Open source securities: Google (GOOGL.US) AI ecosystem continues to improve and is firmly optimistic about the three core main lines of “light, liquid cooling, and domestic computing power”

The Zhitong Finance App learned that Open Source Securities released a research report saying that the Google (GOOGL.US) AI model performed well, accelerated the evolution of self-developed chips and cluster solutions, and brought about incremental demand for liquid cooling. The compatibility between TPU and PyTorch is enhanced, migration costs are reduced, and the market is developed. Google released the TPUv7 and Gemini 3 series models, which have leading performance and are expected to accelerate the deployment of TPU clusters. Companies such as Anthropic and Meta plan to rent TPU to drive ASIC demand. AI is driving demand for optical fiber cables, and prices are expected to recover.

The main views of Open Source Securities are as follows:

Google's AI ecosystem continues to improve, and overseas TPU and other ASICs are expected to accelerate deployment

Google's AI model performed well, accelerating the evolution of self-developed chips and cluster solutions, and is expected to bring more incremental demand for liquid cooling. On December 18, 2025, according to Reuters, Google is promoting full compatibility between TPU and PyTorch, which is expected to drastically reduce enterprise migration costs, apply more developers, and accelerate market development. On April 9, Google released TPUv7 Ironwood. The single chip reached 4614 TeraFlops, the memory capacity was 192 GB, and the bandwidth was up to 7.2 Tbps. The single cluster could be expanded to 9216 chips, and a liquid cooling solution was used. On November 19, Google released the Gemini 3 series model, which includes two core versions: the Pro preview version for everyday applications and the Deep Think model, which integrates the three core competencies of reasoning ability, multi-modal understanding, and intelligent body ability, and performed well in multiple test benchmarks. The bank believes that the release of Gemini 3 marks an important breakthrough for Google in the field of AI applications. It not only achieves intergenerational improvements in technical performance, but also enriches the application ecosystem and accelerates the closed loop of the business model. It is expected to accelerate the deployment of TPU clusters in the future.

A number of companies have successively expressed their intention to rent TPU, continuing to drive ASIC demand

TPU is in high demand. In addition to Google's own self-developed model training such as Gemini, companies such as Anthropic and Meta are successively planning to rent TPU. On October 23, Anthropic announced a partnership with Google to deploy up to 1 million Google TPU chips to train Claude, its AI model. The expansion plan is worth tens of billions of dollars, and the estimated computing power capacity will reach 1 GW in 2026. On November 25, according to TheInformation, Meta is in discussions with Google to rent chips from Google Cloud as early as 2026 and start using Google TPU chips in its own data centers in 2027. This investment may be worth billions of dollars. Recommended targets: Invec (self-developed leader in the entire liquid cooling chain), benefiting targets: Feilong Co., Ltd., Dayuan Pump Industry, Lingyun Co., Ltd., Yinlun Co., Ltd., Tongfei Co., Ltd., Shenling Environmental, Gaolan Co., Ltd.

AI continues to drive demand for fiber optic cables, and prices are expected to recover

As the AI model continues to iterate on the training side, the implementation of inference side AI applications accelerates, and parallel strategies continue to increase the requirements for internal networks in the cluster, driving an increase in the usage of optical fiber cables, an increase in demand for DCI and urban optical cable connections in external data centers, and may drive a recovery in the price of optical fiber cables. Recommended targets: Zhongtian Technology, Hengtong Optoelectronics; Beneficial targets: Changfei Optical Fiber, Yongding Co., Ltd., Fiberfire Communications, etc.

Firmly optimistic about the three core main lines of “light, liquid cooling, and domestic computing power”, and focus on satellites and end-side AI

Recommended targets: Zhongji Xuchuang, Xinyisheng, Invek, Yuanjie Technology, Eurocom, Huagong Technology, ZTE, Shengke Communics-U, Tianfu Communications, Aofei Data, Dawei Technology, Xinyi Network Group, Ziguang Co., Ltd., Guanghetong, Zhongtian Technology, Hengtong Optoelectronics, etc.; Benefiting targets: Cambrian, Haiguang, Moore Thread-U, Inspur Information, Changguang Huaxin, Shijia Guangzi, Yongding, JPT, Zhishang Technology, Tengjing Technology, etc., Optical Bank Technology, Juguang Technology, Taichenguang, Dekley, Huilu Ecology, Jiayuan Technology, Haige Communications, Zhenyou Technology, Xinke Mobile-U, Tongyu Communications, Chaojie Co., Ltd., etc.

Risk warning: the amount of computing power chips falls short of expectations, AI development falls short of expectations, industry competition intensifies, etc.