I. Technical Breakthrough: Dual Strategy of Ecological Compatibility and Domestic Substitution
1. Core Technical Parameters and Performance
Alibaba’s latest AI chip is an upgraded version of the Hanguang 800, positioned as a dedicated inference chip, utilizing domestic 5nm process technology (manufactured by SMIC), with key breakthroughs in the following areas:
Computing Power and Energy Efficiency: In inference tasks involving large models with hundreds of billions of parameters, the latency is only 70% of that of NVIDIA’s A100, and power costs are reduced by 50%. Tests show that generating a single image with Stable Diffusion takes 1.8 seconds, which is 40% cheaper than NVIDIA’s solution.
Ecological Compatibility: A translation layer for instructions achieves seamless compatibility with the CUDA ecosystem, allowing developers to migrate applications without modifying code, with performance loss controlled within 8%. This design significantly lowers the technical barrier for switching from NVIDIA platforms, especially suitable for enterprises needing rapid deployment of AI inference services.
Process and Packaging: Utilizing SMIC’s 5nm process, the yield has improved from an early 55% to 75%; the packaging process is handled by JCET using Chiplet technology, integrating HBM3 memory from Changxin Technology, achieving 100% autonomy from wafer to packaging.
2. Differentiated Competition in Technical Routes
Unlike Huawei’s Ascend, which relies on a fully self-developed ecosystem (requiring adaptation to the CANN architecture), Alibaba has chosen a pragmatic route of CUDA compatibility and domestic manufacturing:
Market Positioning: Focused on the inference market (which accounts for over 70% of AI computing power demand), avoiding the high-barrier competition of training chips.
Cost Advantage: By reducing dependence on TSMC through domestic supply chains, the chip price is 30% lower than NVIDIA’s H20, and there are no CUDA licensing fees.
Future Layout: Simultaneously developing a 4nm training chip (targeting 500 TOPS computing power) and a photonic AI chip (which is 1000 times faster than GPUs), forming a full-stack technology matrix for “frontier exploration in inference and training”.
3. Application Scenarios and Commercial Progress
Internal Applications: Deployed in Alibaba Cloud servers, supporting core businesses such as Taobao product image classification (processing speed reduced from 1 hour to 5 minutes) and intelligent customer service.
External Output: Providing computing power services to enterprise clients through elastic bare metal instances, with rental costs 30% lower than NVIDIA instances. By June 2024, a leading server manufacturer has secured orders for hundreds of thousands of Alibaba chips.
II. Core Beneficiary Companies: Full Industry Chain Coverage from Design to Application
1. Manufacturing and Packaging
SMIC (688981): Responsible for manufacturing Alibaba’s 5nm chips, with expected revenue contribution reaching 12% by 2025. Its N+2 process (equivalent to TSMC’s 7nm) has been validated by Alibaba and may be used for the production of 4nm training chips in the future.
JCET (600584): Exclusively provides Chiplet packaging services, achieving a yield of 99.999%, with monthly production capacity exceeding 5 million units. Its 2.5D/3D packaging technology can improve chip energy efficiency by over 20%, directly benefiting from Alibaba’s demand for high-end chips.
2. Design and Tool Support
Chipone (688521): Provides Alibaba with RISC-V IP cores and one-stop design services, with jointly developed GPU IP already applied in the Hanguang chip. In 2024, its RISC-V related business revenue is expected to grow by 45% year-on-year.
Huada Empyrean (301269): As a leading domestic EDA provider, its tools support the design of Alibaba’s 5nm chips, with post-simulation testing efficiency improved by 35 times compared to traditional solutions. The Empyrean ALPS® series tools have been standardized for SMIC’s 28nm production line.
3. Server and Computing Power Output
Inspur (000977): Integrates Alibaba’s chips to launch AI server product stacks, with order volume in Q3 2024 increasing by 180% year-on-year. Its NF5488M6 server, equipped with the Hanguang NPU, offers three times the inference performance compared to traditional solutions.
Sugon (603019): Collaborates with Alibaba to develop high-performance computing clusters for large model training. Its liquid cooling technology reduces PUE to as low as 1.15, with single project orders reaching 15 billion yuan.
4. Ecosystem and End-User Applications
Runhe Software (300339): Jointly releases multiple RISC-V based development platforms with Alibaba, achieving large-scale applications in smart grids and industrial internet. In 2024, its OpenHarmony related business revenue share is expected to increase to 35%.
Haier Smart Home (600690): Utilizes Alibaba’s AIoT chips to create a smart home ecosystem, with voice interaction response times for products like refrigerators and air conditioners improved to 0.3 seconds, and sales of related product lines expected to grow by 60% in 2024.
5. Supporting Components and Service Providers
Invec (002837): Provides precision temperature control solutions for Alibaba’s data centers, with single cabinet heat dissipation power reaching 25kW, saving 40% energy compared to traditional solutions. In 2024, its data center business revenue is expected to grow by 58% year-on-year.
DataPort (603881): As a deeply integrated IDC operator with Alibaba Cloud, has deployed over 100,000 Hanguang chip servers to support Alibaba’s internal computing power needs.
III. Market Impact and Risk Warnings
1. Replacement Effect on NVIDIA
The CUDA compatibility of Alibaba’s chip makes it a direct competitor to NVIDIA’s H20. Estimates suggest that if Alibaba’s chip captures 20% of the Chinese inference chip market by 2025 (approximately 12 billion yuan), it will lead to a 15% reduction in NVIDIA’s related revenue.
2. Synergistic Effects of Domestic Industrial Chain
Equipment and Materials: Upstream companies such as Northern Huachuang (etching equipment) and Shanghai Silicon Industry (silicon wafers) will benefit from SMIC’s capacity expansion.
Policy Support: The National Big Fund Phase II has invested in SMIC and JCET, with expectations that the proportion of domestic equipment procurement will increase to 70% by 2025.
3. Potential Risks
Technology Iteration Risk: NVIDIA plans to launch the H30 inference chip in 2025, with a 50% performance improvement over the H20, which may weaken the competitiveness of Alibaba’s chip.
Ecosystem Development Challenges: Despite CUDA compatibility, Alibaba needs to continuously optimize its software toolchain (such as RatelNN) to attract developers and avoid repeating Huawei Ascend’s ecological dilemma.
Geopolitical Risks: If the U.S. further restricts HBM exports, Alibaba may face storage performance bottlenecks and needs to accelerate the mass production process of HBM3 at Changxin Technology.
Disclaimer:
1. I write this public account solely to record my trading insights and share with everyone. Please do not blindly follow my trades. Articles not marked as [Original] are not written by me and are usually reposted, and do not represent my views.
2. My trading style changes with the market, and this public account can only post one article after the market closes each day. I need to respond flexibly during trading hours, so mentioning individual stocks does not constitute any investment advice!
3. Investment involves risks; please proceed with caution.
If you like it, please click below【Looking】