The computing power infrastructure industry chain will continue to benefit from the AI model arms race of overseas internet giants.
The Zhitong Finance app learned that Minsheng Securities released a research report saying that Google's strong performance in the big model arms race will, on the one hand, drive the continuous escalation of computing power demand, and at the same time bring more catalysts to the subsequent launch of big models, including GPT-5. The bank believes that the computing power infrastructure industry chain will continue to benefit under the AI model arms race of overseas Internet giants, and recommends the optical module industry chain, such as Zhongji Xuchuang (300308.SZ), Xinyisheng (300502.SZ), Tianfu Communications (300394.SZ), Shijia Photonics (688313.SH), Guangku Technology (300620.SZ), and Tengjing Technology (688195.SH).
Event: On December 6, 2023, Google Deep Mind CEO Demis Hassabis announced the official launch of Gemini 1.0, making Gemini the most powerful and versatile model ever built by Google. Gemini 1.0 is divided into three levels (sizes): (1) the Gemini Ultra version has the strongest performance and is the largest, suitable for complex tasks (2) Gemini Pro is optimized in terms of cost performance and is suitable for efficient multi-tasking (3) Gemini Nano can be deployed on edge sides such as mobile phones.
Minsheng Securities's views are as follows:
The arms race is in smoke, and Gemini's performance tests have reached new highs in a number of performance tests.
Gemini excelled in tests in various fields such as natural imagery, audio and video understanding, and mathematical reasoning. Among them, Gemini Ultra's score rate in the MMLU test (large-scale multi-tasking language understanding data set) reached 90.0%, surpassing human experts and GPT-4 for the first time (previously GPT-4's best score was 87.3%); GeminiUltra was concentrated in 32 academic benchmark tests widely used in large-scale language model development, and the performance of 30 of these test sets surpassed GPT-4. The bank believes that Google's strong performance in the big model arms race will, on the one hand, drive the continuous escalation of computing power demand, and at the same time bring more catalysts to the subsequent launch of big models, including GPT-5. Looking at the horizontal comparison, in terms of performance, GeminiUltra > GeminiPro > Geminino.
Pixel8Pro is the first to be applied, and Google products are expected to be fully upgraded.
Google product manager Brian Rakowski mentioned that the optimized Gemini can run on all devices from data centers to smartphones. Pixel8 Pro is the first smartphone to run Geminino. Currently, Geminino can be used to generate meeting minutes offline and GBoard to respond intelligently. Over the next few months, Google will launch Gemini-related services one after another, including search, ads, Chrome, DueTai, etc. Google has the most comprehensive downstream product matrix and has a large number of PC and mobile phone users. The bank believes that the launch of Gemini is expected to push the big model to achieve further breakthroughs in application.
TPUV5p was released simultaneously, catalyzing AI acceleration.
On December 6, Google simultaneously released TPUV5P, which is Google's most powerful and efficient TPU chip to date. TPUV5P is designed to improve performance, scale, and flexibility. Its training performance is 2.8 times faster than TPUV4, and the training speed of embedding intensive models is 1.9 times faster. Meanwhile, compared to TPUv4, TPUV5P has increased flops by two times and memory bandwidth by three times. The bank believes that Google, as a new force for self-developed computing power, is expected to intensify healthy competition in the computing power market, thereby reducing the cost of using computing power.
Risk warning: Overseas Internet giants' AI model capital expenses fell short of expectations, and computing power demand investment fell short of expectations.