IndustryAtlas

SUPPLY CHAIN HANDBOOK

AI hardware: cloud, client and edge systems

Prepared by IndustryAtlasEditorial principles & corrections

Place compute chips within servers, data centers, PCs and edge devices, then examine memory, power and cooling. Different applications have different constraints; a cloud configuration does not apply to every client device.

Most recent source review:

Reading guide revised:

Why do cloud, PC and edge AI require different hardware?

The difference starts with where work happens and how data enters and leaves. Data centers coordinate many systems through networks, supported by facility power, cooling and management. PCs keep selected tasks on the user's device; edge systems sit near cameras or equipment to process local inputs and respond. Compare software and model support, memory, data movement and sustained operating conditions together, rather than relying on a processor's compute figure alone.

  1. Models and application requirements → Compute location and configuration

    Identify inputs, expected responses and software before choosing cloud, PC or on-site execution. Each location needs an appropriate combination of processors, memory and connectivity.

  2. Compute systems and networks → Data-center service capacity

    Networks and storage coordinate machines; racks, power, cooling and management sustain operation. Chip capability becomes usable application capacity only when these supporting systems work together.

  3. Sensor data and edge inference → Local responses and actions

    Edge equipment runs models on incoming signals, then passes outputs to applications or controls. Robots also require mechanical and actuation systems; recognition alone does not execute an action.

What changes the choice?

Choose centralized or local compute around data location and response needs. Cloud systems pool resources; client and edge devices must fit suitable tasks within local power, space and memory limits.

Sources for this overview

AI devices & infrastructure

AI Applications and Devices

Follow where computation runs before tracing its hardware

AI can run in a data center, phone, PC or field device. Identifying the location and task first makes the required chips, memory, boards, networks and cooling easier to understand than putting every product into one chain.

Entry sources reviewed: 2026-09-26

Training and inference have different jobs

Training creates or updates a model; inference applies it to new inputs. Both may use specialized hardware at different scales.

  • A device can run local inference while relying on cloud updates or larger remote tasks.

Translate AI demand into specific parts

Servers connect to accelerators, HBM, packaging and networks. PCs and phones have different power and packaging trade-offs.

  • CPO is one networking option, not a required part of every AI device. Demand growth does not affect all PCB and CCL products proportionally.

Distinction: Cloud and on-device AI can complement each other, but they do not use one shared bill of materials. TOPS or GPU counts alone cannot compare product value.

Company roles

AI spans applications across industries; this map focuses first on computing-platform and hardware-system roles.

  • Wiwynn · Taiwan · Example role · AI system hardware

    AI computing servers and rack integration solutions.

    Product / role source
  • Advantech · Taiwan · Example role · Edge AI systems

    Edge AI modules, inference systems and industrial computing equipment.

    Product / role source
  • NVIDIA · United States · Example role · AI computing platforms

    Accelerated computing, networking and software for AI factories.

    Product / role source

Sources

Explore this role on the mapSuggest a correctionBack to contents

AI PCs and Workstations

Local computation combined with software and model support

AI PCs use CPUs, GPUs and NPUs for suitable tasks; workstations may include discrete GPUs. Actual capabilities depend on applications, models and hardware support, not simply a product name.

Entry sources reviewed: 2026-10-30

Use the right processing engine

NPUs do not replace every CPU or GPU task. Low-power background work and large models have different needs, and software must support the selected engine.

  • Memory and software compatibility are missing from a simple peak-compute comparison.

Follow the complete product

Compute platform, memory, boards, power and cooling jointly shape the experience. Portable PCs also balance battery life and weight.

  • Devices can combine local and remote processing; common technologies do not imply identical data-center packaging or materials.

Desktop personal computersNotebook computersTablet computers

Distinction: An AI PC neither guarantees that all AI applications run locally nor requires a data-center-class accelerator.

Company roles

  • ASUS · Taiwan · Example role · AI PC brand

    Laptops incorporating AI computing platforms.

    Product / role source
  • Dell · United States · Example role · AI PC brand

    Laptops with AI computing capabilities.

    Product / role source

Sources

Explore this role on the mapSuggest a correctionBack to contents

AI cloud infrastructure

AI Data Centers

Compute, networks and facilities designed as one operating system

An AI data center brings compute systems, networks and storage into a facility that can operate continuously. Servers are one part; racks, electrical infrastructure, cooling and management must work together to turn chips into a usable service.

Entry sources reviewed: 2026-10-05

Separate equipment from the facility

Servers contain chips, packages and boards. The data center coordinates many systems and several network roles rather than one universal high-speed link.

  • Power, cooling and management must support the actual deployment density.

Trace material needs by system branch

Servers and switches lead toward PCB, high-speed laminates and packaging. Cooling and power lead toward other equipment and component suppliers.

  • Each branch depends on actual specifications and the deployment design.

Distinction: A data center is not simply a larger server. Adding servers does not guarantee that network or facility capacity is ready.

Company roles

Lists data-center equipment and platform vendors, rather than cloud-service operators.

  • Wiwynn · Taiwan · Example role · Data-center system integration

    Server and rack integration for data-center computing.

