Two events in 2022 changed profoundly the course of Cold War 2.0. The first was Russia’s full-scale military invasion of Ukraine. The second was the embargo on high-performance semiconductor chips placed by the United States on China. Most people understand viscerally the gravity of the hot war in Ukraine. The deep ramifications of the cold war denial of SCs, and especially equipment to make SCs, against China are somewhat more subtle, but just as critical. SCs are the lifeblood of the modern economy and military. Without them there are no functioning electronics. Without high-performance SCs there are no smartphones and no precision-guided weapons, like missiles. Turns out the two most important events of 2022 were crucially intertwined.
There’s more. It is impossible to develop and deploy sophisticated AI without high-performance SCs. No fancy SCs, no fancy AI. This presents the autocracies with a massive strategic risk. With the SC embargo the democracies have their digital boot on China’s digital throat. Cold War 2.0, barely a decade old, is already at an inflection point. There is a critical election in Taiwan in January 2024. The outcome of that political race could cause China to move against the island before China’s digital deficit in SCs gets inevitably worse. Or perhaps China agrees to become like Japan, a second large Asian power that understands the benefits of a rules-based international order. Such a deal could include China convincing Russia to pull out of Ukraine (including Crimea), and China dropping its own claims to Taiwan, and in return the US embargo on SCs to China would be terminated. The year 2022 was an important one for Cold War 2.0; 2024, a US election year, might be even more eventful.
The last twenty years of growth in the SC industry reflect the significant role they now play in our society. Between 2000 and 2022, annual sales of SCs worldwide increased from $139 billion to $573 billion, a growth rate of 313 percent (roughly 13 percent per year over the last twenty years), while the number of SCs shipped per year increased by 290 percent.1 This growth has been driven by the proliferation of new devices that require SCs. For example, forty years ago cars were still largely solely mechanical devices, but today 20 percent of all SCs end up in automobiles. (TSMC, the world’s leading SC manufacturer, tries to keep all its factories in Taiwan, but it is building several new facilities, including a factory in Germany, to produce SCs specifically for the car market.) Taking all product markets together, in 2021 about 1.15 trillion SCs were made globally.
In the previous chapter AI was discussed as an “accelerator technology.” SCs are accelerators as well, because they have unleashed enormous waves of innovation in many other domains. Research into high-energy physics would not be where it is today without high-performance SCs. The same goes for leading-edge drug discovery, bioinformatics, personalized medicine, nuclear weapons development, or large-area weather forecasting, just to name a few domains of endeavor dependent on high-end SCs. Moreover, the computer, in all its various forms and sizes, has insinuated itself into virtually every nook and cranny of society, triggering massive innovation and efficiency gains at all levels of human endeavor, an achievement made possible only by the SC. Although each SC is tiny, the SC industry is a very big thing.
To grasp the geopolitical ramifications of SCs for Cold War 2.0 it is necessary to understand, at least generally, how SCs are made. In turn, it’s useful to recall how SCs came about; some engineering history is useful here. Computers store data and software as electrical signals and charge. The IBM ENIAC computer of the 1940s (the world’s first “real” computer), contained hundreds of cathode ray (vacuum) tubes. Each tube controlled the flow of electrical charge, which the computer’s central processing unit employed to carry out its fast computations. It made sense to use cathode ray tubes in early computers, because they could be found in radios and televisions, and so their physics and engineering were well understood and their cost was reasonable.
Still, everyone who made computers in the 1950s knew that cathode ray tubes had fundamental limitations. It’s actually very similar to the dilemma with battery technology today. Lithium-ion batteries have brought the world quite a distance in the great energy shift from fossil fuels to renewable sources of electricity, but at the same time scientists and politicians understand that the current battery technology will not be sufficient to get to net zero. The world needs a very significant technology breakthrough in batteries; in effect, the energy storage industry is still waiting for its SC to be invented.
