The artificial intelligence and computing power sectors have been receiving a steady stream of positive catalysts recently. The seventh edition of the AI Computing Conference (AICC) is scheduled to take place on September 21, 2026, at the Zhongguancun International Innovation Center. Since the start of this year, intelligent agents have been accelerating toward explosive growth both globally and within China, rapidly entering production, scientific research, and social services, with the transformative effect of AI converting into real-world productivity becoming increasingly evident. At the same time, the computing infrastructure that underpins AI operations is being elevated to the status of a national strategic asset. The conference will focus on the computing industry transformation driven by intelligent agents, bringing together diverse technology pathways and industrial forces to explore a development path suited to China's AI computing industry, with the goal of pushing AI from a technical capability held by a few institutions to a fundamental utility that is universally accessible.
The 2026 Apsara Conference is set to be held in Hangzhou from September 22 to 24. Under the theme "Intelligence in Action," the event will center on Agentic AI, featuring three main forums, over 120 sub-forums, and more than 1,000 exhibitors. It will showcase Alibaba Cloud's full-stack AI technology across chips, cloud, models, and agents, while also unveiling Alibaba Cloud's core technologies and annual product releases. In addition, on September 23, the Fifth GMIF Innovation Summit, co-hosted by the Shenzhen Memory Industry Association and Peking University's School of Integrated Circuits, will officially open at the Renaissance Shenzhen Bay Hotel. With the theme "AI Storage, Driving the Future," the summit takes a holistic view of the storage industry chain, using the Token economy as a key lens, and brings together leading enterprises, top technical experts, and industry leaders from the global storage and semiconductor supply chain to discuss new trends, new demands, and new opportunities for collaboration in the AI-era storage industry.
Entering 2026, agents have become a red-hot track, with office and daily-life agent products emerging rapidly both domestically and internationally. As agents become increasingly common in daily life, users experience the convenience of "delegating complex tasks to an agent with a single sentence." The International Data Corporation (IDC) recently released its DAA Research Report, which tracks Daily Active Agents. The report shows that the number of active agent units globally stood at 28.6 million in 2025, is projected to reach 79.4 million in 2026, and is expected to surge to 2.216 billion by 2030. According to a March 2026 Gartner report, an agent workflow consumes between five and 30 times more Tokens per task than a standard chatbot, because a single user request typically triggers 10 to 20 sequential model calls. Goldman Sachs forecasts that global Token consumption will grow 24-fold by 2030 compared to 2026, reaching approximately 120,000 trillion Tokens per month, with always-on agents potentially consuming more than 100,000 Tokens daily.
In this round of AI infrastructure investment, the importance of memory chips is rising rapidly. Recently, South Korean brokerage KB Securities analyzed that as AI infrastructure investment expands at an unprecedented pace, the memory chip market is facing a severe supply shortage, with Samsung Electronics and SK Hynix holding less than 10 days of inventory, and actual sellable supply next year may clearly fall short. KB Securities projects that global top-tier cloud providers' AI-related investment will reach $1.3 trillion next year, up 60% year-over-year. The report also notes that memory chips' share of AI infrastructure investment is expected to rise to 40% this year from 14% last year, and further to 57% next year. The analysis suggests that structural changes on the supply side will amplify the supply-demand imbalance. Memory chip makers are shifting significant capacity toward producing the next-generation high-bandwidth memory chip HBM4, which is specifically designed for AI large-model training. Producing HBM4 consumes roughly three times the wafer capacity of traditional DRAM for the same memory capacity. DRAM mainly serves as the operating memory for phones and computers, serving as the core chip for consumer-end general-purpose memory. With limited total wafer capacity, the expansion of high-end HBM4 production will crowd out capacity for standard memory chips, further intensifying supply pressure in the coming period.
China Securities stated that since 2026, the investment logic of the AI industry has been shifting from a race of model capabilities and capital expenditure toward commercial validation based on orders, revenue, and profitability. Overseas agent products are the first to generate incremental revenue, with cloud providers' high-intensity computing investment still supported by cloud revenue and order backlogs. Domestically, the capability gap of models in Coding and agent tasks is narrowing quickly, with inference efficiency, Token call volume, and product revenue all improving simultaneously. Looking ahead over the next one to two quarters, domestic model iteration, agent product upgrades, and the repricing of flagship models are expected to advance together, and the firm recommends continued attention to computing services, domestic chips and supernodes, as well as B-end AI application vendors with proprietary scenarios, data, and enterprise delivery capabilities. Dongguan Securities noted that on the demand side, AI training, inference, and agent applications are expanding steadily, with overseas cloud providers' CAPEX still being revised upward. This drives sustained high prosperity in computing infrastructure construction and continuously pulls demand across memory, CPU, power semiconductors, and semiconductor materials, with some sub-categories already seeing tight supply-demand conditions and price increases. On the supply side, key segments such as semiconductor equipment, high-end GPUs, memory chips, and advanced packaging remain in an accelerated phase of domestic substitution, with the push for supply chain self-reliance and control continuing to strengthen.
