As always in semiconductors, cutting-edge manufacturing technology is central to delivering each new generation of more powerful chips. These chips are the integrated circuits that give intelligence to our favorite consumer devices, power a myriad of cloud-based services, and place the world’s data in the palms of our hands. In this era, the fabless/foundry business model has shown its strength and Taiwan Semiconductor Manufacturing Company (TSMC) has reinforced its role as the top foundry in the world.
For some economic context, the semiconductor market began a multi-quarter slump in the middle of 2022, with sales of the top five chip companies showing year-over-year declines starting in the third quarter. Operating margins for the top processor companies are less than one-third of what they were at the end of 2021. By nature of the business model and their separation from end-user demand, foundries, like TSMC, have been more isolated from this impact. In fact, TSMC only saw its first year-over-year revenue decline since 2019 in Q1-2023.
While the market is struggling somewhat, why is TSMC forging ahead with planned investments on expansion projects in Arizona and the European Union? One reason is the money already on the table.
Today, Yole Intelligence’s Senior Analyst, John Lorenz and his team invites you to get a better understanding of TSMC’s strategy in the field of processors industry. This snapshot is based on Yole Group’s product, Processor Market Monitor (Q2 2023 update). Aim of Yole Group’s Market Monitors is to provide an in-depth coverage of rapidly changing market dynamics and main players’ status and strategy. This monitor subscription provides quarterly updates on unit shipments, wafer production, and revenue on near-term and long-term basis. It provides capacity, CapEx, and supply chain insights into emerging markets and growth rates in mature markets in addition to providing package level forecast. The Processor Market Monitor is also linked to Yole Group’s computing reports such as: Computing and AI for Automotive 2023 – Computing and AI for Data Centers 2022…
Taking the Arizona project first, we see TSMC recently approving the addition of another $3.5 billion to the project where construction of the first phase is underway, aiming for mass production at the end of 2024. In total, the company has estimated the overall investment for the site to reach $40 billion, of which up to $15 billion will be covered through government subsidies.
Once completed, the first phase of the Arizona site will be a front-end manufacturing fab capable of 20,000 12-inch wafer starts per month running on TSMC’s N5 technology node and its derivatives. The site’s second phase, planned for completion in 2026, will be capable of an additional 30,000 wafer starts per month running on the 3nm class of technology nodes. To contextualize, the total monthly volume of 50,000 wafers would represent less than 5% of TSMC’s current wafer production capacity.
John Lorenz
Senior Analyst, Computing & AI at Yole Intelligence part of Yole Group
Certainly, TSMC could have expanded capacity in Taiwan, but in the current environment, foundry customers are looking for a supply-chain more resilient against global risks, and their governments are helping to facilitate. Therefore, US-based manufacturing is important for TSMC’s many US-based customers, such as Apple, Nvidia, and AMD, as the industry adjusts to a more regionalized posture.
The European project is also proceeding, though it is in earlier stages, with TSMC and partners still negotiating with the German government over the size of the subsidy. So far, the expected location is Dresden with an estimated investment of €11 billion. What is not yet known is how much of the €43 billion EU Chips Act the new fab will attract. However, this new manufacturing capacity is much anticipated by TSMC’s Europe-based customers, particularly those in the European automotive industry.
The next reason it makes sense for TSMC to move ahead happens to also be the biggest tech industry news of late: the explosion of interest in generative AI and the corresponding swelling of demand for acceleration hardware that runs the large-language-models (LLMs). The computational requirement for training the kinds of neural networks that drive the newest generative AI applications have been growing more than 100x per year. These transformer models, as they are called, can have over a hundred billion parameters requiring thousands of GPUs working for several weeks to train. Additionally, the sheer size of these models and their tokens have necessitated that the model inferences take place on GPUs as well. One might imagine, this has supercharged the demand for the AI accelerator of the moment, the data center GPU.
To illustrate the sudden impact of this trend, we can simply look at Nvidia’s data center revenues, which grew to $4.3 billion in the last quarter. Citing the demand for these accelerators, Nvidia has guided their next quarter revenue to reach $11 billion, which would match what the company earned for the whole year of 2019. While Nvidia is receiving much of the attention, it’s important to note that others in the acceleration hardware business stand to benefit as well, such as AMD, Intel, and a long list of smaller players who are bringing more customized solutions.
What does all of this have to do with TSMC? Well almost all the firms developing chips for accelerating AI training and inference count themselves among TSMC’s wafer fabrication customers. Taking the leading data-center GPUs right now, Nvidia’s H100 is built on TSMC N4 node and AMD’s MI300 is built on TSMC’s N5. In fact, both of these GPUs use TSMC’s advanced packaging technology CoWoS to make the critical interconnection between dies and to the High Bandwidth Memory (HBM). So, we see that TSMC sits in the critical path for this booming generative AI trend. Additionally, we expect wafer demand for all processors to grow at a 4.7% CAGR long term. Therefore, TSMC is in a good position and wise to add capacity.
The semiconductor industry has experienced some mixed sentiments lately. Consumer-oriented devices have seen a decline since mid-2022, but data center hardware, especially those that accelerate AI training and inference, are experiencing very strong demand more recently. This is the environment that TSMC sees itself continuing ahead with major capacity expansions in the US and the EU. Helping to keep these plans in place are the billions of dollars in government subsidies that TSMC and its peers will make use of. As always, these interesting times in the semiconductor industry cause us to pause and take a look at the wider picture, and for TSMC, the future is bright.
About the author
John Lorenz is a Senior Technology and Market Analyst, Computing & Software within the Semiconductor, Memory & Computing division at Yole Intelligence, part of Yole Group.
John is engaged in the development of market and technology monitors for the logic segment of advanced semiconductors, with a primary focus on processors. Prior to joining Yole, John held various engineering and strategic planning roles over 15 years at Micron Technology.
John has a Bachelor of Science degree in Mechanical Engineering from the University of Illinois Urbana-Champaign (USA), with a focus on MEMS devices.
John Lorenz Senior Analyst, Computing & AI at Yole Intelligence part of Yole GroupCertainly, TSMC could have expanded capacity in Taiwan, but in the current environment, foundry customers are looking for a supply-chain more resilient against global risks, and their governments are helping to facilitate. Therefore, US-based manufacturing is important for TSMC’s many US-based customers, such as Apple, Nvidia, and AMD, as the industry adjusts to a more regionalized posture.
