Guotai Fund: BIS Remote Computing Power New Rules Take Effect, Computing Power Sector Sees Resonance of Self-Controllability and Landscape Reshaping
The U.S. Department of Commerce's BIS released new remote access rules on November 16: overseas data centers providing computing power exceeding 12 million TPP (approximately equivalent to 760 H100 GPUs) to China-headquartered enterprises must first obtain a license, and advanced chip imports in Indonesia, Malaysia, Singapore, Thailand, and Brazil are likewise subject to new licensing requirements. Guotai Fund believes that the narrowing of remote computing power channels has turned self-developed computing power construction from an option into a necessity; IDC data shows that in the first half of 2026, China's AI accelerator shipments reached 2.008 million units, with the share of domestic manufacturers continuing to rise. The article also mentions that Nvidia co-founder Jensen Huang resigned as CEO, Anthropic launched its IPO roadshow, and the Ascend 950DT ten-thousand-card cluster was deployed, among other developments.
Why read this
AI score 59
The article puts the BIS remote computing power new rules, Nvidia management changes, and domestic computing power third-quarter reports together, presenting multiple facets of the changing supply chain landscape.
Topics
Body · AI translation
English text is being prepared.
[BIS remote computing power new rule lands, 12 million TPP included in licensing, how is the domestic computing power substitution logic strengthened?]
On November 16, BIS issued a rule that overseas data centers providing computing power exceeding 12 million TPP to enterprises headquartered in China must first obtain a license, regardless of whether the data center is located in the United States, Europe, or Southeast Asia; advanced chip imports in Indonesia, Malaysia, Singapore, Thailand, and Brazil are also subject to new licensing requirements. The channels for overseas Chinese customers to obtain high-end computing power in the cloud have clearly narrowed.
As remote computing power channels narrow, independent computing power construction has changed from an "option" to a "necessity." The core impact of the new rule is not a short-term computing power supply gap, but an increase in the certainty of long-term computing power self-controllability. If domestic AI companies cannot stably obtain high-end computing power through overseas data centers, the expansion and upgrading demand of the domestic chip and supporting manufacturing industry chain will shift from "policy encouragement" to "realistic constraints," and the increase in domestic computing power share will no longer be just an industry trend, but an inevitable choice driven by supply chain security.
The share of domestic AI accelerator cards continues to rise, and substitution has already been verified by data. IDC data shows that in the first half of 2026, China's total AI accelerator shipments were 2.008 million units, with NVIDIA at 678,000 units, a 34% share, down 44% year-on-year; Huawei at 572,000 units, a 28% share; Pingtouge, Cambricon, Kunlunxin, and Hygon at 7%, 5%, 4%, and 3%, respectively, and 9 of the 11 domestic manufacturers saw year-on-year shipment growth. The change in share structure shows that domestic computing power has moved from "usable" to "large-scale procurement," and the certainty of the industry chain has significantly increased.
Domestic cloud vendors and model vendors will place greater emphasis on the autonomy and security of the supply chain ecosystem. Wanlian Securities pointed out that NVIDIA's management change combined with the new BIS rule will lead domestic cloud vendors and model vendors to place greater emphasis on the autonomy and security of the supply chain ecosystem, which is expected to create conditions for a further increase in domestic computing power share. As cluster construction on different platforms advances, segments such as servers, network equipment, optical interconnects, power supply, and cooling are all likely to see changes in supply share, and manufacturers with multi-platform adaptability, customer validation, and large-scale delivery capabilities are expected to maintain an advantage in share competition.
[NVIDIA CEO change, Anthropic IPO roadshow, why has the overseas computing power competitive landscape changed?]
On November 12, NVIDIA announced that co-founder Jensen Huang would resign as CEO, and CFO Kress would serve as interim CEO. Anthropic officially launched its IPO roadshow, with a fully diluted valuation of up to about $1.4 trillion, and as of the end of September, computing power and infrastructure purchase commitments and leases totaled about $895.4 billion. Cisco raised its full-year AI infrastructure revenue forecast for hyperscale cloud vendors from $7.5 billion to $9 billion.
NVIDIA's management change may prompt changes in the competitive landscape of the AI computing power industry. Google TPU, AMD GPU, and custom ASIC solutions involving Broadcom are expected to gain more validation and procurement opportunities. As cluster construction on different platforms advances, segments such as servers, network equipment, optical interconnects, power supply, and cooling are all likely to see changes in supply share, and manufacturers with multi-platform adaptability, customer validation, and large-scale delivery capabilities are expected to maintain an advantage in share competition.
The purchase commitments disclosed in Anthropic's IPO increase the visibility of AI computing power infrastructure investment. Of the approximately $895.4 billion in purchase commitments and leases, the vast majority is for computing power and infrastructure, with about $69.8 billion due within the next 12 months. Partners in the computing power chip, storage, wafer foundry, and cloud services fields are deepening commercial cooperation with Anthropic through equity investments, further locking in supply, which helps enhance the stability of long-term computing power supply and gives the market a more complete assessment sample of the sustainability of AI infrastructure investment.
