues look very much like the same ones on a modern vehicle program.

Chapter 5: Digital Twins

A digital twin is a working virtual copy of something physical, fed by data from the real thing, and used to test, predict, and decide.

It helps me to think in three levels.

The ‘digital model’ is a virtual representation, which is essentially CAD or a simulation study.

Next, a ‘digital shadow’ adds a one-way data feed, so the virtual copy follows the physical one but cannot influence it.

A ‘digital twin’ closes the loop, meaning insight from the virtual side changes what the physical side does.

Many projects that call themselves twins are actually shadows.

The digital twin parts are similar to the connected vehicle in Chapter 1. Sensors; a gateway that moves the data; a physics or geometry model that represents how the thing behaves; an analytics and AI layer that finds the pattern, and finally, a visualization and decision layer where a human, or increasingly a system acts.

Sensors, controller, network, cloud, feedback. Same architecture, different asset.

I feel several things had to arrive to make this practical. Sensors and industrial connectivity got cheap. Private 5G and time sensitive networking made plant floor data continuous rather than sampled. GPUs, too, accelerated processing, which lets a model run faster than the real process.

Layer on, Universal Scene Description (OpenUSD), it gives a common language to assemble CAD geometry, physics, and materials into a unified 3D scene. This lets complex CAD models move seamlessly between engineering, simulation, and rendering tools without destructive file conversions.

OpenUSD also gave the industry a common way to describe a scene so that a CAD model, a plant layout, and a robot simulation can live in one environment. And AI models made it possible to learn behavior that is too complex to write down as equations, and to generate synthetic data when needed.

Twins fall into four categories.

  • Design, meaning you test a layout, a process, or a product before actually building.
  • Commissioning, meaning you can start up faster because the control code was already tested against a simulated machine.
  • Operations, meaning predictive maintenance, energy optimization, quality root cause, and scheduling.
  • Planning, and scenario testing against disruption before it occurs.

Digital Twins in Smart Manufacturing

Manufacturing is where it operates at three scales. The ‘asset twin’ covers a single machine, a press, a welder, a robot, a pump. It is where predictive maintenance lives. Vibration, current draw, thermal signature, and cycle time compared against a model of healthy behavior, with the deviation flagged before the failure.

The ‘line or cell twin’ covers a process. Control code is tested against a simulated machine, software in the loop first, then hardware in the loop with the real controller driving a virtual line. Programs that do this well can cut weeks off start up and find the integration errors in an office rather than on a floor.

The ‘plant twin’ covers the whole building, and increasingly the site. Layout, material flow, throughput, bottleneck analysis, ergonomics, energy and HVAC load, as well as construction sequencing.

Beyond those three sits the ‘network twin,’ I’ve touched on in the logistics section, modeling suppliers, transport, and inventory – so you can ask what happens if a port closes.

Here the vertical integration point applies again. Hyundai Metaplant America, the Ulsan EV plant, and Hyundai Motor Group Innovation Center Singapore are built as software defined production with their E Forest program.

When the same group designs the vehicle, the robots, the battery, and the plant, the twin does not have to be negotiated across four companies with four data standards and four legal departments.

Samsung and LG run parallel programs on the supplier side. That structural advantage is not a technology advantage, it is an ownership advantage, and it is the recurring theme in every chapter here.

Digital Twins in Semiconductors

A fab is the most heavily instrumented factory, and it is also the most expensive place to learn by trial. A leading-edge fab runs into the tens of billions of dollars, a wafer carries thousands of process steps, and a single error can be measured in six figures. That combination, extreme instrumentation and extreme cost of error, is why chipmakers moved on twins faster than most other industries.

Four twins

  • The ‘tool twin’ models one piece of equipment, an etch chamber, a deposition system, a lithography scanner.
  • The ‘process twin’ models what happens to the wafer. The most valuable form is virtual metrology, where the model predicts a measurement.
  • The ‘fab twin’ models the building.
  • The ‘product twin’ models the chip itself, and its package.

Digital Twins in Robotics

In robotics the twin is not just an analysis tool, it is where the robot learns.

Driving this are simulations because an AI driven robot needs enormous amounts of experience, and gathering it on physical hardware is slow, and expensive.

So, thousands of instances are trained in parallel in simulation, with lighting, friction, mass, and sensor noise randomized so the learned behavior is transferable to reality.

