Soochow Securities: The computing power rally is not over yet, actively position in AI+ and innovative pharmaceuticals on the left side

Dongwu Securities released a research report stating that from June to August, the market's main theme revolved around artificial intelligence. Currently, the rally is mainly concentrated in upstream infrastructure hardware, with the highest prosperity certainty seen in overseas optical modules and PCBs, which took the lead in June. In mid-August, domestic computing power, with Cambricon at its core, officially began to catch up. With ample market liquidity and no obvious flaws in the industrial logic, we do not believe the computing power rally is about to end. However, for investors holding cash on the sidelines, the strong profit-making effect of upstream hardware is causing anxiety, and objectively, the increasing volume of profit-taking will bring some pressure to cash out. Subjectively, the continuous accumulation of gains inevitably makes risk-averse funds anxious.

The core reason for the lag in downstream applications during this round of AI rally is the lack of short-term certainty—there have been no breakout blockbuster products or smooth commercial models. At the listed company level, this is reflected in insufficient earnings visibility, so these companies have not become the first choice for capital in this round. However, from the perspective of the evolution of the technology wave, the ultimate realization of AI empowering everything must be achieved through the application end. In other words, the explosion of applications has medium-term certainty and a broader space than upstream hardware, as already verified in the "Internet+" wave and corresponding stock market rally a decade ago. This also means that the rally in AI applications is only a matter of time.

In the current market environment with ample liquidity and a focus on technology centered around AI, if there is a loosening of chips in upstream hardware (for example, a core stock adjusts by about 20%), then only a few marginal events that attract market attention (such as H20 security issues and DS model's FP8 making the domestic computing power rally go from dark to light) are needed for low-level branches within AI to show strong elasticity. However, joining in at that point would mean losing some odds, facing the same dilemma as now—whether to chase the rally in computing power. Therefore, it is recommended to treat downstream application directions such as AI+ innovative drugs, AI+ military industry, AIGC, edge AI, humanoid robots, and intelligent driving as "call options" based on medium-term industrial logic certainty, and actively deploy on the left side.

Dongwu Securities' main viewpoints are as follows:

The ultimate outcome of the technology wave is to empower everything, which was already confirmed in the "Internet+" era. From the perspective of industrial trend evolution, the rise of AI applications is inevitable, and the latter half of the AI rally will revolve around the application end.

Taking the "Internet+" industrial wave a decade ago as an example, with the technological progress and penetration rate improvement of upstream infrastructure and edge hardware, the explosion of downstream applications is an inevitable trend from the endgame perspective. However, when and in what form it will appear requires post-hoc observation. In addition, from the two dimensions of time and space, the sustainability of the downstream application rally is stronger than that of upstream hardware.

From a time perspective, the upstream hardware rally was the consumer electronics rally around the Apple supply chain from 2009/06 to 2010/12, lasting about 1.5 years. The subsequent server direction followed the applications from 2013 to 2015, but mainly as sporadic rallies. The downstream application rally started in January 2013 and did not peak until June 2015, lasting two and a half years. From a spatial perspective, during the mobile Internet rally from 2009/1/5 to 2015/6/12, the computer and media sectors in TMT had greater upward elasticity, with cumulative gains of 1039% and 710% respectively, higher than the hardware sectors of electronics and communications at 697% and 300%.

Specifically:

From 2009 to 2012, the mobile Internet began to take shape, 3G penetration continued to rise, and the release of the iPhone 4 drove a surge in global smartphone shipments. The leading direction was mainly hardware represented by the electronics sector, with the trading logic being prosperity and growth. In addition, the application end also saw blockbuster products represented by Sina Weibo, but the commercialization model did not work out, and Internet applications were still in the thematic investment stage.

