In 2025, China opened a good year.
In the cultural market, &#34;Ne Zha 2&#34; exploded in popularity, with box office earnings surpassing 10 billion, entering the global movie TOP list, showcasing China's &#34;terrifying&#34; consumer capability to the world. In the AI market, the emergence of DeepSeek, achieving Chat GPT-like effects with lower computing power, directly topped global trending lists.
If the emergence of Chat GPT brought generative AI to the cloud, then DeepSeek has brought generative AI to the edge.
<h2>The &#34;Golden Turning Point&#34; of Edge AI Chips</h2>
The tech industry has been exploring AI hardware products.
From this year's &#34;Spring Festival Gala of Consumer Electronics,&#34; CES, it is very clear. The products launched by major companies at this year's CES are all strongly related to AI, whether it is general PCs, mobile terminals, or robots, glasses, headphones, watches, etc. Almost every product that involves human-computer interaction has manufacturers attempting to combine it with AI large models.
However, these explorations seem to be not very successful. Hundreds of thousands of large models have been launched, and there are even more hardware products, but the results are not very clear. The reason lies in the fact that the computing power provided by local SoCs cannot match the computing power required by large models.
In the ChatGPT era, the distillation of large models is a significant issue. If cloud-based computing power is used to complete AI inference, there are three problems: First, cost. Any operation in the cloud incurs costs, which may not be much, just a few cents in RMB, but it still costs money.
Some mobile manufacturers have revealed that the average cost of calling a cloud-based large model is between 1.2 to 1.5 cents in RMB. Assuming each brand has hundreds of millions of users, and each person calls it 10 times a day, the computing power cost is astonishing. However, if users are charged by the number of calls or monthly subscriptions, under conditions of functional homogeneity, users' willingness to use it will also be difficult to guarantee.
Second, speed. Since it is completed in the cloud, it inevitably requires network transmission, which leads to slow response times for AI at the terminal. In the case of autonomous driving, decisions need to be made within 10 milliseconds based on the current environment, relying on the cloud is overly dangerous. In scenarios like autonomous driving and industrial quality inspection, edge inference latency can be reduced to the millisecond level, improving 5 times compared to cloud solutions.
Third, privacy. This is the most important part, involving content related to healthcare and finance. For users, edge models are the optimal solution.
Everyone has an answer in their hearts: Edge AI is the key to the landing of AI hardware.
The deployment of small-sized models at the edge has already begun.
Since Huawei announced the integration of the &#34;Xiao Yi Intelligent Agent&#34; with DeepSeek-R1, within just over a week, six mobile manufacturers, including Starry Meizu, Honor, OPPO, Nubia, and vivo, have announced their integration with DeepSeek.
It should be explained that the full version of the DeepSeek-R1 model has parameters reaching 671B, and the model file alone requires 404GB of storage space, which no mobile device can meet such hardware configuration needs
However, the distilled versions of DeepSeek (1.5B, 7B) are suitable for use on edge devices such as mobile phones.
To ensure that these small-sized models run smoothly on edge devices and fully leverage their intelligent advantages, powerful edge AI chips are needed to provide computational support.
The market demand for edge AI chips that can accommodate small-sized models is rising sharply.
<h2>Main Players in the Field</h2>
The application market for edge devices is very large. As mentioned earlier, as long as there are hardware terminals for human-computer interaction, manufacturers are attempting to combine them with large AI models. Let's take a closer look here:
In the AI PC field, by 2027, AI PCs are expected to account for 85% of the PC market in China; in the AI mobile phone field, by 2026, the shipment of AI mobile phones is expected to exceed 470 million units, with a penetration rate increasing to 38%; in the AI wearable device field, the market size is expected to grow from USD 41.9 billion in 2024 to USD 120.7 billion in 2028, with a CAGR of 30.3%.
In 2023, the scale of China's edge AI market is 193.9 billion yuan, with an average annual compound growth rate of 116.3% from 2018 to 2023.
DeepSeek primarily drives the demand for AI edge SoC chips.
SoC chips are the main control units of various types of hardware devices, carrying core functions such as computational control, and are the &#34;brains&#34; of the hardware. As AI applications on the edge become more widespread, SoCs will increasingly evolve into system-level chips that integrate artificial intelligence and edge computing capabilities, becoming AI SoCs with computational power reaching tens or even hundreds of TOPS.
In this wave of edge AI, Rockchip has frequently hit the daily limit.