Although what may work perfectly in simulations can fail on hardware, an ‘operation twin’ is added that is a real-time, physics-accurate virtual copy of a physical robot that connects to the real robot using live sensor data.

Compliance

The compliance angle follows what we discussed in the robotics safety section. A twin can prescreen a risk assessment, model speed and separation monitoring, and estimate contact forces before anyone builds the cell.

What it cannot do is replace physical validation. ISO 10218 and ISO TS 15066 still require measurement on the installed system, and power and force limits are verified against a real body region with a real instrument. Simulation is evidence in the file, not a certificate.

The Thread

Every chapter in this book has twins in it – the vehicle has one, used in development and increasingly over the vehicle life. The plant has one. The fab has one. The robot has one. The battery system and the ships have theirs.

They may have been built separately, but they are now converging – and connected.

Chapter 6: Energy and Storage

A battery does not make energy. It moves energy. A generator produces power when you need it. A battery takes power that already exists.

What connects this chapter to the rest is that the vehicle, the fab, and the robot all consume energy, and how it is stored matters too.

I opened the semiconductor chapter with Malone’s 2011 prediction that batteries would be the next game changer. He was right. Cell costs did fall – roughly on an order of magnitude, a Moore’s Law shaped curve. Chips got cheaper through lithography, while Cells got cheaper through scale, and chemistry.

The Working Parts

The parallel to the connected vehicle chapter is closer than most readers expect. A modern battery system is a networked system with controllers.

  • The cell is the electrochemical unit. Modules handle the mechanical, thermal, and electrical integration.
  • The Battery Management System monitors voltage, current, and temperature, balances cells, and enforces safe limits, which makes it the direct parallel to the domain controller.
  • Thermal management provides precise temperature control by generating, moving, or removing heat within a system.
  • The Power Conversion System handles integration into the grid
  • The Energy Management System decides when to charge and discharge.

Chemistry

The chemistry matters, but it’s often in tradeoffs across energy density, cycle life, charge rate, temperature tolerance, safety, and cost. What matters is best matching chemistry to usage.

  • Nickel-rich Nickel-Manganese-Cobalt (NMC) and Nickel-Cobalt-Aluminum (NCA) deliver high energy density and have been the premium for EV.
  • Lithium Iron Phosphate (LFP) is cheaper, longer cycling, and lower density, LFP took the mass market EV and now grid storage.
  • Lithium-Titanium-Oxide (LTO) charges very fast, cycles tens of thousands of times, and tolerates extremes in temperature, trading density and cost per kilowatt hour, which makes it strong in heavy duty, transit, cold and hot climates, and high cycle grid work where cost per cycle is what matters.
  • Sodium Ion/ Solid State are still in early stages. Its opportunity could lie in applications where its characteristics, including high-power and low-temperature applications, and with data centers as an interesting potential market.

The term solid-state describes an architecture of a battery: one that replaces most or all of the liquid electrolyte in a conventional lithium-ion battery with a solid material. In a lithium-ion battery, the liquid electrolyte shuttles lithium ions back and forth: to the anode as you charge your phone and back to the cathode as you use it. That liquid is highly conductive but flammable, and fires from these batteries burn hot and toxic. In solid-state batteries most if not all that flammable liquid is replaced with a new, solid electrolyte: typically, a ceramic, a sulfide, a polymer or some combination.

Storage

When we talk about energy storage duration, we’re referring to the time it takes to charge or discharge a unit at maximum power.

  • Short-Duration – typically 1 to 4 hours.
  • Long-Duration Energy Storage (LDES): Usually spans 10 hours or longer.

Standards and Compliance

Same structure as every other chapter. More capability, more exposure.

  • UL 9540 and 9540A focus on energy storage system safety and fire testing.
  • NFPA 855 governs installation and sets rules for system spacing, location, and fire protection.
  • IEC 62933 covers grid integration and connected systems.
  • UN 38.3 regulates safe transport conditions for lithium batteries globally.

It is worth emphasizing the EU digital battery passport, which is a mandatory electronic record, and applies to EV batteries, light transport batteries, and industrial batteries. Each passport links via a unique QR code to data on carbon footprints, material origins, recycled content, and the health status (expected life and cycle capacity)

EV vs Grid

Electric vehicle (EV) batteries and stationary grid storage cells/ energy storage systems (ESS) share core electrochemical foundations. Both commonly utilize lithium-ion technology (especially LFP chemistries), function through reversible electrochemical reactions to store and release electricity, rely on similar battery management systems.