Since 2013, with the accelerated reduction of 4G costs and the high penetration rate of smartphones, the mobile Internet rally extended downstream. In 2013, WeChat launched its payment function, and the "free game + in-app purchase" model pioneered by "I'm MT" broke through the monetization bottleneck of mobile games. The number of users of blockbuster applications such as mobile payments and mobile games surged during this stage, driving the rally from upstream hardware to downstream applications represented by media and computers. In March 2015, the government work report first proposed the "Internet+" concept, and in July of the same year, the State Council issued the "Guiding Opinions on Actively Promoting the 'Internet+' Action," with top-level design driving the continuous expansion of downstream application scenarios. The "Internet+" concept continued to penetrate the service industry and midstream manufacturing, fully expanding the breadth of the rally. Coupled with the liquidity bull market in 2015, "Internet+" became the absolute main theme of the market at that time. In addition, on the application side, "shovel sellers" (server manufacturers) gradually outperformed smartphones to become the leading hardware direction in this stage.

Looking at software applications alone, the rally can be roughly divided into two stages: the "broad-based rally under the Internet+ narrative" from 2013 to 2015 and the "winner-takes-all leader rally" from 2016 to 2017. From 2013 to 2015, with the full-industry penetration of "Internet+," emerging business models and applications emerged one after another, all with great imagination. At this time, vertical markets were still in the land-grabbing stage, and the issue of market structure had not yet emerged. Coupled with systematic valuation improvement in a liquidity-rich environment, software applications showed a "rise as long as it's related" trend.

From 2016 to 2017, the rally was clearly structural. On one hand, the initial dividend of Internet traffic began to show signs of fading. According to the China Internet Network Information Center, the Internet penetration rates in China in 2016 and 2017 were 53.2% and 55.8%, respectively, with the growth slope slowing compared to 2010-2015. Mobile Internet gradually shifted to a stock market, with leading companies expanding their market share through advantages in capital, technology, and user resources, while the survival space for small and medium-sized companies was significantly compressed. On the other hand, after the "water buffalo" rally peaked in June 2015, there was a significant valuation correction, market risk appetite declined, and the requirement for companies to deliver on performance increased. In the Internet sector, where the Matthew effect is extremely pronounced, leading companies had higher earnings visibility. In 2016-2017, Tencent and Alibaba achieved high growth in revenue and profits, with their stock prices rising accordingly. From 2016/1/1 to 2017/12/31, Tencent in Hong Kong and Alibaba in the US rose by 167% and 112%, respectively.

In comparison, what we are likely to see next is the "AI+ rally" narrative's broad-based stage. On one hand, the growth slope of the AI application sector has not yet slowed, with token usage and AI software user numbers still accelerating. According to the National Data Bureau's press conference on August 14, as of June 2025, China's daily token consumption exceeded 30 trillion, a more than 300-fold increase since the beginning of 2024. QuestMobile data also shows that as of March 2025, the number of active users of AI-native apps reached 270 million, a year-on-year increase of 536.8%. On the other hand, there have not yet been blockbuster AI applications or clear and efficient commercial realization models in China, and the market's understanding of AI applications is not yet sufficient. Before the "winner emerges," the application side is expected to first see a "hundred flowers bloom" rally.

China's AI application end has multiple factors such as policy and resource endowment supporting its qualitative transformation and implementation.

The latest "Artificial Intelligence+ Action Opinions" issued by the central government at the policy level can be compared to the "Internet+ Action Opinions" of 2015. The top-level design clarifies the development goals of "Artificial Intelligence+," and subsequent specific implementation measures by various functional departments are expected to be introduced. The development and implementation of AI downstream applications now have a clear "deadline."

In August 2025, the State Council issued the "Opinions on Deepening the Implementation of the 'Artificial Intelligence+' Action," clarifying that by 2027 and 2030, China aims to achieve penetration rates of 70% and 90% for new-generation intelligent terminals and intelligent agents, respectively. By 2035, China will fully enter a new stage of intelligent economy and intelligent society development, emphasizing the deep integration of artificial intelligence with science and technology, industrial development, consumption upgrading, people's well-being, governance capacity, and global cooperation in six key areas. Previously, "Artificial Intelligence+" was included in the government work report in both 2024 and 2025, but this is the first time that specific development goals and implementation schedules have been set. More local AI support measures are expected to follow, accelerating the development of downstream applications.

China has a natural foundation for the vigorous development of AI applications, with an engineer dividend and a huge user base as the basis for the quantitative to qualitative change of downstream applications. From the perspective of hardware applications, the technology-intensive nature of AI terminal hardware will be further strengthened compared to the previous two technology cycles. China's unique engineer dividend provides stronger high-end manufacturing capabilities than other newly industrialized countries and lower cost advantages than North America. Taking humanoid robots, the largest-scale AI terminal application scenario, as an example, Tesla's Optimus mass production will still favor Chinese suppliers.

