Source: Tianfeng Securities, Author: Pan Yin, Luo Yiyang
1. It is expected that the semiconductor cycle will start to recover in 2024
The average sales of global semiconductors in March increased year-on-year for four consecutive months
Causes of cycles: The unpredictability of supply and demand trends
• Unpredictability of future demand:
It is often new large products that come out, and large manufacturers do not have sufficient goods in advance, such as the demand for semiconductors driven by the replacement of smartphones in history
• Unpredictability of future supply (competitive landscape):
For example, after Huawei was sanctioned, various mobile phone manufacturers stocked up a lot of goods to grab share, which eventually led to oversupply in the industry, such as the shortage of power semiconductors, which led to the industry's oversupply
• Occurrence of unpredictable events:
For example, the Sino-US trade conflict, the disruption to the supply chain (rapid increase in overseas product stocking while speeding up domestic verification), such as the disruption caused by the epidemic to the supply chain (increasing stocking in each link)
• Why do you make a mistake?
Lack of transparency in the industrial chain: Most of them are trade secrets, and they are even more opaque after the Sino-US trade conflict
Long whip effect: The industrial chain is long, from equipment and materials, to manufacturing, packaging and testing, to various IC designs, as well as layers of distribution, there are many links, and there is a long whip effect The unpredictability of technological progress in the technology industry: the industrial chain always believes in the trend when it sees orders
AI is expected to be the main driver of the upward trend in this cycle
The industry inventory is expected to return to the normal level in 1H24, and the AI-driven link has the motivation to replenish the inventory
It is expected that the semiconductor inventory in 1H24 will return to the normal level, and some links are expected to give priority to replenishing inventory. (1) The average semiconductor inventory cycle is 39 months (about 3 years), and the high point of the last round of inventory is in 22Q2-Q3. (2) The average global chip delivery cycle has continued to decline for nearly a year, and has returned to the level of the first quarter of 2021 in February 2024. (3) Some links (such as AI servers/smartphones) give priority to completing inventory depletion, and even start to replenish inventory.
The expansion of production capacity tends to be rational, and the popularization of new demand or the acceleration of expansion progress
(1) The capacity utilization rate of mature processes was maintained at a low level in the peak season of '23, making the expansion progress more rational, such as SMIC (3Q23 77.1%; 4Q23 76.8%), Hua Hong Semiconductor (3Q23 86.8%; 4Q23 84.1%).
(2) The number of new semiconductor factories in the world peaked at 33 in 2022 and declined slightly in 23.
(3) After the global fab capital expenditure reached a stage high in 4Q22, it declined quarter-on-quarter for four consecutive quarters, reflecting the cautious expansion of wafer fabs. If the new demand brought by AI exceeds the judgment of the industry, it may cause a new round of out-of-stock price increases.
On the price side, electronic products represented by storage began to rise in price, and market confidence was gradually established
According to flash memory market data, the NAND price index has increased by about 80% since it bottomed out in August 2023, and the DRAM price index has also increased by more than 30% relative to the bottom.
AIGC is rapidly iterating, and the direction of demand-side innovation is established
According to IDC, in the next three years, ARVR, video analysis, knowledge graph, and natural language processing will become the main focus of AI single-point technology. Video analytics, ARVR, and knowledge graphs require powerful computing power to meet the requirements of rendering, real-time video analysis, and complex computing scenarios, and are the three main single-point technologies with high computing power consumption.
The global AIGC market size was valued at USD 10.79 billion in 2022 and is expected to reach approximately USD 73.16 billion by 2030, growing at a CAGR of 27%.
Second, we are optimistic that AI will drive the semiconductor cycle upward
Semiconductor cycle: It is currently at a relative bottom, and the demand side looks at AI
• Demand side: The core is AI, AI mobile phones, AI PCs, AI + other terminal products
Increment: Focus on the increment of AI to hardware, including cloud computing power, terminal NPU, and storage space increase
In terms of stock, we will pay attention to the pull of the demand for replacement under the combination of AI hardware and AI applications (the mobile phone and PC market is expected to resume growth in 2024, and AI will accelerate growth), which will further drive the whole industry chain beyond expectations
In terms of price increases: After the relationship between supply and demand improves, the varieties of price increases brought about by the new demand for AI are worth paying attention to
• Supply side: Observe the capacity utilization rate of mature process core, and see the expansion of advanced process, as well as the progress of domestic substitution
At present, there is no shortage of mature process capacity (SMIC's capacity utilization rate (3Q23 77.1%; 4Q23 76.8%), Hua Hong Semiconductor's capacity utilization rate (3Q23 86.8%; 4Q2384.1%)), and the expansion of mature process production tends to be rational;
There is a continuous shortage of advanced processes, and the demand for computing power brought by AI is mainly reflected in the field of advanced processes, and the supply and demand pattern is expected to be good, but the expansion of production is constrained by the supply of core equipment in the United States.
Advanced process/advanced packaging, equipment/materials used, it is expected that the demand for expansion will continue to increase, and it has a good competitive pattern because of high technical barriers
• Occurrence of unpredictable events: Sensitivity to major events that may affect industry trends
U.S.-China trade uncertainties
In the AI era, storage is expected to move from "price" to "value" cycle
Storage is a cyclical industry, looking back at the 2019-2023 round of cyclical changes, experiencing oversupply, epidemic shortages, over-falling inventories, etc., and finally ending with the original factory taking the initiative to reduce production and control production. Looking forward to the new cycle from 2024 to 2026, new technologies and applications such as AI PCs will be used to stimulate the potential of storage, get out of the traditional price cycle, and enter the value cycle of new storage.