That said, EV cells are like sprinters. ESS cells are like marathon runners.

While EV sales growth has experienced a relative slowdown and plateau in certain consumer markets, ESS is seeing exponential demand.

Battery manufacturers are increasingly pivoting production lines toward ESS grid-scale and residential storage to stabilize renewable power and offset EV adoption rates.

But as batteries mature and move into increasingly different applications, from passenger EVs, grid storage, and robotics, the optimization problem becomes much broader.

A truck battery doesn’t need the same characteristics as an urban EV.

A grid battery doesn’t need the same characteristics as robotics.

And even two batteries using similar cells can deliver very different real-world performance depending on the pack architecture, thermal management, BMS, software, power electronics and how the battery is integrated into the application.

Chapter 7: Shipbuilding

In my 2014 Hyundai Way: Hyundai Speed edition I mention how Chung Ju-yung, Hyundai’s founder, had gotten into construction, automotive, steel, and shipbuilding.

To elaborate on how he got into shipbuilding that venture came in the late 1960s when Korean President Park Chung-hee wanted Korea to enter the shipbuilding industry and asked Chairman Chung – even though the Hyundai Group had no experience in the sector.

To fund it, Chung needed around $60 million.

The story that has become Hyundai lore is when a British Barclay banker balked, Chung pulled out a 500 Korean won noting that it pictured the 16th century “turtle ship,” the ironclad warship Korea used to beat back a Japanese invasion centuries before.

His pitch was essentially “we were building armored ships when your navies were still using wood.” It worked, and he got the financing. Construction began at the Ulsan, South Korea port in March 1972.

As Chung had secured the Barclays loan for financing, Chung also landed an order from Greek shipowner George Livanos for two 260,000-ton tankers.

Then in what had become Chung’s legacy, instead of building the dry dock first and start on the ships afterwards, crews were pouring the dry dock while the hull sections for the two tankers were being fabricated.

By 1974, Hyundai completed the shipyard and both tankers together, named Atlantic Baron and Atlantic Baroness – launching both in the same ceremony.

Sadly, the oil crisis and production delays resulted in the order from Livanos being acquired instead by Chung’s new venture – Hyundai Merchant Marine.

My takeaway was Korea had never built anything over 10,000 tons, more so, Chung didn’t wait for the facility to be constructed before beginning to build the ships on land. The spirit of this captured in 하면 된다 (ha myeon doen da) the cultural meaning “Even if it’s impossible, it is possible.”

Today, that yard, now HD Hyundai has grown into what’s now the world’s largest shipbuilding operations.

In particular, Chung Ju-yung’s model for shipbuilders has continued to evolve and now is similar to what we not only find at HD Hyundai, but also in Samsung Heavy Industries and Hanwha Ocean’s massive yards – all driven by innovation.

Common among them, we see tens of thousands of workers, multiple dry-dock projects underway, and scores of massive automated cranes hoisting and welding prefabricated sections into place.

Today, shipbuilding along with automotive and semiconductors are Korea’s top exports.

The Architecture: The Yard Becomes a Sensing System

A ship’s hull is not one machine. It is thousands of prefabricated blocks, cranes, welders, and workers all coordinated across the yard, which is the size of a small city.

Connected technology makes that coordination work in real time. All three yards share the same underlying technology, but each is deploying it differently.

Four layers are converging on Korea’s big three yards at once:

  • Digital twins of the yard itself: live 3D models of docks, cranes, and block locations, used to spot bottlenecks before they stall a dry dock.
  • AI vision and autonomous welding: vision-based systems that find a seam and adjust in real time, solving the arc light, spark, and fume problem that blocked automation for years.
  • Physical AI and humanoid robots: piloted for inclined hull surfaces and confined blocks a wheeled machine can’t reach.
  • Autonomous navigation carried back into the yard: the sensor fusion built for ships at sea now trains the systems monitoring sea trials before a vessel ever leaves.

The Compliance Frame: IACS UR E26 and E27

Shipbuilding’s version of R155 and R156. IACS Unified Requirements E26 (Cyber Resilience of Ships) and E27 (Cyber Resilience of On-Board Systems and Equipment) apply to new construction contracted from July 2024 onward.

