According to a research report by CICC, the launch of Google's (GOOGL.US) TPUv7 marks the acceleration of ASIC clusters in self-development on architectures distinct from traditional GPGPUs, leading to heterogeneous restructuring and reshaping of hardware value. This development is expected to drive simultaneous growth in both volume and pricing for AI computing hardware markets, including PCBs, liquid cooling, and power supplies. By 2027, the AI PCB, liquid cooling, and power supply chip markets are projected to reach $21.65 billion, $20.18 billion, and $18.39 billion, respectively.
Key stocks to watch include Shennan Circuits (002916.SZ), Shengyi Technology (600183.SH), Dongshan Precision (002384.SZ), Avary Holding (002938.SZ), Foxconn Industrial Internet (601138.SH), FIT Hon Teng (06088), and Azure Lithium Core (002245.SZ).
CICC's key insights are as follows:
**A Decade of Google TPU Architecture Evolution** Since Google first unveiled TPUv1 in 2016, the architecture has evolved significantly over ten years. From a systolic array designed for inference, TPUs have advanced to training chips capable of supporting nearly 10,000-card clusters, incorporating OCS optical switching architecture and HBM high-bandwidth memory. The latest TPUv7 further breaks through with a dual-chiplet packaging architecture, significantly improving linear acceleration ratios in ultra-large-scale clusters.
**Key Features of Google’s Next-Gen TPUv7** - **Tray Architecture**: Comprises 16 standardized computing trays, each holding four TPU chips. - **Power Architecture**: Adopts a +/-400V high-voltage direct current (HVDC) solution. - **Cooling System**: Features a 100% liquid-cooled design with large cold plates covering four TPUs and VRMs. - **Cluster Scale**: Supports up to 144 interconnected racks, equivalent to a 9,216-TPU cluster.
CICC estimates the total value of Google’s TPUv7 cabinet solution at approximately $730,000, with TPUs, PCBs, liquid cooling, power supplies, and cables contributing $544,000, $40,000, $70,000, $71,000, and $4,000, respectively. As TPU shipments grow and product iterations drive hardware demand, the PCB, liquid cooling, and power supply chip markets are expected to see substantial volume and price increases.
Based on Google’s procurement estimates, the AI PCB, liquid cooling, and power supply chip markets could reach $3.69 billion, $6.06 billion, and $3.1 billion, respectively, by 2027. Considering overall GPU and ASIC demand, these markets may expand to $21.65 billion, $20.18 billion, and $18.39 billion in the same timeframe.
**Risks**: Potential shortfalls in AI and computing infrastructure demand, intensifying industry competition, and supply shortages of critical components.
Comments