TSMC to double AI semiconductor capacity

Global semiconductor giant TSMC is accelerating the production of additional semiconductors for artificial intelligence (AI). With new factories coming on line, etc., it will increase the capacity of important processes to twice the current level by 2024. Against the backdrop of a sluggish global semiconductor market, TSMC will capitalize on the demand for generative AI, which has become the next growth point.

AI is changing people’s lives and work, bringing great opportunities to the semiconductor industry – TSMC Chairman Liu Deyin spoke passionately about the growth of the AI market in a speech at the international exhibition “SEMICON Taiwan” held in Taipei city on September 6.

Generative AI services such as OpenAI’s “ChatGPT” are mostly developed and utilized on the servers of IT giants in the U.S. AI semiconductors are mainly installed on these servers, and NVIDIA in the U.S. holds 80% of the global market share.

TSMC leads the way in mass production and exclusively produces NVIDIA’s high-performance products that are widely used in generative AI. TSMC’s strength in this area lies in a technology known as “advanced packaging,” which is categorized as a “post-process” of semiconductor production.

Semiconductor production is divided into the “pre-process” to form circuits on silicon wafers and the “post-process” to cut the silicon wafers that form circuits into semiconductor chips, assembly and testing.

Semiconductors have previously improved performance by pursuing the microrefinement of circuits in previous processes. On the other hand, cutting-edge packaging, which holds the key to improving the performance of AI semiconductors, is a post-process technology that packages multiple semiconductors with functions such as computing and storage together and enables them to interact efficiently.

In order to increase production of AI semiconductors, TSMC has taken a series of actions to enhance the production capacity of the post-process.

First, a new factory was built in Miaoli County, central Taiwan, in June. TSMC has concentrated all five of its cutting-edge packaging plants, including this one, in Taiwan. In addition to the launch of new plants, it is also promoting the enhancement of existing plants, and plans to increase the overall capacity of TSMC by 2024, which is mainly used by Nvidia and other “CoWoS” mode, to about two times by 2023.

Increase in market size to eight times

Against the backdrop of rapidly expanding demand for AI semiconductors, the lead time for NVIDIA’s mainstay products is thought to be up to 1 year, with an apparent shortage of supply. TSMC’s Liu Deyin told reporters after his speech at SEMICON Taiwan on the 6th that supply will keep up with demand after 1.5 years, showing confidence in measures to increase production.

TSMC is also taking action with an eye toward medium- to long-term market expansion, announcing in July that it is investing about NT$90 billion in a new factory at another site in Miaoli County. Construction is expected to begin in the second half of 2024, with mass production starting in 2027.

The market size for AI semiconductors will increase to more than eight times its 2022 level by 2030, reaching $165 billion (about 24 trillion yen), according to German research firm Statista. This is equivalent to 30% of the overall global semiconductor market (about 80 trillion yen) in 2022, and will be the next engine of the semiconductor market.

Other peer companies are expected to follow in TSMC’s footsteps and enter the AI market.

NVIDIA said at its earnings presentation in late August that in order to increase production, it “will approve new suppliers and gradually increase supply through 2024.” To cope with the further expansion of demand, NVIDIA is going to add foundries other than TSMC.

An executive of a post-process material manufacturer believes that “TSMC has a lot of production experience, but the price is higher. Nvidia and others may focus on cost performance and explore the diversification of foundry parties.”

Other foundries surface

Statistics from French research firm Yole Intelligence show that the top six companies hold more than 80% of the market share in the cutting-edge packaging market, including those oriented toward directions other than AI. In addition to the world’s top three semiconductor companies, TSMC, U.S. Intel, and South Korea’s Samsung Electronics, they also include three Chinese, U.S., and Taiwanese companies that specialize in post-processing.

Among them, is considered the most likely to become NVIDIA’s new foundry is specialized in post-process Taiwan Sunrise Investment Holdings (ASE). Previously, Sun Micron has been focusing on smartphone and PC semiconductors as its mainstay, but plans to establish a production system for special processes used in AI semiconductors by early 2024 to accelerate the capture of AI-related demand.

In addition to Taiwanese companies, U.S.-based Intel has also shown an active stance. The company announced in June that it would invest up to $4.6 billion in Poland to build a new post-process factory. In addition to facing itself, it also plans to take on external foundry orders, aiming for production in 2027.

TSMC and other large semiconductor companies have previously competed fiercely through the microfabrication of the front process. In the future, as the AI semiconductor market expands, the importance of the back-end process increases, and it will become an important focus of competition that will shape the industry.

The semiconductor market has been in a slump centered on smartphones and personal computers, and there are many opinions that a full recovery will not take place until after next spring. Nevertheless, new competition is intensifying around AI semiconductors, which are the next growth driver.