Same lesson as Chapter 1, the process is regulated, not the product.

HD Hyundai

  • Adapting Blackwell GPU digital twin at HD Hyundai Samho shipyard with Nvidia and Siemens, target 2030.
  • Welding automation launching in 2026 that will incorporate humanoid welding robots with Persona AI in the US and Neura Robotics in Germany.
  • Installing Avikus autonomous navigation on 350 plus vessels. This is an autonomous system that detects surrounding vessels or obstacles to avoid collisions.

Samsung Heavy Industries

  • Partnered with Dassault Systemes to establish a smart yard that links a vessel’s virtual twin to the yard floor.
  • XR inspection tools with Samsung Electronics using their Galaxy XR that utilizes their mixed reality headset as an advanced industrial inspection tool to optimize enterprise workflows and quality control.
  • Autonomous welding robots with Rainbow Robotics, which is 35 percent Samsung owned.

Hanwha Ocean

  • 3D digital models plus twice daily from drone photography tracks the yard in real time.
  • Smart Sea Trial Center replaces spreadsheets, pulling from 1,500 plus data points.
  • Starlink added to trial vessels in the second half of 2026
  • Contracted NC AI (the physical artificial intelligence subsidiary of South Korean tech giant NCSoft) to develop a vision-recognition-based autonomous welding system vs. traditional welding automation that relies on heavy robots simply repeating static, pre-programmed physical paths.

Chapter 8: AI

The vehicle became a networked system. Silicon is what the system is made of. Robotics gave that system a body. AI is what runs it.

AI is probably my most challenging topic. I was an early adopter, and am a frequent user. That said, I am sensitive to the concerns many have, although for most of this book, artificial intelligence has shown up as a component: a model tuned for demand forecasting, a vision system inspecting a weld, an optimization layer inside a digital twin. It makes recommendations and flags concerns. A person still had to act.

Agentic AI is different. These systems plan, execute, and act with limited human oversight, and in some deployments with none at all. They interact with each other and with external systems, and they take actions that can be difficult or impossible to reverse. Industry is deploying this now, at scale, inside organizations that have not finished building the governance to manage it.

We begin with Architecture Basics, for which the core components are:

  • Model layer – the reasoning engine, typically a large language model (LLM), that interprets intent and plans a sequence of actions rather than returning a single answer.
  • Data layer – the operational data an agent draws on to act, distinct from the data used to train the underlying model, and often less governed.
  • Action layer – the APIs, tool calls, and system permissions that let the agent actually do something: send an email, place an order, push a code change, move money.
  • Oversight layer – audit logs, monitoring, and escalation paths built into the deployment so a human can see what the agent did and intervene.
  • Governance layer – the accountability structure sitting above all of it: who owns the agent’s performance, and who answers when it fails.

Every one of those layers is also where things go wrong.

Key risk categories:

  • Data governance – agentic output is only as good as the data behind it, and when a system that takes independent action is working from bad data, the consequences can land before anyone notices.
  • Consent drift – existing consent frameworks were built for one-time, bounded interactions. Agentic systems act continuously and evolve, so a user’s original authorization can drift out of scope without anyone re-consenting to what it became.
  • The pleaser problem – agents can be coded to satisfy the user’s apparent intent, which means they can produce a confident, complete-looking answer instead of disclosing a limitation, a gap, or genuine uncertainty.
  • Oversight erosion – judgment atrophies as people stop exercising it, and review turns reflexive, a rubber stamp rather than a real check, the same dynamic that turned cookie consent into a click-through nobody reads.
  • Vendor and third-party risk – agentic systems rarely operate in isolation. They connect to external APIs, third-party models, and partner infrastructure.
  • Systemic exposure – the populations least equipped to challenge an automated decision are also the ones most exposed. As an example, the connected vehicle has its own version of this: subprime and buy-here-pay-here auto lenders have fitted financed cars with third-party GPS/starter-interrupt devices that remotely disable the vehicle when a payment is late, sometimes triggering automatically with no person deciding the individual case.

That’s a different channel than the OEM telematics covered in Chapter 1, but the same exposure. It was also a rules-based trigger with the absence of a real appeal path.

