The global semiconductor market is expected to bottom out in 2023 and begin to fully recover in 2024.
From 2021 to 2022, the global semiconductor market grew rapidly due to the emergence of special needs due to the outbreak of the new crown epidemic. However, with the extraordinary demand caused by the pandemic ending in the second half of 2022, the semiconductor industry fell into one of the worst recessions in history in 2023. However, the Great Recession is expected to bottom out in 2023, with a full recovery coming this year (2024).
The author believes that whether the semiconductor market can fully recover depends on the trend of memory.
In fact, looking at the quarterly semiconductor shipments of various types, logic chips have surpassed the special demand peak caused by the new crown epidemic and hit a new all-time high. In addition, Mos Micro (microprocessors) and Analog (analog devices) are likely to hit record highs in 2024, as the decline due to the end of COVID-specific demand is modest (Figure 1).
Figure 1: Quarterly semiconductor shipments by type and volume (as of Q4 2023) Source: Created by authors based on WSTS data
Among them, Mos Memory bottomed out in the first quarter of 2023 (Q1) and is recovering after a sharp decline, but it still seems to take quite some time to reach the peak of post-pandemic demand. However, if Mos Memory exceeds this peak, total semiconductor shipments will certainly hit a new all-time high. In this way, it can be said that the semiconductor market has fully recovered.
However, judging by the changes in semiconductor shipments, it is clear that this perception is wrong. This is because, while shipments of Mos Memory have largely recovered, the value of logic chips, which have reached a record high, is still at a very low level. In other words, for the global semiconductor market to fully recover, the number of logic chip shipments must increase significantly.
Therefore, in this article, we will analyze the shipment and volume of semiconductors for each type of semiconductor and the total volume of semiconductors. Next, the author takes the difference between the shipment volume and the shipment volume of logic chips as an example to show how TSMC's shipments are sluggish while the shipments of TSMC are recovering rapidly.
Further speculated as to why this discrepancy occurred, and discussed the possibility that the full recovery of the global semiconductor market would likely be postponed until 2025.
The conclusion is that now the semiconductor market seems to be recovering, which is due to the "illusion" caused by the fact that Nvidia GPUs are trading at very high prices. Therefore, it seems that the semiconductor market will not fully recover until foundries such as TSMC reach full capacity and logic chip shipments hit another record high.
Analysis of semiconductor shipments and shipments
Figure 2 shows trends in shipments and volumes of various types of semiconductors and semiconductors as a whole.
After peaking in Q4 2021, Mos Micro's shipments bottomed out in Q1 2023 and are recovering. On the other hand, the number of its shipments has not changed significantly, and it is roughly the same, with a slight decrease from Q3 to Q4 in 2023.
Figure 2 Trends in the overall shipment value and volume of semiconductors by type of semiconductor and semiconductor (as of Q4 2023) Source: Created by the author based on WSTS data
Mos Memory shipments fell sharply from Q2 2022 to bottom out in Q1 2023 and turned upward, but only recovered to about 40% of their peak in Q4 of the same year. On the other hand, the number of shipments recovered to 94% of its peak. That said, it can be assumed that the memory manufacturer's factory run rate is close to full capacity. The question is how much the price of DRAM and NAND flash memory will go up.
After peaking in Q2 2022, shipments of logic chips bottomed out in Q1 2023 and began to rise, and set a new all-time high in Q4 of the same year. On the other hand, after reaching a peak in Q2 2022, the number of shipments dropped to about 65% of the peak in Q3 2023, and remained flat in Q4 of the same year. In other words, in a logic chip, there is a big difference between the behavior of the shipping amount and the number of units shipped.
After peaking in Q3 2022, analog shipments bottomed out in Q2 2023 and then flattened. On the other hand, the number of its shipments continued to decline after reaching a peak in Q3 2022 until Q4 2023.
Finally, after a sharp decline in Q2 2022, overall semiconductor shipments bottomed out in Q1 2023 and turned upward, and recovered to about 96% of the peak in Q4 of the same year. On the other hand, its shipment volume has decreased significantly from Q2 2022 and bottomed out in Q1 2023, but has since continued to be flat at about 75% at its peak.
Looking at the above, if you only look at the volume, Mos Memory is the problem, which is currently only recovering to about 40% of the peak. However, if you look at it more broadly, you can see that logic chips are a major problem, and although shipments have reached an all-time high, shipments have stagnated at around 65% of their peak. The impact of this discrepancy between the shipments and shipments of logic chips seems to extend to the semiconductor sector as a whole.
In short, the recovery of the global semiconductor market depends on the price increase of Mos Memory and whether the number of logic chip shipments will increase significantly. And, since DRAM and NAND prices are on an upward trend, the biggest problem is the increase in the number of logic chip shipments.
Below, we will explain the behavior of TSMC's shipment volume and number of wafers shipped to concretely demonstrate the difference between the number of logic chips shipped and the number of wafers shipped.
TSMC's quarterly shipments and wafer shipments
Figure 3 shows TSMC's sales share by node and sales trends for 7nm and above processes in Q4 2023.
