DeepSeek V4 “杀疯了”：大大提升了国产算力的天花板？
文章语言：
简
繁
EN
Share
Minutes
原文
会议摘要
The dialogue focused on the better-than-expected performance of semiconductor equipment ETFs, recommending attention to storage expansion investment opportunities, while reminding the high-rise sector volatility risk, advocating low or batch layout strategy. The speech discussed in depth the development of domestic and overseas computing power, especially the AI chip market, and predicted that global AI capital expenditure would reach US $793.9 billion in 2026, a growth rate of 45%. Analysis of TSMC, Nvidia and Google TPU roles, emphasizing the growth potential of optical modules, semiconductor equipment and storage market, domestic computing power in the policy support of rapid growth. Recommend ETF products as investment tools, calling for cautious pursuit, grasp the AI large model training and reasoning application prospects.
会议速览
TSMC Revenue and Global AI Capex Growth Correlation AnalysisBy analyzing TSMC's revenue in NVIDIA GB200 and Rubin NVL72 racks, the growth trend of capex in the global AI industry chain is discussed. As a contract manufacturer, TSMC's revenue changes reflect fluctuations in AI chip demand, especially the improvement of GPU and advanced packaging technology, which indicates a significant increase in global AI Capex.
Technology Evolution and Market Prospect of Nvidia and Google in the Field of AI ChipDiscusses the NVIDIA chip roadmap, including the technical details of rubin and rubin ultra, such as the challenges and applications of the fourth generation and wind packaging technology. At the same time, it analyzes the development of Google TPU in the field of AI, emphasizes its competitive situation with Nvidia, and the high requirements of AI chips for TSMC packaging technology, and looks forward to the future market trend of AI chips.
TSMC AI chip foundry revenue and global capital expenditure correlation analysis.Discusses the reasons for Google's own TPU, including hardware and software collaboration and cost savings. The multiplier effect of TSMC's AI chip foundry revenue and global capital expenditure, as well as the proportion of AI hardware in data center construction, reveals the close relationship between TSMC's revenue and global AI investment.
Global AI Capital Expenditure Growth: Forecast and Analysis from 2025 to 2027The dialogue analyzed the rapid growth trend of global AI capital expenditure, which is expected to reach $544.1 billion, $793.9 billion and $1139.8 billion respectively from 2025 to 2027, with a growth rate of about 45%. TSMC's AI chip revenue guidance shows that it will maintain a compound annual growth rate of 55%-60% from 2024 to 2029, indicating that AI capital expenditure may be close to $2 trillion in 2029. Companies such as Google have raised capital spending, reflecting strong market demand and sufficient backlog of orders to support future capital spending growth.
Prospect Analysis of Optical Module and Semiconductor Industry Driven by AIThe dialogue discussed the impact of AI on the optical module and semiconductor industry, and pointed out that optical modules have become indispensable components due to the complexity of data center networking, and the market size is expected to grow rapidly. At the same time, the semiconductor industry was catalyzed by AI, especially the strong performance of GPU and storage, the domestic computing market also ushered in a positive.
Growth Momentum and Localization Challenge of Semiconductor Equipment IndustryThe semiconductor equipment industry has benefited from the growth of the global market and the demand for independent control, and the domestic growth rate is rapid. Lithography, etching and thin film deposition play an important role in wafer manufacturing equipment. Breakthroughs in advanced processes, especially in consumer electronics and AI, have become key to the long-term growth of the industry. Domestic equipment manufacturers need to upgrade technology and narrow the gap with international leading companies to meet the growing market demand.
AI-driven storage demand surges, domestic GPU and computing power development discussion.This paper discusses the driving effect of AI on storage demand, and points out that GPU computing relies on a large number of storage resources, which leads to the rise of storage price and the trend of global storage expansion. Mention that the expansion of domestic storage companies will benefit semiconductor equipment enterprises, emphasizing the investment value of semiconductor equipment as a direct beneficiary of storage expansion. This paper analyzes the domestic GPU market potential and technology gap, and puts forward that through the advantages of abundant power and investment kinetic energy, the successful practice of the DeepSeek model provides the possibility for domestic GPU applications, but in the short term, we need to be alert to the risk of excessive market growth.
