Source: Wall Street News

Driven by emerging technologies such as AI and high-performance computing, the importance of advanced manufacturing processes in the foundry industry is becoming increasingly prominent.$Taiwan Semiconductor (TSM.US)$It still maintains its absolute advantage in the field of foundry and is firmly in the top position in the industry.

On Monday, March 4, driven by strong demand for AI chips, TSMC's Taiwanese stock once rose more than 5% to NT$725, and its market capitalization rose to NT$18.8 trillion, a record high. On Friday, TSMC closed up 4.06% to $133.9 in the US stock market, approaching the highest level in history.

As the supply of AI chips continues to be in short supply, Goldman Sachs believes that the most direct way to solve the shortage of AI chips in the short term is to increase CowOS production capacity, and TSMC is the core of this.

The team led by Goldman Sachs analyst Bruce Lu pointed out in a recently released report that the production capacity of CoWoS (Chip on Wafer on Substrate) packaging technology has become the biggest bottleneck limiting the supply of AI chips in the past year, and is also the key to whether AI chip demand can be met in 2024.

On January 18, when discussing advanced packaging issues at the TSMC Law Conference, TSMC CEO Wei Zhejia pointed out that demand for advanced packaging for AI chips continues to be strong. Currently, production capacity is still unable to meet strong customer demand, and the shortage of supply may continue until 2025. TSMC continues to expand its advanced packaging production capacity this year. The advanced packaging production capacity is planned to double this year, and supply is still in short supply. It is estimated that production capacity will continue to expand in 2025.

Goldman Sachs believes that TSMC's foundry model is still the most efficient way to increase AI chip production capacity. Given the advantages of foundry companies in cost optimization and operational efficiency, it is a more realistic solution compared to building their own factories that specifically meet the needs of AI chips.

Goldman Sachs said bluntly that the current shortage of AI chips is mainly$NVIDIA (NVDA.US)$The shortage of H100 chips is manufactured using TSMC's 4nm node and requires the use of CoWOS packaging. The production capacity of advanced processes is not the key to the chip supply problem. CoWOS production capacity restrictions are an important reason for insufficient chip supply. Based on estimates of TSMC's CoWoS expansion plan, production capacity may increase significantly in the third quarter of 2024:

The current demand for H100 chips actually accounts for only about 10% of TSMC's 5nm node production capacity. If TSMC uses 100% of its N5 production capacity to produce H100 chips, H100 production capacity can immediately be increased tenfold. Current advanced manufacturing wafer production capacity is not the key to recent supply issues.

In our opinion, the current bottlenecks are mainly due to CoWoS capacity limitations. The current CoWOS production capacity (16kwpm) is far less than the N5 wafer production capacity (125kwpm). As AI demand begins to emerge in 2023, CowOS equipment vendors will need time to increase their production capacity to meet surging AI demand, while TSMC only requires a capital investment of US$1.5-180 million to double its current CowOS production capacity.

Increasing CowOS production capacity is a direct solution to the shortage of AI chips in the short term

Goldman Sachs believes that demand for AI chips will change dynamically as costs drop. Generally speaking, there is usually an inverse relationship between demand and cost. When costs rise, demand will tend to decline, and when demand and cost rise simultaneously, the industry will seek to reduce costs to meet growing demand, and the AI industry is no exception:

Currently, in a situation where costs remain high, the demand for AI is also rising rapidly. The cost of AI servers is more than 7-30 times higher than that of ordinary servers. Therefore, when considering building multiple fabs to meet strong AI demand, we believe it is necessary to carefully study the dynamics of actual demand.

Undoubtedly, the current demand for AI chips is closely related to the current cost structure. Key factors include: 1) the cost of 1 TOPS (1 trillion operations per second) — that is, the computational performance of AI chips, and 2) the cost of the chip — which is related to the selling price. Since AI development is still in its early stages, the cost reduction curve may drop significantly in the next few years with technology migration, chip design and performance improvements, and the cost of chips may be drastically reduced in the future.

Furthermore, the advent of neuromorphic chips for edge AI (based on pulsed neural networks SNN) will also increase the complexity of evaluating future chip requirements and related cost issues. For us, it is important to identify future and long-term “real AI chip requirements” by understanding changes in AI-related capacity and costs.

Goldman Sachs determined that currently, for the production of AI chips, the production capacity of advanced process processes (such as TSMC's 5nm node) is not a limiting factor. Currently, the main limiting factor is the production capacity of CoWoS (chips stacked and packaged on a substrate):

CoWOS refers to stacking chips and encapsulating them on a substrate. The space required for the chip can be reduced, and the benefits of reducing power consumption and cost can also be achieved. Among them, it can also be divided into a 2.5D horizontal stack (the most well-known is TSMC's CoWoS) and a 3D vertical stack version to stack various chip modules such as different processors and memories into small chips (Chiplets). Because its main application is in advanced manufacturing processes, it is also known as advanced packaging.

