With the landing of AI applications at the terminal, rapid growth in AIoT demand is expected; as the core chip, SOC technology innovation and AI demand are likely to drive both the volume and price of AI SOCs. The rich visual downstream application scenarios are expected to drive the demand for visual SOCs and CIS.
According to Zhito Finance APP, Tianfeng Securities published a research report stating that recently, Volcano Engine released a new upgrade of its Doubao large model family, including Doubao general model, voice large model, and visual large model, with significant performance improvements; in terms of new models, the Doubao visual understanding model was released, and the average price reduction is expected to facilitate rapid adoption. The input price for Doubao visual understanding is 0.003 yuan per thousand tokens, which is 85% lower than the industry average, equivalent to processing 284 720P images for one yuan, marking the formal entry of visual understanding models into the era of subdivisions. With the landing of AI applications at the terminal, rapid growth in AIoT demand is expected; as the core chip, SOC technology innovation and AI demand are likely to drive both the volume and price of AI SOCs. The rich visual downstream application scenarios are expected to drive the demand for visual SOCs and CIS.
Tianfeng Securities' main points are as follows:
The newly released Doubao visual understanding model is expected to rapidly gain popularity due to the reduced average price.
Through the Doubao visual understanding model, users can simultaneously input text and image-related questions. The model can comprehensively understand and provide accurate answers. This will greatly simplify the application development process and unlock more value scenarios for the large models. The visual understanding capability will greatly expand the scenario boundaries of large models, opening up vast application prospects in finance, medical, construction, geography, sports, logistics, and more.
The input price for Doubao visual understanding is 0.003 yuan per thousand tokens, which is 85% lower than the industry average, equivalent to processing 284 720P images for one yuan, marking the formal entry of visual understanding models into the era of subdivisions. At the same time, Volcano Engine will provide higher initial traffic, with RPM reaching 15,000 times and TPM reaching 1.2 million, enabling enterprises and developers to effectively utilize the visual understanding model and discover more innovative scenarios.
The capabilities of the Doubao large model have been enhanced further, and the daily usage is expected to continue to grow.
According to data from Volcano Engine, the daily average token usage of the Doubao large model in December exceeded 4 trillion, which is an increase of over 33 times since its release in May; as the token calling scale increases, the Doubao large model is also rapidly growing in different scenarios. At this conference, Volcano Engine released and upgraded several Other models. Among them, the Doubao general model pro completed a new version iteration, with comprehensive task processing capabilities increased by 32% compared to May, inference improved by 13%, instruction following improved by 9%, code improved by 58%, and mathematical capabilities improved by 43%, while expertise in professional knowledge fields improved by 54%.
The millet economy combined with SOC has vast AI + Hardware application scenarios.
The rise of generative large models is prompting players to explore new gameplay, leading to applications featuring figurines interacting and dialoguing with humans; AI + is expected to become the driving force behind complementary IP souls, allowing IP to not only remain flat or static but to be presented in a more vivid form. According to Transparency Market Research, by 2026, the Global market size for smart toys such as electronic pets, robots, and smart train sets is expected to grow to nearly 70 billion dollars.
The release of AI terminals + technological upgrades drives both the volume and price of AI SOC to rise.
According to MarketsandMarkets, the SoC market size benefits from the demand for AI and machine learning, and is expected to grow from 138.46 billion dollars in 2024 to 205.97 billion dollars in 2029, with an expected CAGR of 8.3%. Taking RayBan Meta as an example, according to Wellsenn XR data, the total cost of RayBan Meta is 174 dollars, of which the cost of the Main Board chip is approximately 99.1 dollars, accounting for more than half of the cost; compared to the previous generation RayBan Stories, the Main Board chip cost has increased by 28.7%. With the drive of technological innovation, the value of the Main Board chip is expected to continue to grow.
It is recommended to pay attention to the symbols.
AI glasses SOC: Bestechnic (Shanghai) Co., Ltd. (688608.SH), Zhongke Lanxun (688332.SH), Juxin Technology (688049.SH), etc.
AI home SOC: Rockchip Electronics (603893.SH), Amlogic (688099.SH), Allwinner Technology (300458.SZ), etc.
AI Vision SOC: Star Technology (301536.SZ), Shanghai Fullhan Microelectronics (300613.SZ), Ingenic Semiconductor (300223.SZ), etc.
AI Toy SOC: Espressif Systems (688018.SH), Longsys (688486.SH), Aojie Technology (688220.SH), Tailin Micro (688591.SH), Beken Corporation (603068.SH), etc.
Supporting Storage/CIS/Solutions: GigaDevice Semiconductor Inc. (603986.SH), Puran (688766.SH), Dongxin (688110.SH), Bestechnic (Shanghai) Co., Ltd. (688416.SH), Will Semiconductor (603501.SH), Sitwell (688213.SH), GeKo Micro (688728.SH), Shanghai Fortune Techgroup (300493.SZ), etc.
Risk Warning: Downstream demand may fall short of expectations, technology upgrades and iterations may not meet expectations, new product promotion may not meet expectations, supplier concentration risk, customer concentration risk, etc.