Compliance, The Global Compliance Landscape

More autonomy also means more regulatory exposure, but here the parallel to Chapter 1 breaks down in one important way. Connected vehicles have R155 and R156. Agentic AI does not have an equivalent yet. No certified, cross-industry, agent-specific management system exists for how an organization stands up, monitors, and shuts down an autonomous agent. The closest analog, ISO/IEC 42001, certifies an organization’s AI management system, but it was not written for autonomous agents. Korea’s AI Basic Act, in effect since January 2026, sets a national framework for high-impact AI, but it does not yet define an agent-specific management system.

In May 2026, EqualAI convened senior executives from financial services, technology, automotive, telecommunications, and consumer companies, including JPMorganChase, Verizon, Amazon, Microsoft, and M&T Bank, to build a practitioner roadmap for agentic deployment in the absence of that regulation. Their central finding: the gap between agentic AI deployment and agentic AI governance is not a technology problem. It is a prioritization problem, and it affects most of the market.

Regulation will not move fast enough to close that gap. Organizations with strong self-governance are better positioned than those waiting on a rulebook, which is the same lesson R155 taught the automotive industry, just arriving at it from the opposite direction: process before product, self-imposed before mandated.

Company Approaches

It appears that the companies furthest along on agentic deployment are converging on a similar set of practices, even without a shared standard to point to. Six priorities came out of the EqualAI roadmap:

  • Build governance before deploying – prioritize high-value, low-risk use cases first, and do not let the deployment timeline get set by internal pressure to move quickly.
  • Explicitly designate accountability – someone owns an agent’s performance and outcomes. Human accountability does not disappear because a task was performed by an agent.
  • Build leadership capacity – boards and C-suites need enough AI literacy to read a performance report on an agentic system and exercise real oversight, not a rubber stamp.
  • Structure governance teams to match the risk – multidisciplinary teams, or a tripartite thinker-builder-fixer model, or risk-tiered matrix governance calibrated to how much autonomy and data access a given deployment has.
  • Account for agentic risk in vendor agreements – liability allocation, audit rights, and least-privilege data access, negotiated before an incident, not after one.
  • Invest in ecosystem readiness – regulator partnerships, third-party benchmarking, and civil society input, since some of this cannot be solved by any single organization.

A Note on Hyundai and Gleo AI

Chapter 1 noted this: Pleos, Hyundai’s mobility software brand, runs on an LLM-based agent called Gleo AI that interprets driver intent rather than executing fixed commands. At the time, that was a compliance detail buried inside a broader discussion of R155 and R156. It reads differently now.

An agent that interprets intent is, by definition, exposed to the pleaser problem described above. It is also harder to audit than a system executing fixed commands, because there is no fixed command to check the output against.

Whether that ownership gives it the same structural advantage for agentic governance, an internal audit trail for a system that is reasoning rather than executing, is the open question, and not one the industry has answered yet.

Chapter 9: Smart Cities

Beyond my support for Connected technology, especially automotive, my passion for the past 20 years has also been with Songdo, a city built on reclaimed land on the coast of Incheon, South Korea. This began with the original developer, NY-based Gale International, and Songdo International Business District. It has now been renamed Songdo International City, one of three areas under the jurisdiction of the Incheon Free Economic Zone (IFEZ).

In addition to serving as an advisor to the developer, I was commissioned for a pictorial account, Chemulpo to Songdo IBD: Korea’s International Gateway; the book showcased 125 years of commerce, trade, and interaction in the nation’s first foreign trade hub, Incheon. In the bilingual work, as the author, I weaved key historical events with photographs and first-hand accounts, past and present.

Several years later, as a follow-up, I was asked to prepare an update on Songdo, now seen as Korea’s city of the future, smart, connected, and adapting. I continue to visit, support, and advise Songdo startups, the Mayor’s office, and the Incheon Free Economic Zone. One huge change: our original Songdo development managed data over Ethernet, and today apps dominate.

I mention in my writings that a “smart city” has no single agreed definition, but the shared goal is consistent: use networked sensors, data platforms, and AI to improve public services and the daily experience of residents. As with connected vehicles, the technology stack matters less than what it is asked to do.