In computers the huge breakthrough came when two engineers at Bell Labs, a private research arm of a US telecommunications equipment manufacturer, innovated the concept of engraving a pattern of copper into a small substrate of sand, or silicon, to form a tiny circuit. The two American inventors of the first SC called the device a “semiconductor” because of the way the copper and the silicon worked to let some electrical signals through but block others. Subsequent improvements produced a more stable technology for the SC that could accommodate hundreds (and ultimately hundreds of millions, and today billions) of tiny circuits.
A single circuit, when it was on, registered as the number 1, and when it was off it was 0 (zero). The computer, in effect, could only speak in the language of 1s and 0s, while our general mathematical language contains ten symbols: 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9. With these ten symbols any number, however long, can be represented. So how can the SC work as effectively (and indeed even better) with only two symbols, namely 0 and 1? The answer, partly, is speed, incredible, unbelievable rates of speed.
The first SCs sixty years ago weren’t all that fast, and in fact weren’t a lot faster than cathode ray tubes, but SCs (even the earliest ones) were much smaller. Computers, because of their SCs, have gotten a lot faster over time. The ENIAC computer circa the 1940s took up an entire midsized room and could do about 5,000 additions per second. By 2008, the first iPhone could do about 200 million additions per second. Today’s iPhone does even much better than that. An Apple iPhone 13 Pro series contains a single chip with the following separate parts to it: two high-performance processors, four high-efficiency processors, a five-processor graphics SC, and a 16-processor neural engine that performs 16 trillion operations per second—and all of this avalanche of processing power is contained in a single smartphone! Small is the operative word when it comes to SCs.
SCs have always been small, but their real magical quality is how they have kept getting ever smaller every few years. Moreover, stunningly, with each generation of miniaturization the SCs also became more powerful by squeezing many more transistors onto the SC. Gordon Moore, one of the founders of Intel, a leading American SC maker, even predicted in 1965 that every two years (or so) the computing power of SCs will double because the size of the transistor continued to shrink. This became coined as “Moore’s Law,” and up to about 2015 the SC industry made good on Moore’s prediction. As if this wasn’t already pretty amazing, their cost, relative to their computing power, has dropped dramatically as well. If a car had undergone the same price-performance improvement ratio over the past seventy years, today’s gas-fueled vehicles would get millions—and probably billions—of miles on a gallon of gas.
It is difficult for the human brain to actually grasp just how miniature the guts of SCs have become. For instance, the different levels of SC performance are measured in “nanometer processes.” The lower the nanometer number associated with the particular SC, the more powerful it is, because that SC has yet more billions of transistors squished onto a tiny surface of silicon. A 7 nm SC (made using a 7-nanometer process) can get between 95 and 115 million transistors per square millimeter. A 5 nm SC can squeeze 125 to 300 million transistors onto the same space. Therefore, today, a 3 nm SC is about 60 percent more powerful than a 5 nm SC.
In one sense, the difference between 5 nm and 48 nm seems fairly academic to the average layperson. The unaided human eye cannot even begin to register these tiny measurements, and the human brain cannot absorb their dimensions either. SC designers and makers, though, have the tools with which to make these minuscule differences meaningful, and so they are important in the SC world. More to the point, the difference between a 48 nm SC and a 5 nm SC (and between a 5 nm and a 3 nm SC) is quite pronounced when it comes to processing certain software and data in the real world. Given the discussion of critical AI in the previous chapter, it is noteworthy, for instance, that AI running on a 48 nm SC will not do nearly as good a job training on large data sets, especially those relating to images, as a 5 nm SC. In short, a company wanting to do state-of-the-art AI today needs a 5 nm SC, and in a few years’ time, the same company will want a 3 nm SC, and so on. The ramifications of this reality for companies in the AI space will become apparent when discussing below America’s sanctions on China for certain high-performance SCs.
There are several kinds of SCs, with the two major categories being processor SCs and memory SCs. The processor ones are, as the name suggests, those that do most of the computer’s heavy lifting; they contain the logic algorithms of the SC, and they actually carry out the secret-sauce functions of the SC. Often traditionally called a CPU, for central processing unit, those SCs that do training of the computer on large data sets for AI are now called GPUs, for graphical processing units, especially if they are expert at capturing and learning from images and videos; computers doing AI functions will also contain a neural processing unit, distinct from the CPU and the GPU. The performance of these processor or logic SCs are measured in nanometer processes, as noted immediately above, such as a 3 nm SC that is more powerful than a 5 nm SC.
