Japan’s Chip Challenge: Semiconductor Policy for the Data Centre Era

This publication is licensed under the terms of the Creative Commons Attribution License 4.0 which permits unrestricted use, provided the original authors and source are credited.

Introduction

Japan used to be the world’s leading semiconductor manufacturer. In the late 1980s, Japan accounted for over 50% of the global semiconductor manufacturing industry, and was home to the world’s three largest semiconductor firms – NEC, Toshiba and Hitachi. Today, however, Japan accounts for about 10% of semiconductor manufacturing and has no firms in the global top ten. 

The erosion of its semiconductor manufacturing industry has become a source of concern for the Japanese Government, given the economic and geopolitical importance of semiconductors. To address this decline, the Ministry of Economy, Trade and Industry (METI) published the Strategy for Semiconductors and the Digital Industry in 2021, updating it in 2023. METI’s strategy aims to bring leading-edge semiconductor manufacturing back to Japan and, in parallel, maintain Japan’s existing strengths in the equipment and materials stages of the semiconductor supply chain. 

The success of Japan’s semiconductor policy will be determined in part by its firms’ ability to adapt to a major new trend in the global semiconductor industry – the growing demand for data centre chips. This article analyses the opportunities and challenges that this trend poses to Japan’s semiconductor industry and concludes with implications for policymakers. 

The Data Centre Era

Trends in semiconductor end-use markets – such as cars, computers and mobile devices – have a major impact on the size and shape of the semiconductor industry. Japan learned this the hard way in the 1990s, when its firms struggled to adapt to the growth of the personal computer (PC) end-use market. The PC era led to a shift in the dominant type of semiconductor – from DRAM chips to logic chips – and the dominant business model – from integrated device manufacturers to fabless-foundry partnerships. In this way, the PC revolution created new winners and losers in the semiconductor industry; Japanese manufacturers, ill-positioned to react to these shifts, declined in market share. Similarly, the smartphone revolution in the 2010s benefited some countries and semiconductor firms, while others lost out. 

Today, data centres – the infrastructure that underpins artificial intelligence (AI) and performs large computational tasks – are beginning to cause profound changes in the semiconductor industry. Advances in AI technology are driving investments into data centres, with the value of the global data centre market projected to grow from $220 billion in 2022 to $418 billion in 2030. Semiconductors are the critical input for any data centre, and revenue from data centre semiconductors is projected to grow 12% per year from 2022 to 2030. 

Data centres rely on a wide variety of semiconductor inputs, but this article focuses on central processing units (CPUs), graphics processing units (GPUs) and accelerators, which together perform most data centre operations. Although these types of semiconductors are also found in some consumer electronics, the specific varieties required for data centres are larger, more expensive, and often more challenging to manufacture. 

Japan is well positioned to capitalise on rapid growth in data centre semiconductors. The Japanese Government and industry have demonstrated their willingness to provide valuable support, fiscal and otherwise, to revitalise the manufacturing ecosystem. The Taiwan Semiconductor Manufacturing Company (TSMC) facility in Kumamoto, which opened in February 2024, is testament to this. Additionally, the presence of major data centre operators in Japan provide opportunities for commerce and cooperation: two Japanese companies – NTT and KDDI Telehouse – are among the world’s ten largest data centre providers, and large foreign data centre operators like Amazon Web Services (AWS), Google Cloud and Microsoft Azure are expanding investments in Japan. Taken together, these factors mean that Japan’s semiconductor industry has the potential to become a major winner of the data centre revolution, but to achieve this it must adapt to significant challenges. 

Challenges

Rapid growth in demand for data centre semiconductors will lead to four main challenges for the global semiconductor industry, with specific implications for Japanese firms. 

1. Barriers to industry coordination

Data centres require highly advanced semiconductors, capable of processing computationally intensive algorithms for next-generation generative AI systems. Historically, advances have occurred at the fabrication stage of the supply chain, by shrinking the size of transistors – the tiny electric ‘switches’ that allow current to flow – and then packing more transistors onto a semiconductor. This is the essence of Moore’s Law, which states that the number of transistors on a semiconductor should double every two years. Recently, however, Moore’s Law has slowed, as fundamental physical limitations make it increasingly challenging to continue shrinking tiny, nanoscopic transistors. Therefore, continued semiconductor advances must come from elsewhere in the supply chain. 

One solution is to tailor semiconductor designs to specific computational workloads. This demands greater industry coordination, as manufacturers need to work closely with designers and their customers to ensure that semiconductors are optimised for their end-use application, specifically the kinds of software that will run on them. These custom semiconductors rely on distinct subcomponents that must be integrated through new advanced packaging technologies, such as three-dimensional stacking. Advanced packaging requires close collaboration throughout the supply chain, amongst designers, front- and back-end manufacturers, equipment providers and materials suppliers.