Currently, Rockchip can provide AIoT chips with different computational power levels ranging from 0.2 TOPS to 6 TOPS, among which the RK3588 and RK3576 come with a 6 TOPS NPU processing unit, capable of supporting the deployment of mainstream edge models with parameter levels of 0.5B to 3B. These can perform functions such as translation, summarization, and Q&amp;A through large language models, and can achieve multimodal search and recognition, effectively addressing pain points in various AIoT scenarios and enhancing product user experience.
Among them, the company's flagship SoC chip RK3588M is one of the few domestic smart cockpit SoC chips that can compete with first-line products abroad.
According to Rockchip, this product has excellent performance, supports multiple screens with one chip, and has outstanding edge AI capabilities. It has already been applied in numerous leading automotive manufacturers, with over 10 mass-produced models and more than 20 targeted model projects under simultaneous development. Additionally, the new product RK3576M is also in the process of customer onboarding.
This is just one aspect of Rockchip's edge AI product applications.
In fact, there are already multiple clients in various fields developing new hardware that supports large AI models on the edge based on Rockchip's main control chips, such as educational tablets, AI toys, desktop robots, computing terminals, and conference hosts.
Allwinner Technology is also a highly regarded SoC company. Last year, the company's net profit attributable to shareholders of the listed company reached 153 million to 190 million yuan, a year-on-year increase of 566.29% to 727.42%. The surge in performance is attributed to a significant increase in product shipments represented by business lines such as robotic vacuum cleaners and smart projectors, resulting in a year-on-year revenue growth of approximately 35%.
When asked by reporters whether Allwinner Technology's products can be adapted to DeepSeek and whether they are planning for DeepSeek, Allwinner Technology stated: &#34;The company's products can provide computing power support for various forms of intelligent terminal products on the edge.&#34;
Espressif Systems' SoC has long been applied in the general IoT field. From the application side, Espressif Systems has achieved over 30% growth in core application markets such as smart home, smart lighting, and consumer electronics.
Both the ESP32-S3 and ESP32-P4 product lines from Espressif Systems have added edge AI capabilities, mainly reflected in voice wake-up and control, as well as image processing functions on the device side. These two series of chips have increased AI acceleration instructions in hardware design; on the software side, they also provide solutions for image recognition and voice wake-up and control.
Wang Jue, Deputy General Manager of Espressif Systems, stated: &#34;The AIoT chip ESP32-S3 with edge AI capabilities is currently growing very rapidly and is also the flagship product being promoted.&#34; The chip ESP32 from Espressif is also used in ByteDance's AI toy &#34;Eye-catching Bag.&#34;
Amlogic has over 15 commercial chips equipped with its self-developed edge AI computing units, with shipments of chips carrying self-developed edge AI computing units expected to exceed 8 million in 2024.
In 2024, the company expects to achieve an operating income of approximately 5.921 billion yuan; the net profit attributable to the parent company is expected to be around 820 million yuan, a year-on-year increase of about 64.65%. The company stated that its 6nm chip S905X5 series can utilize edge AI capabilities to achieve functions such as local simultaneous translation and real-time subtitles, and has received orders from multiple international top-tier operators since its commercial launch six months ago. The 6nm chip is expected to achieve sales of over 10 million units by 2025.
The application scenarios for edge AI audio processors are mostly in the smart IoT field. For example, in smart speakers, edge AI audio processors can support voice wake-up, voice recognition, and voice synthesis functions, enabling natural language interaction between users and the speaker. In smart home systems, it can be used to control household appliances with sound, such as adjusting lighting, air conditioning, and television through voice commands.
Hengxuan Technology's edge SoC has been successfully integrated into products from several mainstream brands, including Baidu, ByteDance, Google, Harman, Anker Innovations, Edifier, and Shokz.
This year, ByteDance's first smart headphones Ola Friend, equipped with the Doubao large model, use the Hengxuan Technology 2700 chip.
** The latest chip from Hengxuan Technology, BES2800, has also been applied in Samsung's newly released Galaxy Buds3 Pro headphones for 2024. The application of multiple headphone models demonstrates Hengxuan Technology's leading position in the field of smart terminal SoC chips.
Regarding the changes in chip requirements for edge-side AI, Hengxuan Technology believes: &#34;The rise of cloud large models not only drives AI mobile phones and PCs but wearables will also benefit from the development of edge-side AI. It will pose new demands on chips, such as the need for stronger environmental perception capabilities in wearables, so the computing power of the main control chip needs to be correspondingly improved. At the same time, long battery life is a necessity for wearable products, so chips must maintain a low power consumption level while enhancing computing power.&#34;