From the software side, China's user scale advantage is evident. According to QuestMobile's "2025 AI Application Market Semi-Annual Report," as of June 30, 2025, the number of monthly active users of mobile AI applications reached 680 million. China's huge user base not only creates enormous potential market demand and more diverse application scenarios, but also provides more diverse and massive data to assist model training and improve vertical application capabilities, empowering the development of the AIGC industry chain.

The lesson from this round of domestic chip rally: as long as the industrial logic is certain, the rally is only a matter of time.

Fundamental visibility affects the priority of sectors in investors' minds. Branches with slightly weaker current prosperity may lag in performance order, but as long as market liquidity is ample and the AI rally continues, the rally in directions with industrial trend logic is only a matter of time. The current domestic computing power rally is a typical example.

The reason why overseas computing power hardware became the leader in this round is mainly due to the most solid current performance, high-visibility medium-term prosperity, and hard-to-disprove future growth. In contrast, domestic computing power, edge, and software applications are "a bit lacking" in these dimensions. Looking back, this round of overseas computing power chain rally started at the end of May and lasted for a long time. The core was that the current and medium-term prosperity of PCB, optical modules, and other computing power communications was continuously verified and revised upward by the performance of domestic manufacturers such as New Yisheng and the capital expenditure guidance of overseas CSP giants. However, during the two months from June to early August, the performance of the domestic computing power sector remained "lukewarm."

But investors familiar with the technology sector know that regardless of whether NV computing cards are allowed to be exported to China, achieving chip self-sufficiency is a matter of survival in the AI era. The establishment of the third phase of the National Integrated Circuit Industry Investment Fund and the innovation layer of the STAR Market all clearly indicate top-level guidance for capital and resource support for "technological self-reliance and self-improvement." In other words, the medium-term logic of domestic chips and upstream manufacturing (foundry, equipment, materials, etc.) is very certain, but there has not yet been a landmark event to attract more capital attention, lacking a catalyst for the rally to start.

On August 8, SMIC's Q3 performance guidance was slightly below the market's optimistic expectations, resulting in significant selling pressure. This shows that just before the official start of the domestic computing power and semiconductor rally, the sector was still in a state lacking profit-making effect. However, based on strong medium-term certainty and ample market liquidity, the sector's bottom center was slowly rising. Still taking SMIC as an example, after being oversold on August 8, funds actively bottom-fished and the rally quickly recovered. Until August 10, when CCTV's Yuyuantantan published an article "How the US Installs 'Backdoors' in Chips" revealing that H20 is "neither environmentally friendly, nor advanced, nor safe," followed by Cambricon's 20% limit-up on August 12, the domestic computing power line officially "came to light."

On August 21, DeepSeek V3.1 used the FP8 parameter architecture to enhance the adaptability of domestic chips. On August 27, the Financial Times reported that Chinese chip manufacturers plan to triple AI processor production next year. The bullish narrative continues to accumulate, strengthening sector confidence. The STAR 50 index showed great upward elasticity: from August 1 to 27, STAR 50 rose by more than 20%, significantly outperforming the CSI 300 and ChiNext indices, with Cambricon, as the leading domestic chip company, rising more than 90% during this period.

The transformation of domestic chips and semiconductors from "ignored" to the market's main theme is essentially a process of building market consensus and accumulating long momentum. However, if you chase the rally after "two bullish candles," you will inevitably lose some odds. If you enter even later and lack a profit cushion, your holding mentality will also be affected. When the medium-term industrial trend is certain, it is impossible to accurately predict when an important catalyst will occur to reverse capital attitudes, but when market liquidity is ample, the downside space for the sector is also very limited. At this time, the better strategy is to use odds thinking and treat relatively low-level sectors with certain industrial logic as "call options" for left-side deployment.