TSMC positions 7nm and beyond as an advanced node, and in Q4 of 2023, 7nm will account for 17%, 5nm will account for 35%, and 3nm will account for 15%, bringing the total number of advanced nodes to 67%. In addition, quarterly sales of the advanced node will start growing in the first quarter of 2021 and decline once from the fourth quarter of 2022, but bottoming out in the second quarter of 2023 and starting to rise again, reaching an all-time high in the fourth quarter of the same year.
Figure 3 TSMC's 2023Q4 sales proportion of each node and the sales trend of advanced nodes Source: TSMC's 2023Q4 financial report slides
In other words, if you look at the sales of advanced nodes, TSMC is doing well. So, what about TSMC's overall quarterly sales and wafer shipments (Figure 4)?
Figure 4: TSMC's quarterly shipment value and number of wafers shipped (as of Q4 2023) Source: Created by the author based on TSMC's historical operating data
TSMC's quarterly shipments and chip shipments charts are roughly the same. The IT bubble peaked in 2000, fell in 2008 after the impact of the financial crisis, and also fell after the memory bubble burst in 2018.
However, the behavior was different after the peak of pandemic-specific demand in the third quarter of 2022. Shipments peaked at $20.2 billion and then fell sharply, but began to recover after bottoming out at $15.7 billion in the second quarter of 2023 and recovered to $19.7 billion in the fourth quarter of the same year, 97% of the peak.
On the other hand, chip shipments in this quarter fell sharply after reaching a peak of 3.97 million units in Q3 of 2022, reaching a peak of 2.92 million units in Q2 of 2023, but remained flat after that, and also remained at 2.96 million units in Q4 of the same year, a decrease of about 1 million units from the peak.
Among the semiconductors produced by TSMC, logic chips are the most numerous. TSMC's sales in the fourth quarter of 2023 hit an all-time high for front-end nodes, recovering to an all-time high of 97% overall. Despite this, chip shipments in the quarter were more than 1 million fewer than at the peak. In other words, TSMC's overall factory operation rate is only about 75%.
In terms of global semiconductors as a whole, logic chip shipments have slipped to about 65% of the peak of special demand for the new crown. In line with this, TSMC's quarterly wafer shipments are more than 1 million fewer than at its peak, and the factory run rate is considered to be sluggish at around 75%.
In the future, for the global semiconductor market to truly recover, logic chip shipments will need to increase significantly, and to do so, the operating rate of foundries led by TSMC must be close to full capacity.
So when exactly?
Forecast major foundry utilization
On December 14, 2023, TrendForce, a Taiwanese research company, held a seminar titled "Industry Focus Information" at the Ariake Washington Hotel in Tokyo Bay. At the seminar, Joanna Chiao, an analyst at TrendForce, introduced "TSMC's global strategy and 2024 semiconductor foundry market outlook". Among them, Joanna Chiao talks about forecasting foundry utilization (Figure 5).
Figure 5: Changes in 12-inch fab utilization rates at major foundries Source: Joanna Chiao (TrendForce), "TSMC Global Strategy and 2024 Semiconductor Foundry Market Outlook"
According to this forecast, by the second quarter of 2022, all foundries will be operating at more than 90%. However, during Q4 of the same year, the operating rate of many foundries dropped sharply.
Among them, only TSMC is firmly at the top of the list with an operating rate of 99%, but it will drop to 76% in Q4 2023. TSMC's operating rate is roughly in line with the operating rate of about 75%, which is derived from the number of chips shipped in the quarter.
In addition, it is expected that the operating rate of major foundries, including TSMC, will not reach full capacity even in Q4 2024. Even TSMC, which has the highest efficiency, is only 85%. Samsung Electronics ("Samsung"), the world's second-largest foundry in terms of sales, expects an operating rate as low as 67%.
That said, according to this forecast, the operating rate of the major foundries will not reach full capacity until the end of 2024. As a result, it is not expected that the number of logic chip shipments will increase significantly by 2024, and as a result, the global semiconductor market will not really recover.
So why are the major foundries, led by TSMC, operating so low?
Year-on-year growth in shipments of various types of electronic equipment
Figure 6 shows the growth rate of the number of electronic devices shipped in 2022, 2023, and 2024 in the previous year. Among them, electronic devices that have a major impact on the semiconductor market, including logic chips, are smartphones, servers, and PCs (notebooks). However, the growth rate of shipments of these three devices is not significant.
Figure 6: Year-over-year growth rate of shipments of various electronic devices (2022-2024) Source: Joanna Chiao (TrendForce), "TSMC Global Strategy and 2024 Semiconductor Foundry Market Outlook"
The growth rate of smartphones decreased by 10.6% in 2022, continued to decline to 2.7% in 2023, and turned positive in 2024, but only by 3.0%. Server sales increased by 4.7% in 2022, decreased by 6.0% in 2023, and returned to positive growth in 2024, but the growth rate was only 2.5%. Laptops will drop significantly to 24.5% in 2022 and remain at -10.2% in 2023 before eventually recovering to +3.2% in 2024.