It is also worth mentioning the AI glasses chip. The industry generally believes that glasses are currently the hottest carrier for AI and large models, and this year the industry will achieve a breakthrough from 0 to 1.
Previously, in 2017, Meta began developing AI glasses, initially in collaboration with Samsung, but after three years with no results, the project was abandoned. Subsequently, Meta gave up on self-developed chips, and Ray-Ban glasses used Qualcomm's AR1 Gen 1 chip.
It can be said that Qualcomm still holds an absolute control position in the ARVR chip field. In the market, the chips available for AI glasses, apart from Qualcomm chips, are Unisoc's TW517.
Unisoc TW517 uses a 12nm process, with a GPU model of IMG8300, and a running frequency of 800MHz. Its main customers are Shanji AI Paitai Mirror and Yingmu Technology.
At the same time, Hengxuan Technology previously revealed that its chips have been applied and released in smart glasses products such as Meizu, and some customer projects are in the onboarding stage.
<h2>Conclusion</h2>
Nezha 2 has surpassed 10 billion in box office, showcasing the &#34;terrifying&#34; consumption capacity of the Chinese market to the world. This means that the scale and consumption capacity of the Chinese market alone are equivalent to the total of all other developed countries globally.
The signal conveyed behind this is very positive.
If we only look at the film market, it may not be fully understood. But if we consider it in the context of technology and manufacturing, the immense value becomes clear.
Moving forward, no matter how Europe and the United States impose trade and technological barriers against us, they cannot stop our industrial upgrade. The biggest obstacle to technology and manufacturing is not the &#34;neck choke,&#34; but rather the lack of a sufficiently large consumer market to support the establishment of related industrial chains.
For example, if a company develops a chip, the cumulative cost from design to mass production is 200 million. If only 100,000 pieces are sold, the cost per chip would be 2,000 yuan, which is too high for the company to make a profit. It would inevitably fall into long-term losses and eventually go out of business. However, if it can sell 100 million pieces, the cost per piece would only need to be 2 yuan. This significant reduction in cost would allow the company to make a profit and continue investing in research and development
Because the market size is large enough and there are enough buyers to help digest costs, companies can make profits and the industry can develop; this is the magic of the Chinese market.
Guosen Securities' research report states that in 2024, the electronic market will switch from &#34;cyclical recovery&#34; to &#34;growth and innovation,&#34; and the industry is expected to enter a year of valuation expansion in 2025. On the application side, AI is revolutionizing human-computer interaction, with AI edge applications centered around voice interaction reaching a critical point for large-scale commercialization, and innovation is frequently catalyzed.
Compared to AI cloud-side, domestic semiconductor companies will achieve higher market participation in AI edge-side innovation, while the self-sufficiency rate of domestic semiconductors remains relatively low. The resonance of both factors establishes the certainty and space for industry growth.
Author of this article: Jiu Lin, Source: Semiconductor Industry Review, Original Title: &#34;The Chip Revolution on the AI Edge&#34;
Risk Warning and Disclaimer
The market has risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this is at one's own risk

DPSK

00981

NVDA

AVGO

INTC

01347

The demand for small-sized edge AI chips has surged, and the Chinese cultural market is expected to perform strongly in 2025, with the AI market also thriving. The launch of DeepSeek has brought generative AI from the cloud to the edge, driving the exploration of AI hardware products. Although many products related to AI were showcased at CES, the performance was subpar due to the mismatch between local SoC computing power and large model requirements. Cloud AI inference faces issues such as cost, speed, and privacy, while edge inference demonstrates higher efficiency and safety in scenarios like autonomous driving

- In 2025, China's cultural market thrived with the film "Nezha 2" surpassing 10 billion yuan in box office, showcasing significant consumer power.  
- The AI market saw the emergence of DeepSeek, which offers lower computational requirements while achieving results comparable to ChatGPT, indicating a shift towards edge AI technology.  
- The demand for edge AI chips is rising as companies explore integrating AI into various hardware, with projections showing substantial growth in AI-enabled devices across multiple sectors.

The Chip Revolution at the AI Edge