In our opinion, the current bottleneck is mainly due to limited CoWoS production capacity. The current CoWOS production capacity (16kwpm) is far less than the N5 wafer production capacity (125kwpm). As demand for AI chips has surged since 2023, CoWoS suppliers are gradually expanding production capacity to meet demand.

Therefore, we believe that from the perspective of industry participants, investing in CowOS will be a key point in solving the shortage of AI chips at this stage. Based on our current estimate of TSMC's CoOS production capacity expansion, assuming that TSMC's entire CoWOS production capacity is used for H100 chip manufacturing, it is estimated that in 2023, 2024, and 2025, TSMC will be able to produce 3.6 million, 7.6 million, and 11 million H100 chips in a year, respectively.

If 100% of TSMC's N5 process production capacity is used to produce H100 chips, then without considering other changes, it is estimated that in 2023, 2024, and 2025, the annual output of H100 chips produced by TSMC will be 36 million, 37.5 million, and 43.1 million pieces, respectively.

Goldman Sachs concluded that increasing CoOS production capacity is the primary solution to meet the surge in demand for AI chips. Increasing CowOS production capacity is the fastest and most direct way to increase AI chip production at this stage. Compared with the expansion of production capacity of advanced manufacturing nodes (that is, 5nm/3nm nodes), the capital intensity of CowOS is much smaller:

Increasing CowOS production capacity is a top priority to meet surging AI demand. According to TSMC's outlook, the company plans to double its CoOS production capacity in 2024, and capital expenditure of approximately $3 billion (10% of total capital expenditure guidance of $3 billion) will be used for advanced packaging/masking/other expansion.

Assuming that 50-60% of the $3 billion capital expenditure is for CoWOS expansion (increase of about 15 kwpm), the capital intensity required for each additional 10k wafer is only about $100-120 million in capital expenditure, which is far less than the capital intensity of 5nm/3nm wafers, which would require 250 million/350 million US dollars for each additional 10k wafer production capacity.

This means that TSMC only needs $1.5-18 billion in capital investment to double its current CoWOS production capacity (from about 15-16 kwpm at the end of 2023 to about 30 kwpm by the end of 2024); $150-18 billion in capital expenditure will increase production capacity to 150 kwpm (about 10 times the current level).

The OEM business model is still the best choice to meet the needs of AI chips

Goldman Sachs believes that the current industry wants to optimize resource allocation to meet the long-term demand for artificial intelligence. In addition to increasing CoWOS production capacity to solve the short-term surge in AI chip demand, from a long-term perspective, the industry may also need more wafer production capacity. In the future, it will face two key choices: 1) establish its own factory specifically to meet AI production capacity requirements; 2) cooperate with existing foundry companies (such as TSMC, Intel, etc.) to expand production capacity. The OEM business model is still the most effective way to expand production capacity:

Looking at the first solution, we believe that if a new fab is built entirely dedicated to AI requirements, the cost of chips may rise further. Considering the industry's precedent of node migration every 2-3 years, depreciation costs are a key element in chip production costs. Depreciation costs typically account for 50-60% of TSMC's chip manufacturing cost structure and are a key profit factor.

Considering the node migration cycle, in order to pursue higher chip performance, power consumption, and higher transistor density, this means that production capacity built 2-3 years ago for AI may need to be phased out. Therefore, if the capital expenditure depreciation cycle is shortened from the current 5-6 years to 2-3 years, it is expected that the fab will face a large cost burden, and the depreciation burden will increase every year.

Furthermore, due to continuous chip migration, continuous reinvestment is required, and production capacity that has already been built may need to be phased out or underused every 2-3 years. Therefore, building a new fab just for AI needs may further increase the overall cost of the chip. Furthermore, in order to determine what kind of node production capacity needs to be built in different years, it is necessary to consider potentially disruptive solutions to reduce costs (e.g. neuromorphic chips) and the evolution of demand for transistors from AI chips, all of which will add more complexity.

Goldman Sachs pointed out that the OEM business model is still the most efficient way to expand production capacity. Considering the advantages of foundries in terms of cost optimization and operational efficiency, as well as the background of rapid development in the AI field, the second option — the foundry business model — is a more realistic solution:

By adopting an OEM business model, companies can avoid the high costs associated with building, owning, and operating a fab. Foundry serves multiple customers, which allows the company to share with other foundry customers and reduce the overall cost burden, such as fixed costs (i.e. depreciation, manufacturing management expenses).

Additionally, the foundry business model also provides greater flexibility and scalability. Companies can expand their production needs by partnering with foundries. These foundries manage capacity planning to keep their plants running at higher capacity. Furthermore, considering technology migration every 2-3 years, the production capacity established in the past few years faces less risk of lower capacity utilization when the company migrates to use more advanced processing technology, because the old production capacity will be filled by other needs of other customers. Overall, we believe that the OEM business model is still the most efficient way to meet the surging demand for AI chips.

Editor/jayden