Overseas computing power and semiconductor industry chain ETF investment analysisThe dialogue discussed in depth the investment opportunities of overseas computing power and semiconductor industry chain, recommended communication ETF515880, chip ETF, semiconductor equipment ETF159516 and integrated circuit ETF, stressed that these ETF can effectively cover the core links of optical module, optical fiber, server, chip industry chain, etc., showing strong prosperity and performance, and suggested that investors pay attention to its development potential in overseas computing power and semiconductor industry.
In-depth analysis of the investment strategy and market performance of chip ETFs and CRE chip designs.This paper discusses in detail the characteristics and investment value of chip ETF, integrated circuit ETF, semiconductor equipment ETF, scientific innovation chip and scientific innovation chip design ETF, emphasizes the importance of listed semiconductor companies on scientific innovation board, suggests investors to pay attention to chip design, storage, GPU, CPU and other sectors, and reminds them to be cautious when chasing high prices, and low or batch layout is more suitable for ordinary investors.
要点回答
Q：In this round of semiconductor market, when will overseas computing power begin to recover and form a resonance market with domestic computing power?
A：The recovery of overseas computing power began in April with a global geopolitical easing and an upward trend in equity market risk investment appetite, which together with domestic computing power formed a significant market recovery.
Q：Why is Taiwan's TSMC revenue relevant to the global AI industry chain?
A：As a large global semiconductor foundry, TSMC undertakes the foundry tasks of GPU, TPU and other AI chips for many North American companies such as Nvidia, AMD and Google, so its revenue situation can better reflect the supply of the entire AI industry chain.
Q：Which of the current mainstream AI chip maker Nvidia's chips accounts for the largest share of TSMC's foundry revenue?
A：Nvidia's Blackwell chips account for a larger share of TSMC's foundry revenue, especially its front-end processes and advanced packaging.
Q：What is TSMC's single-card revenue in the NVIDIA GB200MVL72 rack?
A：In the GB200MVL72 rack, TSMC can obtain a single card revenue of about US $2300 through contract manufacturing related businesses, of which the single card revenue is about US $2000 when only considering GPU-related front-end processes and advanced packaging.
Q：Nvidia's new generation of Rubin rack is expected to bring what kind of revenue growth?
A：With the introduction of Rubin rack by Avida in the second half of the year, TSMC will see a rapid increase in its GPU-related revenue. It is estimated that a Rubin rack can bring TSMC about US $260000 in revenue and a single card about US $3562 in revenue.
Q：Nvidia's chip roadmap shows what kind of development trend?
A：Nvidia's roadmap shows that the Rubin rack will be launched in the second half of this year, and the Rubin Ultra chip may be launched next year. Its chip area is about twice as large as Rubin, and it will use four generations of wind technology to achieve more efficient inter-chip communication and lower power consumption.
Q：For TSMC, what are the challenges of multi-generation wind technology to its packaging technology?
A：If the Rubin Ultra chip fails to adopt the fourth-generation wind technology next year, it may lead to more AI chip shipments using the second-generation wind technology. But in any case, multi-generation wind technology as an industry trend will be landed sooner or later, which will put forward higher requirements for TSMC's packaging technology.
Q：What role does Google TPU play in the development of AI?
A：Google's TPU is one of the important driving forces in the field of AI, Google as a leader in the AI circuit, its proposed transformer architecture is widely used in the current AI model, has a profound impact on the development of global AI technology.
Q：What adjustments has Google made in the development of large models, and how competitive is its TPU chip?
A：Google quickly adjusted the company's route during the rapid development of large models. Its TPU chips have achieved significant growth since their early launch. It is currently competitive with Nvidia in the field of large models. This year, the market is expected to ship nearly 4 million chips, and it is expected to further improve next year. Google developed TPU for two main reasons: first, to achieve R & D synergy between its hardware and software, and second, to reduce costs by replacing expensive NVIDIA GPUs with self-developed chips.
Q：What is TSMC's revenue from AI chips? How does TSMC's revenue from foundry AI chips affect global spending on AI hardware infrastructure?
