On August 7th, the A-share market was reignited by the AI computing power industry chain. During the morning rebound, sub-sectors such as PCB and CPO (Co-packaged Optics) experienced a collective surge, with over a dozen constituent stocks hitting the upper limit of their daily limit up. The market leader, Guangxun Technology, which has a market capitalization exceeding 159.8 billion yuan, also hit the upper limit. The storage chip leader, GigaDevice, with its AH stock, also strengthened, achieving a cumulative gain of over 20% in the last four trading days. In the Hong Kong market, the two "powerhouses" of AI large models, MiniMax and Zhipu, rose by over 18% and 16%, respectively. From computing power hardware to model applications, a structural rally, driven by the dense launch of domestic large models and the capital expenditure of global giants, is unfolding across the board. What is the secret behind the surge in the computing power main line? Reviewing the day's market, the computing power industry chain showed a multi-point blooming pattern: Baoding Technology had 5 consecutive limit-up days in 6 days, Huazheng New Material had 3 consecutive limit-up days in 4 days, Jingwang Electronics had 2 consecutive limit-up days, and Yibo Technology, Norda Shares, and Ouke Technology hit the upper limit, while Shenghong Technology rose by over 11%. The CPO concept and PCB sectors rose together, reflecting the market's optimistic expectations for AI infrastructure demand, which is spreading from merely "chips" to more granular segments like "networking, connectivity, and packaging." Capital is no longer solely focused on the GPU narrative but is seeking flexible targets along the entire computing power construction chain. In fact, the start of this round of rallies can be traced back to the end of July. At that time, the market was skeptical about whether the massive capital expenditure of AI giants could translate into returns, and the sector experienced severe volatility. However, as domestic large models continuously released new versions and overseas tech giants raised their capital expenditure guidance, the expectation of a robust AI industry chain was rapidly restored, and capital flowed back. Judging from the fact that the Wind AI computing power index rose by 2.82% on August 5th, and the performance of Shengshi Technology and Industrial Fanex hitting the upper limit, the computing power main line has transitioned from a oversold rebound phase to a trend-based recovery phase. Domestic large models dominate global usage, changing the industry narrative. One of the core catalysts for this rally is the collective rise of domestic large models. Since August, domestic models such as DeepSeek, MiniMax, Kimi, and Qwen have been intensively iterated, continuously challenging top overseas models with high-level versions, while general-purpose versions are rapidly gaining traction based on the open-source ecosystem and cost-effectiveness. Data from the AI model aggregation platform OpenRouter shows that the top six global large language models in July—Xiaomi MiMo-V2.5, DeepSeek-V4-Flash, Tencent Hy3, MiniMax M3, Zhipu GLM-5.2, and DeepSeek-V4-Pro—are all independently developed open-source models from China. Domestic models have transformed from "followers" to "traffic drivers," directly strengthening the market's long-term confidence in the domestic computing power industry chain and laying the groundwork for the flourishing of AI applications. At the same time, OpenAI announced an 80% reduction in the API calling price for its flagship model GPT-5.6 Luna, which is seen by the industry as a landmark event marking the shift of the large model industry from an "intelligence ceiling arms race" to "computing power industrialization and cost reduction." The avalanche-like drop in inference costs has, on one hand, reduced the profit pain points for AI application companies, and on the other hand, it means that token consumption will increase exponentially—which is not a negative for the total computing power demand but rather a long-term benefit of "trading price for volume." As cited by Bank of China International, the decrease in token prices does not necessarily suppress the total computing power demand, and the higher consumption brought by multi-turn planning and tool calling by Agents is expected to boost inference computing power demand.

