AI Chips Key Takeaways

1. The application of artificial intelligence (AI) technologies to numerous sectors of the real economy is expected to become a growth driver for the semiconductor industry. Specialized AI chips are indispensable for realizing the full potential of AI, and these chips will claim an increasing share of the global semiconductor market.

2. That’s because AI runs on algorithms, data, and computing power. Chip hardware will need to accommodate exponential growth in data volumes and handle the complex, layered matrix calculations needed to teach non-routine human tasks to computers. These use cases range from facial recognition and natural language processing to autonomous driving and recommendation engines, among many others.

3. AI chips are much better suited than traditional CPUs for both training algorithms and for running them in diverse real-world applications. That’s because AI chips are more powerful, designed for specific end-uses, and programmable for narrowly-defined functions. The ability to scale AI applications across industries will depend on advances made in the design and manufacture of these chips. Stable supply chains for AI chips look set to become pivotal to economic growth, technological progress, and national security.

4. The upstream segment of the AI chip supply chain is similar to other semiconductors, consisting mainly of silicon and boron. Both elements are relatively abundant and are not at major risk of supply disruptions. The midstream segment, however, is the most complex and involves suppliers across multiple continents. This segment can be divided into three stages: design, fabrication, and assembly and packaging. These three steps create roughly 45%, 45%, and 10%, respectively, of a chip’s value.

5. The underlying architecture of any chip typically derives from core intellectual property (IP) licensed from a handful of American, British, and Japanese firms (e.g. ARM Holdings). Other companies then create custom chips based on this core IP, including specialized “fabless” design firms that outsource chip fabrication (e.g. Qualcomm, Broadcom), and “integrated device manufacturers” (IDMs) that also fabricate their own chips (e.g. Intel, Samsung). The top design firms are located in Europe, Japan, South Korea, Taiwan, and the United States. Since design happens in-house, this stage is mostly vulnerable to legal risks, such as industrial espionage and regulatory barriers to attaining core IP licenses.

6. Fabrication is perhaps the crucial link in the supply chain, and the one with the most potential for disruption. Barriers to entry are high, involving enormous capital expenditure, atomic-precision manufacturing equipment, and highly skilled personnel. Established players thus enjoy considerable advantages. A parallel can be drawn with commercial aviation, another highly concentrated apex industry.

7. Contract fabrication is dominated by the Taiwan Semiconductor Manufacturing Company (TSMC), which enjoyed a roughly 70% market share in 2018. TSMC is basically the “Foxconn of semiconductors,” as that single company makes the chips designed by most major semiconductor players, from Qualcomm to Apple and Huawei. (This market excludes IDMs like Intel and Samsung, which own their chip foundries. Intel, for example, operates six foundries worldwide, including one in Dalian, China). In fact, Taiwan currently has about 80% of the foundry market.

8. Foundries like TSMC depend on reliable supplies of sophisticated and expensive fabrication tools, known collectively as semiconductor manufacturing equipment (SME). The SME market is concentrated in a few firms headquartered in the United States, Japan, and the Netherlands. The Dutch company ASML, for example, sells around 75% of all photolithography machines, which can cost $120 million each. China has no advanced SME companies and has become the world’s largest SME importer.

9. The final segment in the midstream supply chain is the testing and packaging of completed semiconductors. This segment, referred to as outsourced semiconductor assembly and testing (OSAT), has low margins and is highly competitive. Although some of the biggest OSAT contractors are in Taiwan and China, the industry is dispersed among many companies and geographies. (IDMs like Intel also have their own OSAT facilities, some of which are located in China and Southeast Asia.)

10. The downstream segment of AI chip supply chains refers to the distribution of finished chips for their various end-uses in cloud servers or edge devices. While no Chinese company appears in the top 15 global suppliers of semiconductors—a list dominated by American, South Korean, Taiwanese, Japanese, and European firms—the advent of custom AI chips could present an opportunity for Chinese firms to keep moving up the value chain.

11. Despite being a huge consumer of semiconductors and its government’s provision of generous industrial subsidies, China is still uncompetitive in producing advanced chips, particularly those needed to power AI capabilities. Although Chinese tech giants like Huawei have designed top-of-the-line AI chips, they are still manufactured by TSMC. This imbalance between consumption and production of chips means that China remains vulnerable to disruptions in its AI chip supply chain.

12. China does retain leverage as the world’s largest buyer of semiconductors. Shutting Chinese firms out of global supply chains for AI chips would eliminate a considerable source of revenue for Western semiconductor firms, especially since the consumer electronics industry—and therefore end demand for chips—is essentially concentrated in China. Such a move would limit the ability of these Western firms to sustain the substantial capital and R&D investment—typically around 30% of revenue—needed to maintain technological leadership over competitors like Huawei. Therefore, if the policy preference is for the United States and its allies to retain technological leadership, then a more prudent approach may actually work more effectively.

13. However, some national security considerations may still be necessary in regulating an apex industry like semiconductors. But such actions will likely need to be narrowly focused on the most concentrated sections in the global supply chain. For example, the United States could work with allies in Europe and Asia to bring oversight to the export of the most advanced SME. Another possibility would be for the United States to support key foundries like TSMC to diversify the location of some advanced chip production into various allied countries to mitigate supply chain risks. This would allow semiconductor firms to continue selling to their biggest customer, China, while helping to preserve technological leadership and moderate calls for self-reliance in Beijing.

14. Ultimately, short-term solutions are unlikely to achieve the desired outcome of sustaining US leadership in this highly globalized and complex industry. The most effective long-term solution will be to design policies that incentivize US companies to stay ahead of its competition. This may require public-private partnerships to develop frontier innovations that will render current technologies obsolete, such as quantum computing. It could also involve incentives that encourage US companies to double down on their comparative advantage in software and ecosystem platforms—in order to gain market share. This long-term strategy requires renewed state support for basic research, further incentives for industry R&D spending, government and industry collaboration, open global markets to commercialize and monetize new innovations, and an environment that attracts and keeps the world’s best talent in America.

@MacroPoloChina MacroPolo - Facebook MacroPolo - LinkedIn