A：TSMC's revenue from AI chips is expected to be about US $more than 35 billion this year, about US $24 billion last year, and is expected to reach about US $52.8 billion by 2027, accounting for about 23% to 24% of its total revenue. TSMC completes the foundry and hands over the chips to downstream assemblers, who then hand them over to Nvidia for sale to customers. Among them, TSMC's foundry revenue corresponds to the value of AI chips, while Nvidia sells at a higher price on this basis to earn more profits. In addition, cloud vendors will invest more money to purchase switches, assemble servers, etc. after purchasing AI GPUs, forming a large-scale AI hardware infrastructure expenditure.
Q：What is the relationship between AI chip revenue and global AI capital expenditure?
A：TSMC's AI chip foundry revenue is multiplied by a multiplier effect (about 70% to 80% of data center capital expenditure is spent on hardware construction) and a hardware expenditure multiplier (about 1/70 to 1/80) to arrive at the forecast value of global AI capital expenditure.
Q：What is the growth trend of global AI capital expenditure (CAPEX)?
A：Global AI capital expenditure has maintained a high growth rate, which is expected to reach US $544.1 billion in 2025, US $793.9 billion in 2026 and US $1139.8 billion in 2027, with a growth rate of about 45%. The high visibility is due to the long manufacturing cycle of TSMC chips, which can see the demand in the next few quarters in advance, while the current revenue corresponds to the capital expenditure after three quarters.
Q：In terms of technology investment, especially in the semiconductor industry, what is the future market demand?
A：According to our estimates, global capital expenditure (capex) growth will remain at about 45% in 2026 and 2027. In addition, in the recent performance meetings of some technology companies in North America, their median capital expenditure has been raised. For example, Google raised the median capex by 10 billion US dollars to about 185 billion US dollars. This may mean that I have underestimated the demand forecast for this year and next year, and the actual demand may exceed the expectation.
Q：Judging from the recent North American earnings report, what are the positive signals on the demand side?
A：The recent North American financial report shows that the outstanding commercial orders of many companies are growing very fast. For example, Google reached the order level of about $460 billion in one quarter, and its backlog of orders is enough to cover the capital expenditure in the next few years, which strongly supports their investment needs. Therefore, there is no need to question the current investment in AI areas, and the industry trend is still strong.
Q：Why are optical modules so important in data center networking?
A：Optical modules are very important in the data center network, because the internal network of the data center is complex and large, and the use of traditional copper cable transmission will cause problems such as heat generation, electrical signal loss and data loss. The optical module has the advantage of high-speed transmission, especially suitable for large-scale data center networking needs, is an indispensable device.
Q：How is the growth of the optical module market?
A：The size of the optical module market is growing with the development of AI technology and the growth rate of capital expenditure. The market size is expected to be about US $13.8 billion in 2025 and US $34 billion in 2026, a growth rate of about 150 per cent. The actual market size may be more optimistic than the measured value due to factors such as stock demand and wear and tear, and is expected to reach a growth level of 170 to 200 per cent.
Q：What are the main drivers for the semiconductor industry, especially in the field of semiconductor equipment?
A：The semiconductor industry is being catalyzed by AI, in which the growth in demand for semiconductor equipment is closely related to the expansion of production capacity by chip manufacturers such as TSMC. The sales scale of semiconductor equipment market has steadily increased, and the growth of domestic semiconductor equipment is particularly significant due to factors such as independent control and industry scale expansion. The main demand for semiconductor equipment comes from wafer manufacturing equipment, such as lithography, etching and thin film deposition, while the upgrading of advanced process technology and the demand pull in consumer electronics, AI and other fields are important sources of its growth momentum.
Q：What is the current level of Apple's chip process, and what are the gaps between the domestic foundry and it?
A：Apple's A18 chip uses the 3nm process of Taiwan's TSMC, while domestic foundry companies still have a big gap with TSMC in advanced processes. For example, domestic mobile phone manufacturers may need to face more challenges if they are looking for semiconductor foundry to produce similar-level chips in China.
Q：In the field of AI, what are the advances in NVIDIA's GPU process and what is the status of domestic GPU companies in terms of process?
