On September 3, US tech stocks mounted a robust rebound, with the AI computing chain rallying broadly. Dell Technologies jumped 15.81%, Nvidia rose 3.21%, and Micron gained 2.43%, while the Nasdaq advanced 1.40%. Notably, pure chip stocks were not the leaders in this rally—server vendors showed greater elasticity. As one of the world's core AI server suppliers, Dell's sharp share price increase carries strong industrial signaling, once again confirming that AI capital expenditure is shifting from the chip segment toward a broader infrastructure landscape. Moving forward, the AI industry will need not only more powerful computing chips but also the computing infrastructure capable of supporting large-scale application deployment.

Over the past two years, market attention has centered on large model capabilities and GPU supply, with chipmakers like Nvidia being the primary beneficiaries of the AI wave. Now, the focus is extending to the backend: high-speed interconnect, AI networking, and data center infrastructure are gaining rising importance, with companies like Broadcom continuing to benefit from AI data center construction and surging demand for high-speed connectivity. From GPUs to accelerator chips, and from servers and high-performance storage to network equipment and data centers, the AI infrastructure supply chain is becoming increasingly complete. The underlying driver of this diffusion is a shift in demand structure. Whereas the core question was once "can large models be trained," the current challenge has become "how can large models be widely deployed." Training requires concentrated, high-performance computing power, while inference demands sustained, stable, and efficient compute supply. As AI applications penetrate manufacturing, finance, healthcare, energy, and other physical industries, computing demand is transitioning from phased investment to long-term infrastructure needs.

Overseas capital expenditure changes are first being reflected in China's chip and server segments. Companies such as Inspur Electronic Information (000977.SZ), Foxconn Industrial Internet (601138.SH), and Dawning Information Industry (603019.SH) continue to expand their AI server and high-performance computing offerings, providing foundational support for large model training and industry applications. Cambricon Technologies (688256.SH) stands out as a domestic chip exemplar: amid restricted access to high-end GPUs, its cloud training and inference chips have undergone continuous iteration, with customers covering leading internet firms and smart computing center operators. Revenue is climbing rapidly alongside domestic computing procurement—"domestic substitution" has evolved from a theme into tangible orders.

The acceleration of computing infrastructure construction is now showing up in domestic companies' interim results. Renze Smart Technology (300442.SZ) focuses on AIDC operations: in the first half of 2026, it reported revenue of 3.746 billion yuan, with AIDC revenue reaching 1.995 billion yuan—up 126% year-over-year—accounting for over half of total revenue for the first time, at a gross margin of 47.3%. Net profit attributable to shareholders stood at 1.203 billion yuan, with operating cash flow equivalent to 79% of revenue. The company operates approximately 750MW of data center capacity, has delivered around 100,000 high-density liquid-cooled racks with PUE as low as 1.08, and counts clients including ByteDance, Alibaba Cloud, and Tencent Cloud. It also utilizes REITs to achieve cyclical exit from heavy assets. Xiechuang Data (300857.SZ) is pivoting from smart terminal manufacturing and trading into computing services, with main businesses encompassing equipment sales and computing technology services: first-half revenue hit 12.523 billion yuan, up 153.3% year-over-year, with net profit attributable to shareholders at 1.838 billion yuan and ROE of 34.3%. The company plans a private placement of 8 billion yuan for smart computing centers, targeting 50,000P (FP16 dense) of computing capacity by end-2026, while extending toward MaaS models via platforms like FCloud and TokenShare. Litong Electronics (603629.SH) is built around server resource trading and computing services: first-half net profit attributable to shareholders reached 702 million yuan, up 1,275% year-over-year, with a comprehensive gross margin of 46.6%—the highest in the sector. Computing-related service revenue totaled 1.274 billion yuan, representing 61% of total revenue. The company holds Nvidia Preferred-level certification, operates 38,000P of high-end computing capacity with consistently full data center occupancy, and has signed a three-year, 5 billion yuan long-term agreement with Tencent.

As the AI industry enters the phase of scaled application, Hong Kong-listed companies are also positioning around cloud services, computing operations, and AI infrastructure services. Kingsoft Cloud (03896) has been intensifying its AI cloud service capabilities, upgrading its traditional cloud computing business toward AI infrastructure services. As one of the earliest independent cloud providers in China, Kingsoft Cloud has fully embraced AI over the past two years: leveraging natural scenarios within the Xiaomi and Kingsoft ecosystem, AI-related revenue continues to scale, with smart computing cloud services becoming the primary growth engine. The company is also increasing AI infrastructure investment, strengthening full-stack capabilities from IaaS to MaaS. With rising enterprise demand for AI applications, companies possessing cloud platform capabilities, customer resources, and ecosystem advantages are poised to capture new growth space during AI industrialization. In the computing services arena, GBA AI COMP (01396) is progressively upgrading from a large-scale computing technology service provider into a "super TOKEN factory." The company structures its operations around AI computing services and operations, driving the transformation of computing resources from basic infrastructure into industrial service capabilities through resource integration, computing cloud platform development, and application ecosystem expansion. First-half revenue reached 2.56 billion yuan—ten times the year-ago period—with net profit of 280 million yuan, five times the full-year figure for last year. Currently, its in-operation computing scale exceeds 50,000P (FP16 dense), accounting for over 2% of national total computing capacity, with plans to expand to 80,000–100,000P by year-end. Through its "QuantumPai" computing cloud platform, it integrates resource scheduling with application services. The company has signed more than 200 enterprise clients, alongside over 3,000 OPC and individual customers.

The sustained strength of the computing chain reflects AI capital expenditure expansion in microcosm, with capital deployment extending from chips to servers, data centers, and computing operations. For companies, scale is merely an entry ticket—order support, delivery capabilities, resource utilization efficiency, and profitability are the variables that ultimately determine long-term value.