Memory SCs specialize in storing data. Their performance is measured differently from that of processor SCs, because memory SCs are generally trying to expand their memory capacity on the same sized chip space, and they do so by adding many layers on top of each other. Today, the best-performing memory SCs are at the 128-layer level, but some SC makers are working on coming out with a 176-layer memory SC in a few years, which will improve performance significantly.
Making an SC is a complex, laborious, and expensive process. It is not for the faint of heart, or the impecunious. First, the SC is designed by making a “blueprint” of how the millions of tiny circuits on a minuscule piece of silicon wafer should be arranged. This can only be done using very complex and very expensive SC design software from one of only three software companies in the world specializing in making this unique electronic design automation (EDA) software. There are also some companies that own “templates” useful for making some elements of these designs. These templates can speed up the design process significantly if they are used as part of the base of the new SC’s design. A British company, ARM, has the leading template for smartphone computer processor SCs, while Intel has a leading template for anyone wanting to make an SC for x86 type computer servers.
Once the design for the SC is complete, it is sent to an SC fabricator, or “fab,” where the actual SC will be made. The fab etches the design onto a wafer of silicon, creating the minuscule circuits. The machines used for the etching processes are some of the most complex in the world. The fab operator doesn’t make these machines, but knows how to operate them in conditions where the light, heat, and other conditions are very carefully calibrated for quality control. Operating all the complex equipment in the fab, and producing world-class SCs is very, very difficult. Even major SC manufacturers can get the etching process wrong, and then ruin the performance of the resulting SCs.2
It used to be that only large manufacturers like Intel designed SCs, and then they designed general SCs that everyone else would use more or less for the same purposes. Those days are long gone. Today Intel still designs SCs, but so do two other categories of SC designers. First are SC companies that do nothing but design SCs, and then sell them widely to other companies once the SC is fabricated by someone other than the SC designer. AMD and Nvidia are the two leading such SC design firms in the world for processor SCs. The other SC designers are tech companies who want to design SCs for their own use exclusively. Apple is a good example. It has very particular demands of the SCs it uses, for example for its smartphones, so that it designs its own SCs, which it then has manufactured for it exclusively. Amazon is also starting to design its own SCs. The designers of SCs, like AMD, Nvidia, Apple, or Amazon, use the EDA software from one of three EDA suppliers: Synopsys, Cadence, or Siemens EDA (formerly Mentor).
The SC designers, like Nvidia or Apple, do not have the factory able to actually make the SC they have designed. For this stage of the process, the designer goes to a “fab” operator, like TSMC, and TSMC makes the SC at its fab but to the strict requirements of the designer. The fab operator needs to undertake about 1,000 process steps to get the silicon wafer ready for etching. This could require about fifty types of very specialized equipment (called SME, or semiconductor manufacturing equipment).
As part of the SC manufacturing process in the fab, for the most advanced process nodes (such as 5 and 3 nm), the fab operator will take the “blueprint” design (created with the EDA software), and load it into a huge machine that is at the heart of the SC fabrication process for the most powerful SCs in the world. The machine, made by ASML in the Netherlands, is about the size of a large truck. This machine can “read” the SC design in the blueprint, and then etch that design onto a piece of silicon wafer using a tiny beam of light. It sounds simple described in these general terms, but this single process is perhaps the most complicated industrial activity known to humankind. There is only one such etching machine (made by ASML) in the entire world that can manufacture SCs in the 7, 5, and 3 nm process nodes, and it uses a method called Extreme Ultraviolet Lithography (EUL).
The scientists and engineers at ASML in the Netherlands—and the many engineers and customers in the rest of Europe, America, Taiwan, Japan, and South Korea who assisted—who built the EUL etching machine took twenty years to wrestle with the related basic science, and then the very finicky engineering, to get the machine to work.3 As the innovation process around the EUL machine took thirteen years longer than anticipated, ASML had to rely very heavily on its optical light supplier (Zeiss in Germany), as well as the three lead customers, namely TSMC in Taiwan, Samsung in South Korea, and Intel in the US, each of which ended up taking an equity stake in ASML. Ultimately, though, all this international collaboration paid off. For several years the price of a single EUL machine was $100 million, but the latest version, which contains even more state-of-the-art innovation, sells for $500 million.