A：Nvidia's GPU process is more advanced. Its GB200 chip uses a 4nm process. It is expected that rubin in the second half of this year may be upgraded to a 3nm process. In contrast, the products of domestic GPU companies are mainly in the process of 7 to 14 nanometers, showing a clear intergenerational gap with Nvidia. Nevertheless, domestic GPU production capacity is still scarce. Due to the rapid growth of GPU demand, advanced processes have a driving effect on the long-term growth of the semiconductor equipment industry.
Q：Why has storage become an important concern in the near future, and how has the development of AI affected it?
A：The growth of AI has had a huge impact on storage requirements, especially in data centers and GPU applications. The need to quickly extract, compute, and store large amounts of data during the operation of AI models has led to a dramatic increase in the demand for high-bandwidth storage (such as HBM) and various types of memory (such as DDR). The rapid ramp-up of AI GPUs has led to simultaneous growth in supporting storage, which is the main reason for the rise in storage prices. The explosion of AI demand has prompted global storage expansion, not only overseas storage companies to expand production, domestic storage companies such as two-tier also listed and expanded production, which will further bring a large number of orders for domestic semiconductor equipment enterprises.
Q：What are the recommendations and reasons for investing in semiconductor equipment?
A：The current semiconductor equipment industry is highly catalyzed, especially benefiting from the expansion of the storage market. Investing in semiconductor equipment ETFs provides a higher degree of benefit and flexibility than ETFs with more storage content. In addition, although the domestic GPU market is growing rapidly, there is a gap between the technical level and North America. However, domestic power is abundant and investment momentum is strong, and domestic computing power can be developed by stacking quantities. Taking the deep-sea series GPU as an example, its release marks the possibility of large model training and reasoning based on domestic chips, and provides a practical basis for the rapid rise of domestic computing power.
Q：Representative products of overseas computing power and why communication ETF515880 are recommended?
A：If you want to invest in overseas computing power, communication ETF515880 is an important choice. The ETF includes core links such as optical modules, optical fibers and servers, of which optical modules account for more than 50%, which can fully reflect the development of overseas computing power as a whole. Compared to QD or other ETFs, communications ETFs are more representative of the fundamentals of overseas computing power and are strong in terms of performance and prosperity, so they deserve continued attention from investors.
Q：What have communications ETFs risen since the 2023 AI hit?
A：As of yesterday, the communications ETF has risen by more than 400, or quadrupled. In 2025, the ETF also rose by 128 percent, compared with about 200 percent in the two-year period from 2025 to 2026.
Q：How does the chip ETF cover the entire semiconductor industry chain? What are the characteristics of the IC ETF?
A：Chip ETF basically covers the whole chip industry chain, including equipment, materials, design, manufacturing, packaging and testing, etc., and it selects high-quality semiconductor companies in the whole market to ensure that the ETF can accurately represent the whole industry. Integrated circuit ETF is also a market-wide selection of chip companies, but it does not include semiconductor equipment companies. The main focus is on the performance of market-wide chip design companies. Although it also includes some manufacturing companies, the content is low.
Q：Which companies are selected for the Semiconductor Equipment ETF?
A：The Semiconductor Equipment ETF focuses on selecting companies related to semiconductor equipment and semiconductor materials, which can directly reflect the product boom of both advanced processes and storage expansion logic, and can be considered for investors who are bullish on storage and unfamiliar with individual stocks.
Q：What is the difference between a CRE chip ETF and a CRE chip design ETF?
A：The selection of semiconductor companies by CRE ETF on CRE Board can reflect the prosperity of semiconductor companies listed on CRE Board. However, CRE Chip Design ETF (Code 589260) is specially selected for companies listed on CRE Board and focusing on chip design. If you like chip design sectors such as storage, GPU and CPU, CRE Chip Design ETF is a product worthy of attention.
Q：What should I be aware of when investing in semiconductor-related ETFs?
A：Although there is more good news on the macro level, some sectors may be volatile. For investors who want to invest in the semiconductor industry, it is recommended to adopt a low-level layout or batch layout, and be cautious about chasing high operations to be more suitable for the risk tolerance of ordinary investors.