Global Giants Increase Capital Expenditure, Demand Spreads from Training to Inference

Overseas markets also provide significant support. Following this round of earnings season, four major tech giants—Microsoft, Alphabet, Meta, and Amazon—have all raised their capital expenditure guidance for 2026: Amazon increased from approximately $200 billion to approximately $220 billion, and Alphabet's full-year guidance increased from $180 billion to $190 billion to $195 billion to $205 billion. TrendForce predicts that the total capital expenditure of the top nine global cloud service providers in 2026 will grow by approximately 90%. The continuous expansion of capital expenditure provides the most direct macroeconomic backing for the prosperity of the AI computing power industry chain.

Guosheng Securities points out that capital expenditure by North American cloud vendors has not peaked. As multimodal large models, embodied AI, and AI Agents deeply penetrate industrial sectors, computing power demand is spreading from the "training end" to the "massive inference end," and the demand slope is increasing rather than decreasing. From Microsoft's earnings data, Azure and other cloud services revenue grew by 43% year-on-year, with annualized revenue exceeding $100 billion, and Microsoft 365 Copilot has over 30 million paying users. AI software company Palantir reported a 93% year-on-year revenue increase to $1.94 billion in its second fiscal quarter and significantly raised its full-year guidance—the performance of these giants is correcting the market's concerns that "AI only burns money and does not make money."

Large Orders Land in Clusters, Accelerating Performance Realization

Industrial deployment data is also impressive. On August 4th, Fengzhu Assistant announced that its wholly-owned subsidiary signed a computing power service contract worth 4.608 billion RMB with Company B, and simultaneously signed a computing power server procurement agreement worth 3.062 billion RMB with Company A. Gaole Co., Ltd. disclosed a computing power service contract worth 3.195 billion RMB, expected to increase revenue by approximately 200 million RMB in 2026. Xingyun Technology disclosed that it had 15.082 billion RMB in long-term in-hand computing power orders as of the end of July, all under contracts exceeding 5 years. The computing power orders worth billions of RMB from listed companies are turning the "AI infrastructure prosperity" from a slogan into real contract revenue, leading to greater market expectations for the mid-year and full-year performance of related companies.

Macro-level policy tailwinds are also continuing. The National Development and Reform Commission recently disclosed that the national intelligent computing power scale reached 2.8 times that of the same period last year by the end of June, the total global downloads of domestic large models exceeded 10 billion times, and over 120,000 high-quality datasets have been built, accelerating the formation of a virtuous cycle where computing power, models, and data mutually promote each other. Relevant institutions estimate that the construction of computing power networks in the "15th Five-Year Plan" period will add 400 billion RMB in direct investment, opening up broad medium-to-long-term space for the industry chain, given that computing power construction is mainly driven by corporate investment.

Institutions Assess the Future: Valuation Repair First, ROI Validation is Key

Regarding the future direction of the sector, institutional views are diverging but generally positive. Xing Ziqiang, Chief Economist at Morgan Stanley China, believes that AI investment has entered a "mid-season rest," and the logic is shifting from chasing upstream computing power chips to two new main tracks: cost reduction and realization in AI applications, and HALO resource support. China Galaxy Securities stated that the market valuation logic is transitioning from capital expenditure expansion to ROI return verification. As domestic large model capabilities approach the first tier globally, domestic AI applications and Agents will enter a period of flourishing, and the inflection point for commercial value realization is here, with performance in early deployment scenarios such as office and government services expected to turn around.

However, the investment banking strategy teams at China Merchants Securities also caution that the core contradiction in the commercialization of AI applications lies in the high enthusiasm for initial trials, while stable payment scenarios and enterprise ROI still require validation. CITIC Securities similarly points out that the narrative of "AI devouring software" is gradually being disproven, and whether the short-term sector can move from valuation repair to reversal ultimately depends on whether AI products can substantively drive an increase in corporate revenue growth. Overall, the driving force behind the current AI computing power trend has shifted from mere thematic speculation to a multi-resonance of "model capabilities + capital expenditure + order realization." For investors, even amidst high sector sentiment, it is necessary to distinguish targets that truly possess the ability to convert orders and exhibit earnings elasticity, rationally grasping structural opportunities rather than blindly chasing high prices.
Detail Page Ad