Since the dawn of this year, DeepSeek has consistently grabbed headlines, attracting substantial investment into sectors such as computing power, chips, and intelligent agents.

The domestic open-source AI large model DeepSeek is sweeping through the technology industry at breakneck speed, with its user base surpassing one billion. As an advanced inference model built on the Transformer architecture, DeepSeek boasts a massive parameter scale, placing stringent demands on hardware computing power, memory capacity, and bandwidth.

However, during a recent earnings call, Qualcomm CEO Cristiano Amon revealed that the newly popular DeepSeek R1 model benefits Qualcomm as its chips can efficiently run locally, eliminating the need for cloud reliance. DeepSeek R1 and similar models demonstrate that AI models are evolving to be faster, smaller, more powerful, and more efficient, now capable of direct on-device execution.

DeepSeek appears to be leveraging its robust local deployment capabilities to support complex AI tasks with low cost and power consumption, driving the full AI transformation of terminals like smartphones. This is anticipated to significantly boost market demand for AI mobile phone SoCs.

In recent years, AI has emerged as the hottest trend in the domestic mobile phone market. According to market research firm IDC, AI mobile phones are characterized by several key indicators: NPUs with computing power exceeding 30 TOPS, SoCs supporting generative AI models, and the ability to run various large models on-device. Over the past year alone, the domestic AI mobile phone market has witnessed rapid growth, with manufacturers like Huawei, Xiaomi, vivo, OPPO, and Honor swiftly integrating their respective cloud-based or on-device AI large models into their products.

Li Weimin, deputy analyst at the Strategic Development Research Institute of China Telecom Research Institute, highlighted that natural language processing capabilities are continually advancing, equating to voice assistants possessing "high intelligence" capable of accurately discerning user intent. Multimodal interaction integrates voice, gestures, facial expressions, etc., making the user interface more natural and enriched. At the software ecology level, intelligent cameras, intelligent translation, and other developer tools are also becoming smarter, enhancing both the efficiency and quality of development and testing.

At Apple's autumn product launch event in 2024, Tim Cook showcased the Apple Intelligence feature. The new iPhone 16 series became Apple's first true AI mobile phone. Shortly after, Huawei officially released HarmonyOS NEXT, deeply integrating AI with the OS, introducing a new Hongmeng native intelligence called Harmony Intelligence, and ushering in an OS experience tailored for the era of AI large models. XiaoYi, the intelligent agent equipped with the Pangu large model, is deeply integrated with the system AI navigation bar and resides on the screen, serving as a readily accessible system-level intelligent agent. At its new product launch event in October last year, Honor introduced smartphones equipped with a new AI system. To demonstrate the power and intelligence of AI functions, Honor CEO Zhao Ming used AI functions to order 2,000 cups of Luckin coffee for on-site guests. Additionally, Xiaomi launched "Super Xiao'ai," vivo released the "Lanxin Large Model," and OPPO unveiled "AndesGPT."

Thus, 2024 is also recognized as the "first year of AI mobile phones."

Apart from terminal manufacturers, MediaTek and Qualcomm's contributions in the AI field are equally noteworthy. Their new products, such as the Dimensity 9400 and Snapdragon 8 Elite, provide robust technical support for the rapid development of AI mobile phones.

The Snapdragon 8 Elite features a newly upgraded Hexagon NPU, offering innovative multimodal model processing capabilities that enable AI computing and computer vision to work in tandem, significantly improving the execution efficiency of AI tasks. For instance, through LMM (Multimodal Model), the NPU can simultaneously process voice commands and image content, achieving faster response speeds. This efficient processing capability allows the Snapdragon 8 Elite to excel in various intelligent application scenarios.

The Dimensity 9400 integrates MediaTek's Dimensity Agentic AI Engine and the eighth-generation AI processor NPU 890. This combination not only significantly enhances the performance of traditional AI applications but also elevates them to a new height of highly intelligent AI applications featuring autonomous perception, "brain" reasoning, and collaborative action. This means that the Dimensity 9400 can fully empower on-device AI, enabling rapid intelligent transformation of traditional applications across various fields like text processing, image processing, and music creation.

However, despite the active responses from mobile phone and chip manufacturers, AI mobile phones have yet to spark the anticipated wave of phone replacements among consumers. Data indicates that although generative AI is viewed as a significant technological breakthrough in the smartphone industry, most consumers currently do not intend to replace their phones due to AI features. This phenomenon persists in both iPhone and Android phones.

Large model-driven AI faces numerous challenges in edge and on-device inference applications, such as the difficulty of meeting the fine-tuning and inference requirements of large models with edge-side computing power and storage capacity. Model compression and lightweighting may lead to accuracy loss, affecting business outcomes.

With the continuous advancement of algorithm technology, including the development of model compression algorithms like model quantization, pruning, and distillation, as well as the emergence of software and hardware platforms specifically designed for edge deployment, the deployment of AI large models on edge devices has become more efficient and convenient. DeepSeek-R1 is a groundbreaking player in this field.

Over the past year, if mobile phone manufacturers have established basic on-device capabilities based on generative AI, the emergence of DeepSeek now shifts the competition among manufacturers from "technical parameters" to "who can embed terminal scenarios faster."

Huawei's system-level intelligent agent "XiaoYi" was the first to integrate the DeepSeek-R1 model on HarmonyOS NEXT (native Hongmeng). After upgrading to the latest version (11.2.10.310) of the XiaoYi APP, the DeepSeek R1 intelligent agent was launched. Users can directly invoke DeepSeek-R1 within XiaoYi for code inference, mathematical calculations, text generation, and even complex logical reasoning tasks.

