file: 04-核心计算与通信硬件/00-概览.md
04 Core Computing and Communication Hardware Layer
1. Industry Chain Positioning and Value Distribution
The core computing and communication hardware layer sits at the absolute heart of the AI industry chain. It is currently the segment with the highest certainty, the strongest value-capture capability, and the greatest capital absorption in the entire AI wave.
On the "smile curve," this layer currently occupies an extremely left-shifted high position. Since computing power is a rigid requirement for model training, and high-end GPU supply is constrained by advanced process nodes and advanced packaging (e.g., CoWoS), supply and demand are severely imbalanced. Pricing power rests entirely with computing chip vendors led by NVIDIA (英伟达), who not only enjoy exceptionally high gross margins of 70%+, but also indirectly control the pace of downstream development (cloud service providers and large model companies) through GPU allocation quotas.
2. Core Tracks and Technical Bottlenecks
This layer is not merely a single "GPU," but rather a "computing network system" composed of three key elements: compute, storage, and communication.
- GPU / NPU (Computing Heart): The bottlenecks lie in increasing single-card compute density, breaking through memory bandwidth limits, and circumventing TSMC's CoWoS capacity constraints. In addition, the CUDA software ecosystem is NVIDIA's deepest moat; other vendors (such as AMD's ROCm or various ASIC startups) are struggling to catch up.
- HBM High-Bandwidth Memory (Data Blood): With the explosion in large model parameter counts, traditional GDDR memory can no longer meet data throughput requirements. HBM has become the standard. The current bottlenecks are HBM's extremely low yield and high cost; SK hynix currently holds an absolute advantage in this domain.
- Optical Modules and Network Switching Equipment (Nervous System): In 10K-card or even 100K-card clusters, data synchronization among GPUs is critical. 800G and even 1.6T optical modules, along with InfiniBand / advanced Ethernet switches, are key to solving the "computing island" problem. The bottlenecks here are power consumption control and optoelectronic conversion speed (silicon photonics is the future breakthrough).
3. Domestic Substitution Status and Timeline
- AI Computing Chips: Under U.S. export restrictions, domestic substitution is a must. Huawei's Ascend (昇腾) series, backed by a relatively complete CANN software stack, currently leads the domestic first tier; Cambricon (寒武纪) and Hygon (海光信息) are active in specific markets. The challenge is that advanced process nodes are restricted, so a "curve overtaking" strategy must be pursued through advanced packaging or chiplet technologies.
- Optical Communications: This is the only segment where Chinese companies hold absolute global dominance in the industry chain. Zhongji Innolight (中际旭创) and Eoptolink (新易盛) together account for more than 50% of the global 800G optical module market share. There is no obvious "bottleneck" issue here; instead, these companies are enjoying the dividends of global computing power expansion.
4. Cross-Layer Linkage Effects
- Upward Linkage (Model Layer): The evolution of model architectures (e.g., the rise of Mixture-of-Experts / MoE architectures) directly changes the demand ratio for memory capacity (HBM) and communication bandwidth (optical modules).
- Downward Linkage (Manufacturing & Materials Layer): GPU iteration speed is constrained by the yield of TSMC's 3nm/2nm processes and the delivery schedule of ASML's next-generation High-NA EUV lithography systems.
5. Benchmark Companies in This Layer
The following are the dominant or representative core players in this layer globally and domestically. These figures are automatically synchronized and rendered from the company database:
Cerebras
Unlisted | United StatesCore Business: Develops the world's largest AI chip—the Wafer-Scale Engine (WSE)—where an entire wafer is a single chip.
⚔️ Moat: The wafer-scale chip (Wafer-Scale Engine) technology route is unique—the entire wafer is not diced but directly used as a single superchip, providing efficiency far exceeding that of traditional GPUs in certain AI training scenarios.
Key Competitors:
HiSilicon
Unlisted | ChinaCore Business: Develops the Ascend series of AI processors and full-stack software/hardware AI solutions, serving as the absolute main force in domestic substitution for AI computing chips.
⚔️ Moat: The only company in China with full-stack self-developed AI capabilities (chips + CANN operator library + MindSpore framework). Despite the most stringent US export controls, it has maintained product iteration and built the largest AI chip ecosystem in China.
Key Competitors:
Iluvatar CoreX
Unlisted | ChinaCore Business: General-purpose GPU chip design with self-developed architecture, covering all AI training and inference scenarios
⚔️ Moat: Self-developed TianGai 100/ZhiKai 100 architecture + comprehensive adaptation to domestic large models such as DeepSeek + deep integration with government and enterprise customers
Key Competitors: Hygon, Moore Threads, Cambricon, Huawei Ascend
Cambricon
688256:SH | ChinaCore Business: Develops cloud and edge AI chips and supporting foundational system software platforms; it is a rare independent AI chip architecture design company in China.