    Product / role source
  • Delta Electronics · Taiwan · Example role · Power and cooling infrastructure

    Data-center power, cooling and management solutions.

    Product / role source
  • NVIDIA · United States · Example role · AI data-center platforms

    AI Factory platforms cover accelerated computing, networking and software.

    Product / role source

Sources

Explore this role on the mapSuggest a correctionBack to contents

Data Center Power

Conversion and distribution from the facility to the chip

Power reaches data-center equipment through facility, rack, server and board-level conversion and distribution. Dense computation increases the importance of efficiency, space, heat and stable delivery.

Entry sources reviewed: 2026-10-08

Several levels form the power path

Facility distribution and backup support overall operation; rack and server power supplies serve equipment; board regulators respond to local chip loads.

  • Architecture and voltages depend on the system. New approaches do not imply immediate replacement of every existing installation.

Connect efficiency to cooling and boards

Conversion losses create heat, while boards must meet current and thermal-cycle requirements. Power boards and high-speed signal boards emphasize different properties.

  • Identify the board function before inferring material demand.

Explore power-supply manufacturing

The power-electronics maps trace Si/SiC/GaN fabrication, PSU components and materials while distinguishing modules, shelves and board regulation.

Distinction: Processor power specifications are not complete system specifications. Racks and facilities must include other equipment and conversion losses.

Company roles

  • Delta Electronics · Taiwan · Example role · Data-center power

    Data-center power supply and power management solutions.

    Product / role source
  • Lite-On · Taiwan · Example role · Server power supplies

    Data-center server power supplies and rack power solutions.

    Product / role source

    Sources reviewed: 2026-09-27

  • Vertiv · United States · Example role · Power infrastructure

    Modular power and infrastructure for high-density computing.

    Product / role source

Sources

Explore this role on the mapSuggest a correctionBack to contents

On-device & edge AI

Edge AI Devices

Processing near sensors, users and physical operations

Edge AI places some computation near data sources or users. It can appear in inspection equipment, smart gateways or embedded computers. The defining questions are how data arrives, where it is processed and how the system responds.

Entry sources reviewed: 2026-10-03

Edge and data center can cooperate

Models may be developed or updated elsewhere while local equipment handles selected inference tasks. Offline capability depends on the implementation.

  • Devices have different model, power and hardware limits.

Start the hardware chain at the interface

Sensors connect to processors, then storage, networking or control systems. Industrial environments can demand different reliability and product lifetimes from consumer electronics.

  • Boards and modules must support signal, power and thermal requirements together.

Distinction: Edge AI does not mean every task stays off the cloud, and not every edge device is a robot.

Company roles

  • Advantech · Taiwan · Example role · Edge AI equipment

    AI inference systems, modules, edge servers and gateways.

    Product / role source
  • NVIDIA · United States · Example role · Embedded AI platforms

    Jetson offers AI computing modules and development tools for edge devices.

    Product / role source

Sources

Explore this role on the mapSuggest a correctionBack to contents

Robots and Intelligent Equipment

Sensing, computation, control and mechanisms working together

Robots combine sensing, computing, control and mechanics. AI can support recognition, planning or interaction, but model outputs must reach controllers and actuators. This map emphasizes electronics and computing rather than a complete mechanical-transmission chain.

Entry sources reviewed: 2026-09-29

An AI chip does not complete the robot

Sensors, processing platforms, controllers and actuators each have distinct responsibilities. Cameras are only one sensing option.

  • Compute must fit response-time, power and thermal constraints at system level.

Development and operation use different systems

Simulation, training and deployment can use different computing environments; operating robots may use edge modules.

  • A prototype, a product feature and broad deployment represent different levels of maturity.

Distinction: Robotics and AI are not synonyms. One demonstration does not prove autonomous performance in every real environment.

Company roles

  • Techman Robot · Taiwan · Example role · Collaborative robots

    Collaborative robots integrating robot arms, vision and AI.

    Product / role source
  • HIWIN · Taiwan · Example role · Robotics and motion control

    Industrial robots, single-axis robots and motion-control components.

    Product / role source

    Sources reviewed: 2026-09-27

  • FANUC · Japan · Example role · Collaborative and industrial robots

    The CRX family comprises collaborative robots.

    Product / role source
  • NVIDIA · United States · Example role · Robotics computing and software

    Robotics AI computing, simulation and development platforms; not a robot-arm manufacturer.

    Product / role source

Sources

Explore this role on the mapSuggest a correctionBack to contents

Long-form guides (Chinese)

AI 伺服器供應鏈：GPU、HBM、板材、散熱與網路的分工

從 AI 運算需求理解伺服器供應鏈，串起 GPU、HBM、先進封裝、PCB、散熱、供電與資料中心網路。

散熱產業鏈：從導熱材料、冷板與 CDU，到 AI 資料中心

看懂氣冷、冷板液冷與浸沒式的分工，串起 TIM、熱管、均熱板、風扇、冷卻液、快速接頭、泵浦、CDU、整櫃及機房排熱。

無人機產業鏈：從共用零組件到消費、工業與軍用應用

用同一張地圖看懂無人機的機體、推進、能源、飛控導航、通訊與任務酬載，串起整機整合、三類應用和營運服務。