China's Computing Power May Reach a Historic Inflection Point

Affected by the release of DeepSeek V4, the domestic compute power sector is celebrating today, with Hua Hong Semiconductor and SMIC's Hong Kong stocks both exceeding 10%. Hygon Information rose by over 8%, and Cambricon increased by over 2%.

The most core architecture change in V4 lies in "replacing calculation with storage", through a memory mechanism called Engram, moving about 80% of static knowledge out of expensive graphics card memory, into lower-cost system memory and hard drives, with less than a 3% throughput loss.

With the rapid expansion of caches brought by millions of contexts, the hardware bottleneck in AI inference is shifting from "piling more memory" to "using larger and cheaper memory". Calculations have shown that servers using new memory expansion solutions can expand model parameter capacity from 671 billion to 1.2 trillion, improve inference performance by 20%, and increase throughput by up to 43%.

This means that the proportion of memory chips, storage devices, high-speed interconnects, and other links in AI hardware expenditures will rise structurally.

The most critical industry signal in the V4 release is that DeepSeek officially stated that due to high-end computing power limitations, the throughput of Pro services is very limited, and the price is expected to drop significantly after the mass availability of Ascend 950 supernodes in the second half of the year. This indicates that domestically produced chips are deeply involved in the inference deployment of cutting-edge models, moving from "being able to run" to "commercial use", and DeepSeek is actively integrating its product roadmap closely with the pace of domestic supply.

The scale of trillion-level parameters is also forcing server deployment to be upgraded from individual cabinets to ultra-large scale clusters, which poses higher requirements for whole machine delivery capabilities.

SMIC, as the core manufacturing platform for domestic advanced processes, directly undertakes the expansion needs of domestic compute power chips, and is an irreplaceable manufacturing base in the "national model, national core" closed loop.

Hua Hong Semiconductor focuses on specialty processes, with production advantages in power devices, analog chips, and other AI server supporting links, also benefiting from the scale expansion of domestic compute power clusters.

Compute power chips can pay attention to Cambricon, Hygon Information, and Xirui Shares.

Suppliers related to Ascend, pay attention to Aerospace Appliance and Huafeng Technology, the full-year shipment expectation is revised to 1.1 to 1.2 million chips.

For chip manufacturing and packaging, focus on SMIC, Silex Microelectronics, and Jiangsu Changjiang Technology. For memory and storage, pay attention to Lanti Technology, GigaDevice, and Biwin Storage.

For server whole machines, focus on Huaxin Technology, Sugon, and Inspur Information.

Unlike the sharp market fluctuations caused by R1 in early 2025, V4 is more about confirming trends rather than surprises, but it truly establishes the turning point for domestic compute power to enter performance verification from concept speculation.<\p>