NPU helps smart terminals enter the AI era
Compared with CPUs and GPUs, NPU has obvious advantages in the following aspects: High performance: AI tasks require NPU to have high-throughput computing power to support fast data processing and real-time inference. For the next generation of AI PCs, Intel proposed a performance requirement of at least 40 TOPS at the AI Summit. It is foreseeable that AI will also put forward higher requirements for the computing power of mobile phone NPU
Energy efficiency: A high energy efficiency ratio is essential for longer battery life and less heat generation. This requires low-power design and energy efficiency optimizations such as dynamic voltage and frequency scaling (DVFS), dedicated power delivery rails, and microarchitecture optimization to achieve high performance while maintaining low power consumption
Memory management: Currently, the token generation efficiency of large language models is limited by memory bandwidth. Therefore, the NPU needs to have a high memory bandwidth to support a large number of data reads
For example, high-performance memory interfaces such as LPDDR5x are used to meet the memory requirements of AI models. At the same time, the NPU also needs to be highly memory efficient to reduce latency and power consumption for data movement, such as using techniques such as memory hierarchies to further optimize overall performance.
The Ryzen 8040 series processors are equipped with a 4nm Zen 4 CPU core architecture with up to 8 cores/16 threads and a clock speed of up to 5.2 GHz, an RDNA 3 graphics core architecture with up to Radeon 780M iGPU (12 compute units), and an XDNA NPU with 16 TOPS of AI computing power, improving AI performance by 60% compared to the previous generation. The TDP range is 15-54W.
It is optimistic that AI will promote mobile phones to open a new wave of replacement
In the face of sluggish growth in the smart terminal market, mobile phone companies hope to start a new round of growth cycle with the help of generative AI mobile phones. According to data from market research institute IDC, global smartphone shipments will decline by 3.2% year-on-year in 2023, hitting a new low in a decade. Canalys, a market research institution, predicts that the demand for end-side AI capabilities is expected to stimulate a new round of replacement, which will help increase the average selling price of devices, and AI capabilities will become an effective force for Chinese mobile phone manufacturers to promote high-end. Generative AI will bring about the biggest user interaction change since the transition from feature phones to smartphones, and is expected to become a new growth driver for the smartphone market, and will significantly change the market size and accelerate the pace of upgrades.
The penetration rate of AI PCs is expected to continue to increase, and the wave of replacement machines will drive the industry chain back to growth
AI empowers other terminals such as smart speakers/smart watches
The application of AI technology and large models in smart hardware such as smart speakers and smart watches is gradually increasing. The combination of smart speakers and AI large models has become one of the future development trends. The semantic depth understanding of models such as ChatGPT is expected to be applied to voice interaction terminals such as smart speakers to improve user experience. Smart watch manufacturers are integrating generative AI into their device ecosystems in a targeted manner, with a focus on expanding health and fitness tracking application scenarios. As vendors recognize the important role these features play in driving user engagement and loyalty, they are developing AI-driven solutions to deliver actionable insights and personalized recommendations.
Representative new products
Galaxy Ring: Samsung Electronics plans to produce 400,000 units of the Galaxy Ring, the first generation of its smart ring product, which was released for the first time this year. A smart ring is a wearable device that is worn like a ring. With the various sensors installed on the smart ring, you can check your body and health information using a smartphone app. The initial quantity is the quantity that is the first to be sold in a major global market immediately after the product is launched.
Rabbit R1: From the perspective of product form and interaction, Rabbit R1 can be regarded as a variant of smart speakers. This product has an interactive Q&A function, and after binding Uber, Spotify and other accounts, users can play music, take a taxi, order food, plan a trip, etc. through voice operation. The Rabbit R1 also features a visual recognition feature that identifies the food in the fridge and plans low-calorie dinners for the user. The product uses its own Large Motion Model (LAM) to learn and repeat what the user is doing on the PC. Users only need to give voice commands, and the Rabbit R1 can repeat the functions that were previously available on the PC.
OPPO Watch X: OPPO has added a blood glucose watch face to the OPPO Watch 4 Pro, which supports connecting to a CGM continuous glucose meter, which allows you to view blood glucose data on the watch face, and the OPPO Watch X also continues this function. OPPO Health Lab has established the industry's first digital wrist data model library to simulate blood circulation and tissue optics under different physiological states, thereby greatly expanding the sample size of wrist models and reducing the error of watch monitoring in different skin tones and wrist states.
3. Potential risk warning
The implementation of AI applications is not as expected: The promotion of AI algorithm technology and its application faces some challenges, and large models have hidden dangers in terms of data security, accuracy, ethical issues, and potential model theft and attacks. The commercialization path of AI is still unclear, the charging model of consumer-based large models is still being explored, and the charging level of enterprise interfaces and future business models are also uncertain.
R&D progress is not as expected: Storage product technology continues to upgrade, if related companies cannot accurately grasp the market demand, continue to develop new products to maintain competitiveness, there is a risk of losing market share.
Intensified competition in the industry: In recent years, new manufacturers have continued to enter the storage industry, and if the number of new manufacturers continues to increase and the supply continues to be excessive, it will put some pressure on the product shipments and prices of related companies.
Macroeconomic downside risk: Affected by the overall macroeconomy, international geopolitical conflicts and the downward trend of the semiconductor cycle, the consumer electronics market has been greatly impacted, and the domestic and foreign market demand has shown varying degrees of weakness.