In addition to these three types, the growth rate in 2024 is also disappointing, with tablets down 0.1%, TVs down 0.3%, and monitors down 1.5%.
In this context, the high growth rate of AI servers and electric vehicles (EVs) is remarkable. Specifically, AI servers were zero in 2022, but grew by 36.9% in 2023 and 44.2% in 2024. This AI server uses AI semiconductors. Moreover, Nvidia's GPUs account for more than 80% of this AI semiconductor.
The reason why the shipping amount of the logic chip does not match the shipping quantity
Among NVIDIA's GPUs, the H100 is currently the highest-performing GPU, which is in high demand and is sold at an extremely high price of 5 million to 6 million yen per unit. This is believed to be the reason for the record high in logic chip shipments.
However, as Figure 6 shows, with the exception of AI servers and electric vehicles, many electronic devices are growing at an extremely slow rate, so shipments of logic chips used in these devices will be at a low level. As a result, the operating rate of the major foundries that produce logic chips will also be sluggish.
So, how much will Nvidia's GPU and other AI semiconductors be shipped in the future? How much is it expected that global logic chip shipments will increase? In other words, will AI semiconductors become the savior of market recovery?
Changes in AI server shipments
Figure 7 shows the number of AI server shipments from 2022~2026. According to the forecast, AI server shipments in 2022 will be 877,000 units, which will increase by 36.9% to more than 1.2 million units in 2023 and further increase by 44.2% to more than 1.7 million units in 2024. Approximately 2.7 million units are expected to be shipped by 2026.
Figure 7: AI Server Shipment Trends Source: Ken Kuo (TrendForce) Data, "Global Memory Market Analysis Predicts the Future of AI Next Year"
So, what is the percentage of all servers in this AI server (Figure 8)?Based on this number, it is expected that AI servers will account for 6% of the total in 2022, 9% in 2023, 13% in 2024, 14% in 2025, and 16% in 2026.
Figure 8: Number of servers shipped, proportion of AI servers, proportion of AI chip wafers Source: Joanna Chiao (TrendForce), "TSMC's Global Strategy and 2024 Semiconductor Foundry Market Outlook"
Generative AI such as ChatGPT is exploding in popularity around the world. As a result, the demand for AI servers seems to be very high. However, in reality, the shipment of AI servers is not as large as imagined.
AI server bottlenecks
It is believed that the reason for this is the speed of supply of AI semiconductors. Currently, Nvidia owns about 80% of the GPUs of AI semiconductors in TSMC for pre- and back-end. In the subsequent process, a package called CoWoS is carried out, but the capacity of this CoWoS becomes a bottleneck.
In addition, in CoWoS, HBM (High BandwidthMemory) is configured around the GPU with multiple layers of DRAM, but this HBM is also considered to be one of the bottlenecks. Because it is mainly SK hynix that supplies HBM to NVIDIA, while Samsung, the king of DRAM, is lagging behind.
In this way, it is difficult to say that HBM is also fully supplied, in addition to the insufficient capacity of CoWoS. In addition, the lead time for CoWoS is about half a year, after which it will take another half a year to assemble the AI server. In short, it takes about a year to ship an AI server.
In addition, in the figure on the right side of Figure 8, the percentage of chips for AI semiconductors in the capacity of the world's advanced processes (possibly 7nm or later) is written. AI semiconductor chips, which were only 4% in 2022, are expected to double by 8% by 2026.
When will logic chip shipments increase?
Is this 8% big or small? Although this is a delicate question, even by 2026, the remaining 92% of wafers will be consumed by chips that are not AI semiconductors. Most of them are logic chips. Therefore, in order to increase the shipment of logic chips and the operating capacity of major foundries led by TSMC, the demand for electronic devices such as smartphones, PCs, and servers must increase.
All in all, AI semiconductors represented by Nvidia GPUs will not be a savior. As a result, it will be difficult for the global semiconductor market to truly recover in 2024 and is expected to be postponed until next year 2025.
However, there is another (positive) possibility to reverse this forecast.
The AI semiconductors that have been explained so far mean that they are all mounted on servers. However, the movement of AI processing on terminals (edges) such as PCs, smartphones, and tablets is now emerging.
For example, the AI PC advocated by Intel and the AI smartphone that Samsung is about to make, etc. If they explode in popularity (i.e., innovation), the AI semiconductor market will expand rapidly. In fact, Gartner, an American research firm, estimates that 240 million AI smartphones and 54.5 million AI PCs will be shipped by the end of 2024. If this prediction is realized, the demand for cutting-edge logic chips (shipments and shipments) will also increase, and the utilization rate of wafer foundries such as TSMC will also increase. In addition, the demand for MPUs and memory is bound to expand dramatically.
In other words, if such a world comes, AI semiconductors will be the real savior. Therefore, in the future, I will focus on the trend of edge AI semiconductors.
*Disclaimer: This article was created by the original author. The content of the article is his personal point of view, and our reprint is only for sharing and discussion, and does not mean that we agree or agree, if you have any objections, please contact the background.