We begin with Architecture Basics, for which the core components are:

  • The Internet of things (IoT) sensor layer – traffic counters, air and water quality monitors, smart streetlights, parking and waste sensors embedded across the physical city.
  • Smart City Platform – the integration layer combining Geographic Information Systems (GIS) and Information and Communication Technology (ICT) with IoT feeds and AI analytics, giving city operators an operational view.
  • Citizen-facing app layer – transit, parking, and municipal service apps that put the platform’s data in residents’ hands.
  • Backend cloud and data platform – municipal or partner-operated servers handling data storage, analytics, and interagency access, the civic equivalent of an OEM’s backend cloud.

Platform Requirements

The core requirements are:

  • “Interoperability” across agencies – traffic, utilities, emergency response, and sanitation systems that feed one platform rather than operate as isolated silos.
  • Real-time response capability – the platform must support live emergency dispatch and traffic management, not just historical reporting.
  • Citizen data governance – clear rules on what is collected, how long it is retained, and who can access it, since much of the sensor layer touches identifiable movement and behavior data.

Data Privacy and Security

Every sensor and platform integration point above is also an attack surface, and a data privacy question. Smart cities face the same categories of exposure as any large networked IT system, but at municipal scale and with a public trust dimension.

Key risk categories

  • Surveillance concern – CCTV, camera and sensor networks raise legitimate questions about tracking and profiling of residents.
  • Platform consolidation risk – a single integrated platform is efficient, but a breach or outage there can cascade across traffic, utilities, and emergency systems at once.
  • Vendor lock-in – cities that build on one vendor’s proprietary platform may face the same supplier-dependency exposure OEMs face with Tier 1 suppliers.
  • Data privacy – location, transit, and utility usage data are personally identifiable at the household level even when anonymized in aggregate reporting.

City Approaches

Cities are converging on a similar architecture, GIS plus ICT plus IoT plus AI, while differing sharply in governance and ownership. The real differentiator, as with connected vehicle OEMs, is who controls the data.

Songdo International Business District

Songdo is greenfield, build-from-scratch on reclaimed land within the Incheon Free Economic Zone. It is the flagship example: 1,500 acres and roughly 167,000 residents, designed from the ground up around what is often described as a 15-minute city, where residents can reach 90 percent of daily needs by bike or public transit in fifteen minutes or less.

Smart infrastructure is embedded rather than retrofitted: traffic sensors, underground pneumatic waste collection tubes that remove the need for street-level pickup, and integrated apps for transit and parking. Having been part of the original team that built the IFEZ-Incheon, I watched this infrastructure get planned alongside the industrial buildout, not layered on after the fact, which is a meaningful difference from cities retrofitting smart systems onto legacy infrastructure.

Seoul

Seoul, Korea’s ancient capital, is a retrofit case: the city built its integrated Smart City Platform between 2017 and 2019 onto centuries of existing infrastructure that today combines GIS and ICT with IoT and AI to support emergency response, traffic management, sanitation, and facility maintenance from a single operational view.

The platform gained international attention when it was adapted for COVID-19 contact tracing during the pandemic, and it has since been showcased as a model other cities study directly.

Incheon

The Incheon Free Economic Zone Authority announced an expanded smart city vision at CES 2024, with Mayor Yoo Jeong-bok framing the goal as eco-friendly, people-centered development built on collaboration between technology companies and government, rather than a technology rollout led by either side alone.

Autonomous Vehicle Testing

Songdo and Incheon are also where the connected vehicle chapter’s technology meets the smart city’s infrastructure directly.

Hyundai Mobis, the Incheon Free Economic Zone Authority, and Incheon Technopark signed an agreement to run Level 4 autonomous driving trials across a 60-kilometer course spanning Songdo and neighboring Yeongjongdo, with roughly 20 startups, partner firms, research institutes, and universities participating. Testing runs through urban streets rather than closed highway routes, which is the harder and more data-rich environment for validating city driving.

The choice of location is not incidental. Songdo’s sensor-equipped streets and integrated traffic platform, built for the smart city use case, double as ready-made infrastructure for autonomous vehicle validation, giving Hyundai Mobis V2I data other test sites would have to build from scratch.

Governance Models

Songdo, Seoul, and Incheon each demonstrate a different ownership structure for the same underlying architecture. Songdo’s systems were built in from the start by the master developer and IFEZ authority together. Seoul operates its platform directly as a municipal function. Incheon’s approach leans on formal public-private partnership, with the city setting requirements and technology partners building to them.