Despite the strong need for enhanced industry coordination, it faces at least three barriers. First, there is a lack of standards and procedural infrastructure for heterogeneous chip design and advanced packaging. Second, industry-leading companies are not incentivised to address these standardisation challenges, as this would promote competition and new market entrants. Third, the siloed structure of many firms makes it difficult for engineers to address cross-functional challenges. For Japan to capitalise on increased demand for data centre semiconductors, its semiconductor industry must learn to work better together. 

2. Rising R&D costs and capital expenditure

The pursuit of leading-edge data centre semiconductors contributes to rising R&D costs across the supply chain. The greater sophistication in semiconductor design and greater complexity in manufacturing and packaging are hugely expensive. The design costs of an advanced AI accelerator can reach $540 million. At the fabrication stage, researchers are developing new technologies – such as gate-all-around (GAA) transistors – to overcome physical limits to shrinking transistors, and breakthroughs in materials and equipment are also required to support advanced fabrication and packaging techniques. This could be a source of concern for Japanese firms, which typically spend less on R&D (as a proportion of operating revenue) than their international counterparts (Figure 1). In the highly innovative semiconductor industry, there is a risk that some Japanese firms could be left behind: lower R&D expenditure is often linked with lower profitability, which, in turn, can constrain further re-investment in R&D. 

Figure 1: Semiconductor Firms’ R&D Expenditure as a Proportion of Operating Revenue (Authors’ analysis, based on their aggregation of financial data from a selection of 235 current and former publicly listed firms that have significant lines of business in the semiconductor value chain).

The surge in demand for data centre semiconductors is also causing a sharp increase in capital expenditure, particularly for leading-edge manufacturing. Today, the construction costs for a leading-edge 3-nanometer (nm) fab – a semiconductor manufacturing plant – are approximately $20 billion; a next-generation 2nm fab is expected to cost $28 billion. Capital expenditure amongst Japanese front-end manufacturers is well below international firms (Figure 2). This will need to be reversed if Japan is to re-establish its manufacturing competitiveness. 

Figure 2: Semiconductor Firms’ Capital Expenditure as a Proportion of Operating Revenue (Authors’ analysis).

3. Manufacturing capacity constraints

Given the high costs and technological complexity, only three companies – TSMC, Samsung and Intel – are capable of manufacturing leading-edge data centre chips at commercial scale. Unsurprisingly, demand for data centre semiconductors exceeds supply, leading to shortages and long wait times. Japan has an opportunity to service some of this unmet demand. It is doing this, in part, by subsidising major international semiconductor manufacturers to relocate to Japan, illustrated by TSMC’s new fab in Kumamoto. Beyond this, however, the Japanese government is focusing on Rapidus, a joint venture between eight Japanese companies and IBM that aims to mass produce next-generation 2nm semiconductors for data centres by 2027. 

Rapidus faces steep technological challenges. Prior to recent foreign investments, Japan’s most advanced fabs could only manufacture semiconductors at the 40nm node, a technology that is more than 15 years old. Rapidus aims to catapult Japan’s semiconductor manufacturing from 40nm directly to 2nm, skipping at least eight generations of node technology in a very short time period in what would amount to an “unparalleled technological feat.” Nevertheless, Rapidus could become technologically competitive due to extensive R&D support from IBM – the first company in the world to develop the 2nm technology in 2021 – and imec – a leading R&D centre for nanoelectronics.

The biggest challenge facing Rapidus is its commercial viability, specifically its ability to attract and retain customers. Many potential customers are hesitant to redesign their semiconductors to be compatible with different foundry processes. Rapidus is an unproven fab, and uncertainty about the yield of its manufacturing process increases the already high costs of switching manufacturing partners, dissuading prospective customers. Additionally, Rapidus’s atypical business model means that it will manufacture ‘hot lots’, small volumes of specialty semiconductors with fast cycle times. A lack of customers and small production volumes will challenge Rapidus’s ability to benefit from economies of scale and learning by doing, two factors that are essential for a fab’s commercial success. 

4. Worsening talent shortages

Increased demand for data centre semiconductors leads to increased demand for a workforce capable of designing and manufacturing them. This requires new types of skills, including familiarity with new materials (for example, gallium nitride), new advanced packaging technologies, and specialised AI applications. Moreover, the semiconductor workforce as a whole needs to grow, as some forecasts estimate a shortfall of 400,000 semiconductor workers worldwide by 2030. Domestically, the Japan Electronics and Information Technology Industries Association projects a talent gap of 40,000 semiconductor engineers.

The shortage of semiconductor skills is partly a global problem, but is exacerbated in Japan by domestic factors. As Japan’s semiconductor manufacturing industry was eroded from the 1990s onwards, much of its talent left for Taiwan, South Korea and China. Japan’s remaining semiconductor workforce is ageing and will be badly affected by retirements over the next 15 years. Decades of declining birth rates have caused a decline in the number of graduates, reducing the pipeline of new semiconductor workers. Moreover, Japan’s performance in STEM subjects (science, technology, engineering and mathematics) is starting to lag behind competitor countries, with only 35% of Japanese students graduating with a STEM degree, compared with 38% in the United States, and 42% in South Korea and Germany. Hideki Wakabayashi, professor at the Tokyo University of Science, summarises the problem: “it is often said that semiconductors are lacking, but the bigger shortage is engineers.”