Honor also announced the launch of the HONORALPHA PLAN, further deploying around the AI ecological niche. Relevant strategies and technologies will be unveiled at the Mobile World Congress 2025. Prior to this, Honor stated that the DeepSeek-R1 online version has officially been launched, with the first batch of supported models including the Honor Magic7 series, foldable V2 series, etc. The "HONORALPHA PLAN" is regarded internally at Honor as a critical turning point in AI strategic layout.

vivo officially announced that its Lanxin Large Model will be integrated with DeepSeek to further enhance the AI experience of its intelligent assistant "Lanxin XiaoV." From the images officially released by its OriginOS, after the integration is completed, vivo's OriginOS pre-installed application Lanxin XiaoV will support deep thinking (R1) capabilities, providing functions such as image generation, AI text creation, and AI question answering.

OPPO announced that the upcoming globally thinnest folding flagship, the OPPO Find N5, will officially integrate DeepSeek-R1. Users can directly activate it through voice commands via XiaoBu Assistant without downloading or multiple steps.

Nubia has fully embedded the 671 billion-parameter DeepSeek into its system, enabling interconnection with devices such as multimodal AI headphones. Currently, the Z70 Ultra is undergoing internal testing. By fully embedding DeepSeek-R1, the Nubia Z70 Ultra can directly invoke DeepSeek-R1 through the Nebula intelligent conversation interface, avoiding the cumbersome operations brought by multiple entry points.

Furthermore, Meizu's new version of the voice assistant integrated with DeepSeek-R1 has been fully rolled out to users of the Meizu 20 series, 21 series, and Lucky 08. Users of the above models with the Flyme 11 system version who receive the 11.3.19 version update of the Aicy voice assistant app can smoothly use DeepSeek-R1.

DeepSeek's technological breakthroughs in algorithm optimization and other aspects enable large-scale model training at a lower cost and enhance inference capabilities. These breakthroughs are pivotal for reducing the cost of AI model training and inference. Following the wave of DeepSeek adaptation in the AI chip and cloud computing fields, smartphones are also comprehensively embracing the DeepSeek integration trend. DeepSeek is igniting a new battle in mobile phone AI.

The first stage of AI mobile phone development focused on improving or enhancing existing mobile phone functions. The second stage will introduce new functions presented in the form of AI intelligent agents, where mobile phones truly differentiate themselves from users downloading third-party AI apps through built-in localized on-device models.

Industry insiders stated that based on the open-source nature of DeepSeek, mobile phone manufacturers' AI capabilities will be significantly enhanced in the second half of the year. "First, everyone's AI integration costs have decreased, especially for small and medium-sized mobile phone manufacturers. Without the need to independently develop large models, they can quickly make up for shortcomings in AI capabilities, potentially accelerating the popularization of AI functions in low-end models and driving the penetration of the AI mobile phone market."

The trend of mobile phone "AIization" is a double-edged sword. On one hand, AI mobile phones offer the possibility of an almost "fully automated life": a single command can order coffee, automatically navigate routes, and send AI WeChat red envelopes in groups. On the other hand, while enjoying the benefits of technology, privacy concerns are also on the rise. The challenges faced by AI mobile phones stem not only from technological limitations but also from security and regulatory issues.

Although most Chinese mobile phone manufacturers have already integrated DeepSeek, primarily through cloud deployment, this method involves close interaction between users and model data, often touching upon personal privacy. Some netizens have expressed concerns that "cloud-side AI functions are powerful but pose high privacy risks. Frequent data uploads to the cloud are like leaving a backdoor for hackers. I choose on-device AI because, although its performance is slightly weaker, I feel more at ease knowing that my data stays on my phone."

Theoretically, on-device AI can minimize data transmission and avoid uploading sensitive user information to the cloud. When managed properly, it presents a "more private" solution. Due to privacy concerns and other factors, as the cost of computing power further decreases in the future, on-device AI may be deployed locally.

Among these, the AI performance of mobile phone SoCs is crucial. In the realm of AI chip design, NVIDIA stands as the undisputed king.

Recent news indicates that NVIDIA is deepening its collaboration with MediaTek, planning to launch AI PC chips in the second half of 2025. Additionally, it is developing an AI mobile phone chip to expand its presence in the mobile market. The collaboration between NVIDIA and MediaTek is expected to bring innovation to the smartphone market. Leveraging their respective strengths, they have the potential to occupy a significant position in the custom chip market. Although details about the mobile chip are still scarce, this collaboration undoubtedly deserves close attention.

In fact, NVIDIA ventured into the mobile chip market many years ago but ultimately failed. This AI revolution in smartphones will also present NVIDIA with new opportunities in the mobile chip field.

Beyond mobile phone SoCs, HBM technology may also benefit from the AI wave. Benefiting from the surging demand for AI chips, SK Hynix became the second-largest market capitalization company in South Korea last December, second only to Samsung Electronics. However, due to high costs and technical difficulties, HBM technology was previously mainly used in data centers, with NVIDIA being one of its largest customers.

As NVIDIA's primary supplier of High-Bandwidth Memory (HBM), SK Hynix's share price soared by over 50% over the past year, reaching a market capitalization of 97.27 trillion Korean won (approximately $73.84 billion).

Currently, Samsung Electronics is developing the next generation of DRAM to maximize the computing performance of AI smartphones and PCs. To ensure the performance and stability of mobile HBM, Samsung Electronics previously announced that it will use copper pillars to connect stacked DRAM.

It is reported that Samsung Electronics will launch its next-generation mobile memory in 2028, a new low-power wide I/O (LPW) DRAM optimized for AI on devices. The new LPW DRAM, also known as low-latency wide I/O (LLW), is referred to as "mobile HBM" memory optimized for high performance and low power consumption. Samsung aims to become a leader in the mobile memory market with its next-generation LPW DRAM specifically designed for on-device AI.