- The preview version of DeepSeek-V4 is now officially launched and open-source, supporting a super long context of millions of words. Its core technological breakthrough lies in achieving deep adaptation to domestic chip architectures like Huawei's Ascend for the first time.
- Boosted by accelerated expectations for the de-CUDA process of computing infrastructure, China's A-share GPU and semiconductor index soared during Friday's trading, with XGPT Information (688041:CH) rising over 9%, followed by other industry chain stocks such as SMIC (688981:CH).
- Institutions like CITIC Securities (600030:CH) have upgraded their long-term industry guidance, predicting a year-on-year (YoY) growth rate of over 100% in shipments of domestic computing power chips by 2026, significantly driving demand for advanced packaging and switching chip orders.
Acceleration of Domestic Computation Alternatives
The V4 series models released by DeepSeek come in two versions, Pro and Flash. Their most significant industry impact is the demonstration of the feasibility of top-parameter large models within non-NVIDIA (NVDA:US) ecosystems. For a long time, the CUDA ecosystem barrier has been a key pain point restricting the large-scale commercial use of local computing power in China. This breakthrough in connecting DeepSeek-V4 with Huawei's Ascend architecture signifies that the underlying compiler and operator library adaptation has reached a commercial-use standard. If the Ascend 950 supernodes can be delivered in bulk as scheduled later this year, the inference cost for the Pro version is expected to drop significantly, offering enterprises more cost-effective computing power options.
Market Liquidity and Capital Contest
Observing the capital flow on the trading floor, Friday saw a significant net inflow of funds into the semiconductor and computing hardware sectors of the A-share market. As a representative of domestic X86 and DCU architectures, XGPT Information's intraday gains briefly exceeded 10%, reflecting the market's optimistic pricing for the expansion of the local AI chip market share. Furthermore, the overall rise of stocks like Loongson (688047:CH), Cambricon (688256:CH), and Moore Threads illustrates that funds are not restricted to a single technology route but are systematically allocating across the entire self-controlled computing power pool. In the current environment of relatively loose macro liquidity, technology hardware sectors with a high certainty of order growth expectations have become reservoirs for incremental funds.
Pricing Logic and Forward Valuation Reevaluation
Brokerage reports indicate that the explosion of multimodal applications and agents is fundamentally changing the model of computing power consumption. The exponential growth in token invocation volume has widened the supply-demand gap on the inference side, providing a strategic capacity filling window for local manufacturers. CITIC Securities expects that shipments of domestic computing power chips will more than double within the year, prompting sell-side analysts to revise upward their forward price-to-earnings (PE) and price-to-sales (PS) models for related companies. Due to the high capital expenditure barriers in hardware iterations, companies that can achieve volume production early and establish a software-hardware ecosystem stand to enjoy higher valuation premiums in the future.
Race between Computing Power Bottleneck and Capacity Expansion
Despite the explosive demand on the demand side, the actual delivery capacity of the supply chain still faces multiple constraints. The bottleneck for high-end chips currently lies not only in the design of logic chips but also in the elasticity of domestic foundry and advanced packaging capacity following restrictions on overseas fabs like TSMC (TSM:US). The market performance of companies like Tongfu Microelectronics (002156:CH) in testing and packaging reflects the pricing of funds on the scarcity of high-density packaging capacity such as CoWoS. If the yield rate of wafer-level packaging climbs as expected, the overall gross profit margin level of the domestic computing power industry chain is expected to see substantial recovery in the next two quarters.
