As of the market close, the Sci-Tech Innovation 50 Index rose by 8.59%, reaching a new high not seen in nearly three and a half years. Industry leaders such as Cambricon and Haiguang Information saw their stocks hit the daily limit of 20%, leading to a frenzy across the entire semiconductor industry, with a net capital inflow exceeding 15.8 billion in a single day.The surge in semiconductors is primarily attributed to two factors.
First, last night, the domestic large model enterprise DeepSeek released version 3.1, which for the first time supports FP8 precision and the next generation of domestic chip architecture, marking the initial formation of a “domestic chip – domestic model” technology closed loop.
This means that as long as we are willing, we can achieve complete autonomy in the field of large models.
As we all know, we are relatively weak in the field of high-end chips, so currently, most domestic AI large models rely on imported chips from companies like NVIDIA for training.
However, due to the competitive relationship between China and the U.S., the U.S. has been restricting us in the high-end chip sector, leading our tech companies to use NVIDIA’s stripped-down low-end chips.
Therefore, the support for domestic chips by this domestic large model can be seen as a strong boost in this field.
From a practical application perspective, the Cambricon Siyuan 590 chip has completed adaptation work with the aforementioned V3.1 version, and its inference performance has improved by 40% compared to the previous generation; the Haiguang Information Deep Computing No. 2 chip has achieved international mainstream levels of computing density at FP8 precision and is now included in the procurement lists of several national supercomputing centers.
As the computing model continues to upgrade and expand, the upstream and downstream of the industry chain, whether it be liquid-cooled servers or chip packaging and materials, will inevitably welcome a vast domestic replacement market. Just the replacement demand alone will bring substantial profits to related enterprises.
Second, last night, it was revealed that NVIDIA’s H20 chip has been suspended from production.
The H20 chip is based on NVIDIA’s Hopper architecture, utilizing advanced CoWos packaging technology, equipped with 96GB of HBM3 memory and 4.0TB/s bandwidth, with performance reduced by about 28% compared to the H100 chip, specifically designed as a “stripped version” AI chip to comply with U.S. export restrictions on chips to China.
In April, during the phase of U.S.-China trade conflict, the U.S. government decided to prohibit NVIDIA from selling the H20 chip to the Chinese market to curb China’s rapid development in the AI field.
On July 15, NVIDIA CEO Jensen Huang visited China for the third time this year, emphasizing “high importance on the Chinese market” to secure domestic orders, and the H20 chip, which had been halted for three months, resumed sales to China.
On July 31, the National Cyberspace Administration of China interviewed NVIDIA, requiring it to explain the security risks of backdoor vulnerabilities in the H20 computing chip sold to China and submit relevant proof materials, directly addressing issues of “tracking and positioning” and “remote shutdown”.
The incident originated in May when several U.S. lawmakers jointly proposed a bill called the “Chip Security Act,” which primarily requires all high-end AI chips subject to export controls to be equipped with location verification mechanisms within 180 days, and further demands research into secondary security measures, including a “one-button destruction switch” for remote failure or performance degradation.
U.S. lawmakers also stated that NVIDIA’s “technology” has matured.Although there have been no publicly confirmed large-scale cases of exploiting GPU hardware-level backdoor vulnerabilities, such operations are entirely feasible in reality, given the extremely complex architecture of GPUs, most firmware being closed-source, and the lengthy verification chain. Once the signature system is breached, remote persistent control can be achieved, technically possible.
And once the chips we purchased at a high price are controlled on a large scale through backdoor vulnerabilities during a critical phase of U.S.-China confrontation, the consequences would be unimaginable.
Therefore, the adaptation of domestic large models to domestic chips, combined with the revelation of NVIDIA’s H20 chip being suspended from production for the Chinese market, provides ample imagination for domestic semiconductor replacement, pushing it directly to the core of the market.