Among the related concept stocks, Z.AI (02513): In early September, CMB International released a research report stating that the company's annualized recurring revenue growth trajectory is steady with strong prospects, and the disclosure of new user scale and engagement further reinforces its positive view. Driven by stronger annualized recurring revenue growth, the firm raised its 2025-2028 revenue compound annual growth rate forecast from 199% to 275%, and lifted the target price from HK$1,503.9 to HK$1,985, maintaining a "Buy" rating. The firm believes the company has struck an effective balance between model iteration pace, intelligence level, and per-task cost, maintaining its position at the Pareto frontier of the large language model industry, providing ample room for long-term monetization. The report noted that Z.AI completed six GLM iterations in 11 months, with its intelligence index rising from 32 to 60; MaaS platform registered users grew 144% to 7.4 million, paying daily active users grew 603%, Token call volume has increased more than 40-fold since the start of the year, while the average API selling price rose approximately 101%. As of August 2026, the annualized recurring revenue from its open platform and API reached $1.6 billion, up from $1 billion in early July, with management guidance of reaching $2.4 billion by year-end. The deployment of a domestic 100,000-card chip cluster and the rebuilding of the inference stack have reduced per-Token costs by approximately 80% since the start of the year, with the compute multiplier up 14-fold year-over-year, and the open platform gross margin reaching 24.6% in the first half of 2026. CMB International raised its 2025-2028 revenue CAGR forecast from 199% to 275%, and expects total revenue of RMB 38.1 billion in 2028.
ILUVATAR COREX (09903): The interim results show that thanks to the concentrated explosion in cloud-side AI inference computing demand, the Zhikai series achieved breakthrough growth, with revenue of approximately RMB 654 million in the first half of this year, a sharp increase of 651.8% year-over-year; gross profit was approximately RMB 233 million, with a gross margin of 35.6%, becoming the core driver of revenue growth. The Tiange series, positioned for high-end training scenarios, maintained steady growth, with first-half revenue of approximately RMB 262 million, up 38% year-over-year.
KINGSOFT CLOUD (03896): The company has been continuously strengthening its AI cloud service capabilities in recent years, driving the upgrade of its traditional cloud computing business toward AI infrastructure services. As one of the earliest independent cloud service providers in China, KINGSOFT CLOUD has fully embraced AI over the past two years. Leveraging the natural scenarios of the Xiaomi and Kingsoft ecosystems, AI-related revenue continues to scale, with intelligent computing cloud services becoming the primary growth engine. At the same time, it continues to increase AI infrastructure investment, strengthening its full-stack capabilities from IaaS to MaaS. As enterprise demand for AI application continues to rise, companies with cloud platform capabilities, customer resources, and ecosystem advantages are expected to gain new growth space in the AI industrialization process.
粤港湾智算 (01396): The company released interim results for the six months ended June 30, 2026, showing tenfold revenue growth: operating revenue of approximately RMB 2.56 billion, a surge of approximately 1,020.9% year-over-year, reaching a new revenue milestone with robust growth momentum. Profit achieved a leapfrog rise: period profit of approximately RMB 281 million, reaching five times last year's full-year net profit (approximately RMB 56 million), with profitability achieving a historic breakthrough. Its computing scale remains at the industry's forefront: FP16 dense computing power exceeds 50,000 PFLOPS, solidly positioned at the industry's top tier, building the underlying strength for a 100,000-card ultra-large computing cluster.
商汤-W (00020): In late August, CMB International released a research report maintaining a "Buy" rating on the company and raising its total revenue forecast for 2026-2028 by up to 7% to reflect the strong performance of its generative AI business. First-half revenue increased 23% year-over-year to RMB 2.91 billion, broadly in line with market expectations. CMB International noted that the company disclosed for the first time that recurring revenue (RR) for the first half reached RMB 1.14 billion, up 124% year-over-year, accounting for 39% of total revenue, up from 22% in the same period last year, reflecting a successful transition from one-off project deliveries to a recurring revenue model. During the period, generative AI revenue increased 28% year-over-year to RMB 2.33 billion, accounting for 80% of total revenue, mainly driven by strong demand for Token Factory and intelligent agent application services; visual AI revenue increased 14% year-over-year to RMB 497 million, accelerating from the 3% growth rate in the previous full year.
Comments