None of the three depends on a single dominant vendor the way Hyundai’s connected vehicle stack depends on AutoEver, Pleos, and Mobis. That is the structural difference between smart cities and connected vehicles: a city cannot vertically integrate its own supplier base the way an OEM can, which makes interoperability standards and clear data governance the real point of leverage, not vendor selection.

Chapter 10: Biotech and Biopharma

As noted in Chapter 9, Smart Cities, I was part of the team that developed the land where today is the IFEZ-Incheon bio-pharma cluster. That was back when Songdo’s ambitions were still mudflats and master plans rather than the world’s largest single-site biomanufacturing campus. Likewise, I have followed and reported on the buildout as the bio cluster developed.

A “bio-pharma cluster” is a geographic concentration of manufacturing capacity, research institutions, and regulatory infrastructure built to attract the same class of company to the same location.

IFEZ’s own foreign direct investment target – 600 million dollars for 2024, up from 432 million dollars in 2023 – was built substantially around this bio-industry pitch, part of a broader push to reposition Songdo from a residential new-town project into a global business hub competing for the same investment as Dubai, Singapore and Hong Kong.

BIOSECURE

The single biggest driver of the cluster’s growth was not a Korean policy but an American one. The BIOSECURE Act, enacted December 18, 2025, bars U.S. federal agencies and contractors from using biotech services from companies of concern, principally the Chinese CDMOs WuXi AppTec and BGI Group. Estimates put roughly 30 billion dollars in contract manufacturing work at stake over the next decade, work that has to relocate somewhere.

With U.S. domestic capacity limited and European capacity constrained, Songdo was positioned to absorb a share of that demand, not as a temporary landing spot but as a structural decoupling from Chinese suppliers. That is the throughline worth watching: this is a geopolitical supply chain story wearing biotech’s clothes, the same dynamic reshaping semiconductors and shipbuilding elsewhere in this book.

Company Approaches

As with connected vehicle OEMs and city governments, the companies building out Songdo’s cluster share a broadly similar physical architecture while differing in ownership and strategy.

Samsung Biologics

Samsung Biologics is the cluster’s anchor tenant and the world’s third-largest CDMO by revenue, reporting 3.2 billion dollars in 2025 revenue and 845,000 liters of global capacity, with 25.8 percent year-over-year growth and a 99 percent batch success rate in the first quarter of 2026.

In 2026 the company acquired roughly 187,000 square meters of land in Songdo’s 11th zone for 248.7 billion won (about 169 million dollars) to build a third campus, with plans for approximately 7 trillion won in investment by 2030. The new campus is intended to diversify the company beyond its core antibody manufacturing business into cell and gene therapies, antibody-based vaccines, and peptides, consolidating manufacturing, quality, and technical functions across all three Songdo campuses.

Samsung projects the expansion will create over 10,000 direct and indirect jobs and roughly 12 trillion won in economic impact, and follows the company’s recent spin-off as a pure-play CDMO.

Lotte Biologics

Lotte Biologics is the cluster’s most aggressive new entrant, targeting a position among the world’s top ten CDMOs by 2030 on 4.6 trillion won (roughly 3.3 billion dollars) of investment. Three planned Songdo plants are expected to add 400,000 liters of capacity, built in under a decade, a pace only a cluster with existing infrastructure, permitting familiarity, and a trained labor pool could support.

Celltrion

Celltrion, historically a biosimilar developer, has pivoted into CDMO services as a second business line, investing 700 billion won to expand its own manufacturing footprint and acquiring capacity in New Jersey, giving it a foothold inside the U.S. market alongside its Korean base.

Eli Lilly and the Research Layer

Eli Lilly’s announcement of a Gateway Labs research incubator inside Samsung’s Bio Campus II is the detail worth watching most closely. A global pharmaceutical company locating an early-stage research function in Songdo, rather than only contracting for manufacturing capacity, signals that multinational drugmakers are starting to treat the cluster as a strategic innovation anchor, not simply a lower-cost place to fill bioreactors.

Cold Storage

Biologics, vaccines, and the newer wave of cell and gene therapies are temperature-sensitive by nature – many degrade or lose potency within hours outside a set range, and some ultra-low products (below -60°C to -80°C) have almost no tolerance for a break in the chain. As Songdo’s CDMOs scale up antibody, vaccine, and cell/gene manufacturing, the product has to move from bioreactor to patient without a gap in temperature control.