⚔️ Moat: Possesses a self-developed AI instruction set architecture and independent underlying innovation capabilities in NPU chip design. It holds a place in the non-Huawei domestic AI chip market and has been deployed in multiple AI computing center projects.
Key Competitors:
Moore Threads
Unlisted | ChinaCore Business: Develops domestic full-featured GPU and AI computing chips, providing complete solutions from graphics rendering to AI acceleration.
⚔️ Moat: Self-developed MUSA architecture and complete software stack, achieving CUDA compatibility and rapid adaptation of domestic large models, forming a software-hardware collaborative ecosystem barrier.
Key Competitors: Cambricon, Hygon, Biren Technology, Jingjia Micro
Hygon Information Technology Co., Ltd.
688041:SH | ChinaCore Business: Development and sales of x86-based general-purpose CPUs and AI-oriented DCU co-processors
⚔️ Moat: x86 instruction set compatibility + independent iteration capability, enabling seamless adoption of the mature software ecosystem
Key Competitors: Intel (x86), AMD (x86), Huawei Ascend, Cambricon, Zhaoxin, Phytium
NVIDIA
NVDA:US | United StatesCore Business: High-performance GPU computing platform and software ecosystem for AI training and inference.
⚔️ Moat: CUDA software ecosystem + leading architecture iteration + deep integration with TSMC advanced packaging, forming a triple moat across hardware, software, and manufacturing.
Key Competitors: AMD (MI series), Intel (Gaudi series), Google (TPU), Huawei (Ascend)
NVIDIA
NVDA:US | United StatesCore Business: The world's largest AI chip and computing platform company, designing high-performance GPUs and providing the CUDA unified computing platform, as well as data center networking solutions based on InfiniBand/Spectrum-X.
⚔️ Moat: The CUDA software ecosystem has extremely deep barriers, with millions of developers creating high switching costs; the co-design of hardware and software systems (GPU+NVLink+InfiniBand) delivers unparalleled cluster performance advantages, resulting in de facto ecosystem monopolization.
Key Competitors:
Core Business: In-house TPU chips and cloud AI compute services
⚔️ Moat: TPUs are deeply integrated with Google's AI software stack, owning one of the world's largest TPU clusters and forming an extremely high ecosystem barrier
Key Competitors: NVIDIA, AMD, Intel, Amazon (Trainium), Microsoft (Maia)
Core Business: Self-developed AI chips (TPU) and the world's largest AI computing infrastructure and cloud services
⚔️ Moat: Deep coupling of self-developed TPUs with TensorFlow/JAX, backed by the world's largest AI clusters and data center network, builds an integrated hardware-software ecosystem moat
Key Competitors: NVIDIA, AMD, Intel, Amazon, Microsoft
AMD
AMD:US | United StatesCore Business: The only company globally that competes head-on with Intel and NVIDIA in both CPU and GPU markets. Provides data-center-grade MI series AI accelerator cards and EPYC server CPUs.
⚔️ Moat: Possesses extremely strong chip design capabilities (CPU+GPU integration), making it the only effective choice for cloud providers to counterbalance NVIDIA. The ROCm software ecosystem is rapidly catching up with CUDA, offering significant cost-performance advantages.
Key Competitors:
Tongfu Microelectronics Co., Ltd.
002156:SZ | ChinaCore Business: Provide advanced chip packaging and testing services, focusing on Chiplet 2.5D/3D packaging technology
⚔️ Moat: One of the few domestic manufacturers capable of large-scale Chiplet packaging mass production, deeply tied to top customers such as AMD, with processes covering FCBGA, 2.5D TSV, and 3D stacking
Key Competitors: JCET, Huatian Technology, TSMC (CoWoS)
Marvell
MRVL:US | United StatesCore Business: Provides semiconductor solutions for data infrastructure, including PAM4 DSP (core chips for optical modules), Ethernet switching chips, and custom AI ASIC foundry services for cloud providers.
⚔️ Moat: It holds a dominant position in the high-speed optical module core chip (PAM4 DSP) market; it also has strong customization capabilities for advanced-process chips and is an important partner for cloud providers such as Amazon in developing their own chips.
Key Competitors:
TSMC
2330:TW / TSM:US | Taiwan, ChinaCore Business: Globally leading semiconductor wafer foundry and advanced packaging service provider
⚔️ Moat: World-leading advanced process technology, exclusive CoWoS packaging supply to AI chip giants, and extremely strong customer stickiness
Key Competitors: Samsung Electronics, Intel, UMC
Qualcomm
QCOM:US | United StatesCore Business: Designs, develops, and supplies wireless communication chips and mobile processing platforms, along with technology patent licensing.
⚔️ Moat: Globally leading CDMA/5G communications patent portfolio + Snapdragon mobile platform ecosystem, forming a software-hardware integrated edge AI computing moat.
Key Competitors: Apple (in-house chips), MediaTek, Samsung Exynos, Intel (modems)