Cisco raises its AI infrastructure revenue forecast, and demand for AI computing power infrastructure is still in an expansion phase. Cisco raised its AI infrastructure revenue forecast for hyperscale cloud vendors for fiscal year 2027 from the previous $7.5 billion to $9 billion; in the previous fiscal year, revenue on this basis was about $4 billion, and if the new target is achieved, the scale will be more than double that of the previous fiscal year. As the scale of training and inference clusters expands, the bandwidth demand for high-speed interconnects within clusters and cross-data-center transmission rises simultaneously, which is expected to drive continued growth in demand for high-speed switches, optical modules, and coherent optical transmission equipment.
English text is being prepared.
Hygon Information's revenue for the first three quarters was 14.280–15.580 billion yuan, up 50.47%–64.17% year on year; Inspur Information's net profit attributable to shareholders was 3.8–4.1 billion yuan, up 156%–177% year on year; and Sugon's net profit attributable to shareholders was 1.724 billion yuan, up 78.5% year on year. On October 16, China Telecom and Huawei launched the world's first Ascend 950DT ten-thousand-card cluster in Inner Mongolia, and China Mobile launched an Ascend 950DT thousand-card super node in Beijing.
The localization process has shifted fully from policy-driven to market-driven. Hygon Information's announcement noted that downstream customers' willingness to purchase domestic CPUs and DCUs and their stickiness will continue to strengthen. The domestic computing power industry is entering the stage of performance delivery, and the certainty of the industrial chain has significantly increased. Profit growth is significantly faster than revenue growth; Inspur Information's profit growth is significantly faster than its revenue growth, reflecting the product mix optimization brought by the increased share of AI server revenue, as well as the dilution of expense ratios under economies of scale.
Super nodes expand computing units from a single card to an entire rack and even a cluster, with system integrators and ODM manufacturers benefiting directly. Huawei's super node scale has been upgraded from 384 cards to 8,192 cards, and the next-generation 960 super node debuts the NPO solution and can scale to 4,096 cards. System integrators and ODM manufacturers such as Inspur Information, Sugon, and Unisplendour, as the bearers of system integration, benefit directly from large-scale deployment. The focus of competition has shifted from "whether a single card exists" to "system integration capability," and the collaborative optimization of chips, interconnects, optoelectronic co-packaging, advanced packaging, power supply and cooling, and scheduling software has become key.
The gap in single-card hardware parameters is narrowing, and what truly determines actual output is the efficiency of computing power utilization. Soochow Securities noted that from 384 cards to 8,192 cards, from a single rack to a ten-thousand-card cluster, and from electrical interconnects to NPO/CPO, the technology roadmap is still iterating rapidly. The key to competition is no longer "whether domestic cards exist," but system-level integration capability, which places higher demands on links such as system integration, interconnects, optical modules, and power supply and cooling, and also brings incremental demand across the entire computing hardware chain.
[Guotai ChiNext Computing Power ETF (158055) covers communications, electronics, computers, and power supplies, capturing the resonance of independent controllability and landscape reshaping]
Guotai ChiNext Computing Power ETF (158055) tracks the ChiNext Computing Power Infrastructure Index, covering core links such as communications equipment, IT services, semiconductors, and power supply equipment.
It covers the entire computing hardware industry chain and combines growth elasticity with structural balance. Looking at the index constituents by sub-sector, optical communications accounts for 20.46%, network equipment for 16%, and PCBs for 16%; in addition, there are software, cloud computing, and GPUs, as well as some distribution in power supplies, storage, and liquid cooling. Compared with the STAR AI Index, which is concentrated in the semiconductor sector, and the cloud computing index, which leans toward the application layer, this index achieves full-chain coverage of computing hardware.
Second, resonance across multiple points provides performance support, and the third-quarter reports of domestic computing power players have begun to deliver. Orders for optical components are scheduled through 2029, NPO is expected to be delivered next year, the third-quarter reports of domestic computing power players are being delivered intensively, and the large-scale deployment of super nodes is pushing competition from single cards to system-level integration, with resonance across optical communications, PCBs, and storage. The high performance growth of domestic computing power names such as Hygon Information, Inspur Information, and Sugon verifies the trend of domestic substitution shifting from policy-driven to market-driven, and also provides support for the earnings expectations of the index constituents.
Note: Mentioned individual stocks are only for industry event analysis and do not constitute any individual stock recommendation or investment advice. Short-term index performance and similar movements are for reference only, do not represent their future performance, and do not constitute a commitment or guarantee of fund performance. Views may be adjusted as the market environment changes and do not constitute investment advice or commitments. The risk-return characteristics of the mentioned funds differ, and investors are kindly requested to carefully read the fund legal documents, fully understand the product elements, risk levels, and income distribution principles, choose products that match their own risk tolerance, and invest prudently. For fund fee rates involved, please consult the legal documents.
AI Trends