Currently, in the AI rally, the downstream application end is significantly lagging behind upstream hardware, making it a potential allocation direction with odds advantage. The rally is only a matter of time. Since the "golden pit" formed by the market after the 4.7 tariff incident, if we take the STAR 50 as a benchmark, only the upstream hardware end has outperformed, followed by consumer electronics and robotics, while the software application sector has lagged the most. From the performance from April 7 to August 26, the upstream hardware sectors of optical modules/PCB/high-speed copper connections/servers outperformed the benchmark by 67.8/17.3/14.6/6.3 percentage points, while only the game sector in the mid- and downstream software and application end outperformed the benchmark by 14.3 percentage points. Other sectors such as AIGC/intelligent agents/cloud computing/humanoid robots/AI wearable devices/intelligent driving/e-commerce underperformed the benchmark by 13.1/12.5/16.5/13.3/15.1/19.4/14.7 percentage points, respectively.

Looking at the domestic software application direction, insufficient model capability leads to limited vertical application capability, and the market has not yet seen the possibility of earnings realization or the imagination space brought by blockbuster applications in the software application sector. From the "overseas mapping" perspective, currently, US-listed AI software applications are also only showing single-point performance in companies such as Palantir and AppLovin, without a large-scale rally. The core is that the "singularity" of industrial development still needs to be awaited, which is also one of the concerns for many investors, so the AI software application sector is the most obviously lagging.

However, according to the previous analysis, the implementation of applications is the inevitable result of the technology wave. Before leading companies emerge, there will first be a comprehensive rally based on industrial logic narrative. The trigger may be a leap in the capability of a domestic foundational model and a surge in token calls, a steep rise in monthly active users/rankings of a certain application, or a specific R&D or implementation subsidy given by an "AI+" policy.

In fact, in a healthy "slow bull" pattern, it is difficult for excess returns within a sector to widen indefinitely. Some investors holding cash tend to "set another table" in low-level branches due to fear of heights, while some holders' profit-taking mentality is amplified as profits accumulate. If there is a loosening of chips in high-level upstream hardware (volatility/adjustment) in the future, the spillover of liquidity will also help enhance the upward elasticity of low-level branches. For funds that missed out on upstream hardware, there is motivation to deploy in the low-level downstream application direction with low short-term earnings visibility but certain medium- and long-term endgame and cost-effectiveness at the current position. Based on odds thinking, it is recommended to actively deploy investment opportunities in downstream application directions such as AI+ innovative drugs, AI+ military industry, AIGC, media games, AI edge, humanoid robots, and intelligent driving sectors.

Based on this, the following AI application end directions (including hardware and software applications) are strongly recommended:

AI+ innovative drugs: The application of AI in the pharmaceutical field is expected to significantly reduce the cost and time cycle of drug discovery, accelerate the target development and validation process, and reduce the risk of initial trial failure through simulated clinical trials.

AI+ military industry: Artificial intelligence empowers the informatization of the military industry by effectively and in real time integrating multi-source intelligence data such as satellites, radar, and drones to build a comprehensive and accurate battlefield situation map and revolutionize command systems. Unmanned equipment and autonomous combat systems such as robot wolves and robot dogs are another key area of AI+ military industry.

AIGC: The endgame narrative is complete, but in the short term, the launch of blockbuster applications is still pending, and earnings visibility is low. Pay attention to catalysts such as upgrades in the capabilities of domestic large models and progress in the AI Agent industry.

Humanoid robots: The largest-scale terminal application scenario for AI, with domestic robot manufacturers gradually entering the order verification stage. Focus on updates to Tesla Optimus V3's new blueprints.

Consumer electronics: New products will be launched intensively after September. Focus on the Apple consumer electronics new product launch event on September 10 and feedback on Meta AI glasses products.

Intelligent driving, vehicle-road-cloud: The VLA technology paradigm is reshaping the landscape of car companies, and cloud-vehicle collaboration competition is entering a white-hot stage. This is also an important branch of edge AI, but short-term elasticity is relatively insufficient due to the competitive landscape among car companies.

AI+ others: AI+ finance, AI+ agriculture, AI+ logistics, AI+ law, AI+ government affairs, AI+ e-commerce, AI+ programming, etc.

Risk Warning

Domestic economic recovery is slower than expected; Federal Reserve rate cuts are less than expected; macro policy strength is less than expected; technological innovation is less than expected; geopolitical risks.