That makes cold storage capacity at the point of export a constraint on the cluster’s growth. A biologic made in Songdo but stuck at ambient temperature during customs or transfer is lost inventory. This is why the airport and the port each need their own dedicated cold chain infrastructure, and why global logistics players are building it out now rather than relying on general-purpose warehousing.

The Smart Manufacturing Tie-In

Cold storage is increasingly wired into the same digital manufacturing push driving CDMOs. Connected sensors on packaging, pallets, and vehicles track temperature, humidity, and location continuously over cellular networks, with live alerts when a shipment drifts out of range.

Marken Korea

  • Opened a cold chain logistics center, inside Incheon International Airport’s Free Trade Zone. 1,411 square meters, with multiple temperature zones: standard refrigeration (2-8°C), freezing (-15 to -25°C), and ultra-low storage (below -60°C).
  • Meets Good Distribution Practice standards, with emergency power and centralized monitoring. Built for biopharmaceuticals, clinical trial drugs, and cell and gene therapies. The pitch is direct airside connectivity for customs clearance and fast import/export of temperature-sensitive shipments, reinforcing Korea’s role as a regional pharma hub.

UPS

  • Announced a 48 million dollar investment to upgrade 27 temperature-controlled cargo cross-docking facilities worldwide, explicitly including expansion of its Korea Incheon air hub. Driven by rising demand from cell and gene therapies, mRNA platforms, and GLP-1 drugs. Facilities are pursuing IATA Pharma certification with 24-hour control centers.
  • Follows UPS’s earlier acquisitions of European and North American healthcare logistics companies, part of a push to build out global cold chain capacity for biopharma.

Korea Ultra-Low Temperature Incheon

  • Selected as primary developer for the Incheon New Port cold chain special zone: 118,658 sqm of a 231,821 sqm total site.
  • The facility recycles LNG cold energy from KOGAS’s Incheon terminal to chill seafood, agricultural products, and biopharmaceuticals through the supply chain – a port-side (rather than airport-side) cold chain hub.

Marken and UPS anchor to the airport (Incheon International Airport Free Trade Zone); Korea Ultra-Low Temperature Incheon anchors to the port (Incheon New Port); these two logistics nodes both serve the biopharma cluster.

What the Cluster Has Not Yet Produced

Korea leads the world in biologics manufacturing capacity but has yet to originate a global blockbuster drug of its own. That gap is the honest caveat to the growth numbers above: Songdo has built the world’s largest production base for other companies’ molecules, but the cluster’s own drug discovery and development pipeline remains the less mature half of the bio-pharma story.

Governance Models

As in the smart cities and connected vehicle chapters, the real differentiator across Songdo’s bio-pharma tenants is not the underlying technology but who owns and controls the stack.

Samsung Biologics owns land, plant, and increasingly the full range of modalities under one corporate umbrella, the same vertical integration logic Hyundai applies to connected vehicles through AutoEver, Pleos, and Mobis. Lotte Biologics is building the same integrated model from scratch, betting that speed into an existing cluster beats organic growth elsewhere.

Celltrion took the opposite path, converting an existing biosimilar developer into a CDMO rather than building fresh capacity, and Eli Lilly’s Gateway Labs represents a third model entirely: a multinational renting a research foothold inside someone else’s campus rather than building its own.

Conclusion

Connected does not just showcase the architecture in cars, smart manufacturing, semiconductors, robotics, digital twins, energy, ships, AI, cities, and biopharma. It shows how each of these sectors can apply that architecture to the greater ecosystem.

Three threads ran through every chapter. First, the same architecture: sensors, controllers, a network, a gateway to the cloud, and software that updates. Second, regulators increasingly certify the process, not just the product, from R155 and R156 in vehicles to ISO 10218 in robotics, IACS E26 and E27 in shipbuilding, and the process documentation biopharma regulators require. Third, ownership is the advantage: the groups that own more of the stack, as Hyundai does, integrate faster and govern more easily.

Every chapter carries the same lesson in a different material: the technology is never the hard part. Governing it does.

Request a private briefing.

Confidential strategy engagements, C-suite briefings, and executive workshops. Limited availability.

Request a briefing Text or call 310-866-3777