Skip to content
Core Subtopic 1: Computing Power Core Layer – The Fundamental Role of GPU and AI Servers
Definition and Explanation
- GPU (Graphics Processing Unit): The core engine of AI computing power, processing massive amounts of data through a parallel computing architecture, serving as the “brain” for AI training and inference.
- AI Server: A hardware platform that integrates components such as GPU, storage, and networking, serving as the physical carrier of computing power output, designed to meet high-density computing demands.
Key Facts and Trends
- Explosion in GPU Computing Power Demand: The global AI chip market is expected to reach $53.6 billion in 2023, projected to grow to $92 billion by 2025, with a compound annual growth rate exceeding 30%. NVIDIA dominates the market, while domestic chips (such as Cambricon) accelerate independent research and development to promote alternatives.
- Integration Trend of AI Servers: A single AI server can accommodate 8-16 GPUs (such as NVIDIA H100 clusters), with increased computing density driving hardware architecture upgrades from “general-purpose” to “specialized”. Companies like Foxconn have become core suppliers for leading cloud providers.
Real-World Cases
- Internet giants like ByteDance are purchasing AI servers in large quantities, driving growth in related supply chains (such as ASIC chips and power modules).
Core Subtopic 2: Data Transmission Layer – The High-Speed Interconnection Value of CPO
Definition and Explanation
- CPO (Co-Packaged Optics): Co-packaging optical modules with switch chips to reduce signal loss and enhance interconnection rates within GPU clusters in data centers, addressing the “computing power-transmission” bottleneck.
Key Facts and Trends
- Accelerated Rate Iteration: Optical modules are upgrading from 400G to 800G and 1.6T, with global demand for 800G optical modules expected to exceed 20 million units by 2025, becoming a mainstream configuration.
- Technological Route Competition: CPO and LPO (Linear Direct Drive Optical Modules) coexist, with CPO being more suitable for AI computing clusters due to its low latency and low power consumption advantages, despite higher costs; domestic manufacturers (such as Aoyuan Guangtong) enhance competitiveness through customized solutions.
Controversy Points
- In the short term, copper cables may partially replace CPO due to cost advantages, but in the long run, as GPU computing power doubles, CPO remains the core technology to meet high bandwidth demands.
Core Subtopic 3: Hardware Connection Layer – The “Nervous Center” Function of PCB
Definition and Explanation
- PCB (Printed Circuit Board): The physical carrier connecting components such as GPU, CPO, and power supply, responsible for signal transmission and power distribution, with its density and reliability directly affecting AI server performance.
Key Facts and Trends
- Surge in Demand for High-End PCBs: AI servers drive demand for high-density interconnect boards (HDI), IC substrates, and other high-end products, benefiting the PCB industry from increased unit value (the value of a standard server PCB is about $500, while AI server PCBs can exceed $2000).
- Industry Chain Rotation Characteristics: As a hardware foundation, PCBs exhibit significant performance elasticity during the AI computing expansion cycle, but are subject to rotation adjustments influenced by market sentiment (e.g., the sector is in an adjustment phase in December 2024, while long-term demand expectations remain unchanged).
Core Subtopic 4: Cooling Assurance Layer – The Essential Position of Liquid Cooling Technology
Definition and Explanation
- Liquid Cooling Technology: Efficiently removes heat generated by high-power components such as GPUs using liquid (water or specialized coolant), divided into cold plate (indirect contact) and immersion (direct contact) solutions.
Key Facts and Trends
- Rapid Increase in Penetration Rate: GPU power consumption has surged from 300W (A100) to over 700W (Blackwell Ultra), making liquid cooling a necessity rather than an option, with direct liquid cooling and immersion solutions accounting for over 60%.
- Investment Hotspots in the Industry Chain: Liquid cooling servers are in the “0→1” penetration stage, with high technical barriers in upstream cooling liquids, CDU (Cooling Distribution Unit), and cold plates, accounting for over 50% of value. Companies like Highlan have entered the ByteDance supply chain.
Core Subtopic 5: Cross-Layer Synergy – Technological Interaction and Industrial Resonance
Definition and Explanation
- Synergy Mechanism: The closed loop formed by GPU computing power, CPO transmission, PCB connection, and liquid cooling heat dissipation creates a cycle of “computing power generation – data transmission – structural support – heat dissipation stabilization”; any shortcoming in one link will limit overall performance.
Key Facts and Trends
- Synergistic Iteration of Computing Power, Transmission, and Cooling: For every doubling of GPU computing power, the CPO transmission rate must be upgraded simultaneously (e.g., from 800G to 1.6T), and liquid cooling power must increase in tandem, forming a “triple constraint” relationship.
- Industry Rotation Logic: The AI hardware sector exhibits a rotation characteristic of “GPU/Optical Modules → PCB → Liquid Cooling”, with liquid cooling becoming the next strategic hotspot following optical modules due to its low penetration rate and non-excessive supply side.
Recommended Resources
- “China AI Chip Industry White Paper” (China Academy of Information and Communications Technology, 2024)
- “Liquid Cooling Server Technology and Market Analysis” (ZOL News, 2025)
- “CPO Technology Evolution and Data Center Applications” (Aoyuan Guangtong Technical White Paper, 2025)
- Xueqiu Column: “In-Depth Analysis of AI Hardware Industry Chain Rotation Logic” (December 2024)
- NetEase Finance: “Comprehensive Overview of Leading Companies in the Computing Power Industry Chain” (August 2025)
Intelligent Summary (Executive Briefing)
- GPU is the Core Engine: The demand for AI computing power drives the GPU market size to double in three years, with domestic substitution and international competition coexisting.
- AI Server Integration is Key: The number of GPUs per machine has increased to 16, driving hardware architecture towards specialization, significantly benefiting supply chain manufacturers (such as Foxconn).
- CPO Solves Transmission Bottlenecks: Demand for 800G optical modules is expected to exceed 20 million units by 2025, with CPO being a long-term technological direction, though facing short-term pressure from copper cable alternatives.
- Liquid Cooling Becomes a Necessity: The high power consumption of GPUs has made liquid cooling a standard requirement, with upstream components like cooling liquids and CDU having high barriers and elasticity.
- Synergistic Iteration is Key: Computing power, transmission, and cooling form a “triple constraint”; the industry chain rotation is shifting towards low penetration rate segments (such as liquid cooling), necessitating attention to technological compatibility and supply-side dynamics.