BOCI has published a research report indicating that the domestic AI computing power supply chain is transitioning from a phase of thematic anticipation to one of tangible earnings delivery. The brokerage believes that upward revisions to capital expenditure by domestic CSP vendors such as Tencent, coupled with price increases for large language models led by DeepSeek, are emerging as two core engines driving sustained demand for domestic computing power. The industry chain is poised to enter a new stage characterized by robust demand, accelerated supply expansion, and continuous earnings release. Based on recent interim report feedback, earnings at the foundational hardware level, including core chip manufacturing, AI computing chips, switch chips, and server/switch assembly, are now entering the delivery phase. The combination of LLM price hikes and CSP capex increases serves as the primary engine for sustained growth in domestic computing power. BOCI asserts that the domestic AI computing power supply chain has entered a period of high certainty and high-growth realization.

The expansion trend in domestic CSP capital expenditure is now firmly established. In the first half of 2026, Tencent's capital expenditure reached RMB 84.72 billion, a substantial 82% year-on-year increase, with a significant portion directed towards expanding computing power procurement. Given that Tencent has historically maintained relatively modest AI-related capex among domestic CSP players, this scale of expansion further reinforces the overall upward trend. The brokerage anticipates that other CSP vendors, including Alibaba and ByteDance, may also enlarge their capital expenditure budgets this year. The increased investment in AI infrastructure by leading cloud providers like Tencent reflects growing demand for LLM applications and cloud services. This also signals that domestic AI computing cluster construction is evolving from isolated projects to large-scale, sustained procurement. Considering that CSP capital expenditure typically leads server shipments and data center construction, these upward revisions are expected to transmit sequentially through to AI chips, servers, switches, and network interconnect components, thereby enhancing order visibility across the supply chain. With internet firms, telecom operators, and industry clients collectively raising their AI-related investments, the demand base for domestic computing power is diversifying, which should further bolster the sustainability and certainty of the sector's prosperity.

The recent significant price adjustment by DeepSeek for its API services serves as evidence that the economic value of computing power is improving. DeepSeek has adopted a peak-valley pricing model, with the flagship model DeepSeek-V4-Pro now priced at RMB 27 per million output tokens during peak hours, an approximate 350% increase from current levels. During off-peak periods, the price is RMB 13.5 per million output tokens, representing 2.25 times the existing rate. The industrial signal from this price increase is strong. On one hand, LLM inference costs rise alongside increases in parameter scale, context length, and user invocation volumes; price adjustments help improve model vendors' per-unit computing power returns and commercialization capabilities. On the other hand, enhanced model capabilities and maturing payment models are pushing the LLM industry from a "low-price customer acquisition" strategy towards "value-based pricing," thereby improving the commercial viability of AI services. BOCI believes that rising model call prices do not indicate weakening computing power demand. Instead, they validate the tight supply-demand dynamics for high-performance models and the scarcity of premium computing resources. As inference demand expands from internet applications into office, finance, industrial, and government sectors, computing power requirements will shift from a training-centric focus to a balanced emphasis on both training and inference. This will drive continuous infrastructure expansion across servers, switches, optical interconnect, PCBs, high-speed connectors, power supplies, and liquid cooling systems.

Domestic semiconductors are set to embark on a growth trajectory driven by AI, with equipment expansion leading the way. From an industry chain perspective, the primary constraint on domestic computing power currently lies not in demand but in the insufficient supply of high-end chips and advanced manufacturing capabilities. The accelerated iteration of AI computing chips will correspondingly drive expansion in wafer fabrication, advanced packaging, equipment, components, and materials. The front-end process benefits from the construction of advanced and specialty process capacities, while the back-end process benefits from the heightened demand for advanced packaging, packaging substrates, and high-density interconnect driven by high-performance computing chips. As the leading indicator of capacity expansion, semiconductor equipment is expected to see order momentum materialize first, subsequently transmitting to core components, thin-film deposition, etching, cleaning, inspection, and packaging equipment. Additionally, catalysts such as memory industry expansion and the anticipated listing of Yangtze Memory Technologies Co. are likely to raise the profile of domestic equipment and materials sectors.

Key risks include domestic internet companies and telecom operators potentially missing capital expenditure expectations, growth in AI applications and token demand falling short of projections, and domestic AI chip performance iteration or capacity release underperforming expectations.