Policy Implications

The four main challenges facing the semiconductor industry in the data centre era require a response from policymakers. The five policy implications set out below contain both general principles and specific suggestions for the Japanese context. 

First, government has an important role to play in facilitating industry coordination and reducing information asymmetry between different stages of the supply chain. Critically, government coordination efforts should complement industry-led efforts, rather than duplicate or undermine them. This is the approach that Taiwan’s government-supported Industrial Technology Research Institute (ITRI) adopted, with considerable success. 

In the case of Japan, the government could expand public-private partnerships that serve coordination functions. In 2022, the Japanese Government created the Leading-edge Semiconductor Technology Center (LSTC), an R&D organisation serving as a collaborative forum for academic and industry research roadmapping. The LSTC’s coordination remit could be expanded to ensure that data centre end-use customers are represented within the organisation’s deliberation and planning exercises, and to be more proactive in coordinating R&D decisions between Japan’s materials and equipment industries. 

Second, government support is critical to deal with the rising R&D costs associated with data centre semiconductors. Literature suggests that, compared with grants or subsidies, tax credits usually lead to more diverse, risk-taking R&D projects, which is particularly important given the need for diversified R&D pathways to support novel semiconductor designs. R&D tax credits are well suited to the Japanese context given that Japanese semiconductor firms’ R&D expenditure lags behind their international competitors. The Japanese government took a step in the right direction with tax reforms announced in December 2023, but Japan’s R&D tax credits are still less generous than, for example, South Korea’s. 

Third, government should aim to carefully expand production capacity for leading-edge semiconductors. The Japanese government has already provided Rapidus with 330 billion yen (approx. £1.6 billion) in subsidies and earmarked at least another 646 billion yen (approx. £3.2 billion) for the project; in contrast, the eight Japanese companies that co-founded Rapidus have only contributed 7.3 billion yen (approx. £40 million). The Japanese government could consider making additional public funding contingent on a significant increase in private funding from the eight co-founders. Matched funding contributions between government and private firms have a good track record of supporting innovative projects. Additionally, given that securing a steady stream of customers is critical to Rapidus’s commercial viability, the Japanese government may want to offer demand-side subsidies to incentivise data centre firms to purchase leading-edge semiconductors from Rapidus. 

Fourth, government should help to provide the affordable public infrastructure on which semiconductor fabs rely. This can take the form of providing abundant water supply, as advanced fabs consume approximately 5 million gallons of water per day, the same volume that a small city consumes in a whole year. This also involves streamlining the permitting process to accelerate fab construction. In the case of Japan, the most valuable public infrastructure support relates to electricity. Semiconductor manufacturing is extremely energy-intensive; powering high-precision tools, etching intricate circuit patterns and regulating fabs’ ambient temperature all require large quantities of reliable electricity. However, electricity costs for fabs in Japan are double those in the U.S., Taiwan, and South Korea. Therefore, this could be an argument for subsidising electricity costs for Japanese semiconductor fabs. 

Fifth and finally, government should develop semiconductor skills and talent, alongside the private sector. To increase the domestic skills pipeline, increases in STEM funding for universities and technical colleges are crucial. Additionally, regional consortia that bring together governmental bodies, educational establishments and industry representatives can design new semiconductor-related curricula and develop a database of semiconductor skills requirements and job vacancies. However, the declines in Japan’s birthrate, graduates and workforce suggest that the domestic talent pool is insufficient to address the workforce shortage. This points to a need for immigration system reforms to increase the number of international semiconductor workers in Japan. Currently, the majority of participants in Japan’s Specified Special Worker program must be proficient in Japanese and can only work in Japan for five years, which severely limits the pool of international talent able and willing to immigrate to Japan. In response, the Japanese government could consider a new, dedicated visa for international semiconductor workers. 

Conclusion

The challenges and policy implications arising from data centre chips are not unique to Japan. Indeed, other countries are grappling with similar issues and Japan is actively cooperating with many of them: in 2022, Japan and the U.S. set out their Basic Principles on Semiconductor Cooperation; in 2023, Japan and the European Commission published their Memorandum of Cooperation on Semiconductors; and also in 2023, Japan and the UK struck a Semiconductor Partnership. Although none of these agreements focus solely on data centre chips, their strong emphasis on R&D collaboration and semiconductor workforce development could allow Japan and its partners to capitalise on emerging opportunities from data centre chips.

The views expressed in this article are those of the authors, and do not necessarily represent the views of The Alan Turing Institute or any other organisation.