Once the ASML machine etches the required pattern for the billions of circuits onto the SC, various chemical processes deposit just the right amount of semiconducting material in the appropriate places on the SC. Subsequently, individual SCs have to be cut out of the wafer, and then each SC can be cleaned up, tested, and sold. (If you are AMD or Nvidia; if you are Apple, you will keep all your custom SCs coming out of the fab for inserting into your own devices, like iPhones.) As noted above, making an SC is a complex process.
The market and supply chains for SCs are fairly complicated and quite global. An SC might be designed in the US using American EDA software, but also some intellectual property from ARM in the UK, and then the SC is made in Taiwan at a TSMC plant, but then packaged and finished in a facility in China. Throughout this process, the supply chain would touch various specialists. South Korea’s Samsung specializes in making memory SCs that reside in all sorts of computers; they store the data relevant to the various processes being asked of the SC. In the United States, there is the former world leader, Intel, who used to make the dominant SC in the marketplace, namely the processor chip for the personal computer.
In the United States, Intel is still the largest SC manufacturer, but there are other players as well, including Micron, which makes memory chips for the most part. Then there are a number of other designers of chips in the US. Most American SC designers have their SCs made by a fab owned and operated by Intel or Micron, except for their highest-performance SCs, which need to be made at one of TSMC’s fabs in Taiwan (though soon also at a TSMC facility in the US), because the world’s most advanced fabs are on that small island, 160 kilometers across from a quite hostile China. What’s wrong with this picture, especially if the ultimate customer for the particular SC is the United States military?
In China, SMIC (Semiconductor Manufacturing International Corp) operates fabs that can make SCs in the 25 to 14 nm range. There are a few smaller companies that can make less powerful SCs in the 48 to 25 nm range. Yangtze Memory Technologies Corp. (YMTC) makes memory SCs, including at the 128-layer performance level. All these fabs, though, use the ASML etching machine, or less powerful etchers from one of two Japanese companies, Nikon and Canon. In short, China is not self-sufficient when it comes to the SC supply chain. In 2022 China imported about $400 billion worth of SCs, many from the fabs in Taiwan. As for the less powerful SCs made by fabs in China, they currently rely almost completely on SC manufacturing equipment from Japan, the United States, and the Netherlands. It is estimated that Chinese companies produce less than 10 percent of the SC manufacturing equipment being used in fabs in China.
The M&A market is critical to the leading suppliers of SME’s and others in the SC supply chain ecosystem. Some recent major acquisitions have been AMD acquiring Xilinx, and Intel acquiring Tower Semiconductor. At the same time, though, Nvidia’s acquisition of ARM was blocked by regulators in early 2022, but other than that, most deals involving buyers and sellers in democracies have been allowed to proceed. On the other hand, Chinese firms will have an increasingly tough time buying companies in the democracies, just as Tsinghua Unigroup’s acquisition of Micron Technologies was blocked by the US government in 2015, Fujian Grand Chip’s acquisition of SME Aixtron was blocked by the German government in 2016, and a Chinese company’s proposal to buy the Newport Wafer Fab in 2022 was blocked by the UK government.
Here is a list of the key players in the global SC industry, indicating where they are active in the vertical supply chain of the industry, and their sales and market capitalization. Also noted is the country’s market share of the particular portion of the SC industry. Note that for purposes of this list, Taiwan is assumed to be in the camp of the democracies.
Democracies |
||
United States, 70 percent market share |
Revenues |
Market Cap |
Synopsys |
$5.2 billion |
$65 billion |
Cadence |
$3.6 |
$62 |
Mentor (now Siemens EDA, Germany) |
$3.1 (Siemens $79 B) |
$138 (Siemens) |
Autocracies |
||
China, 10 percent |
Revenues |
Market Cap |
Empyrean Technology Co. |
$0.115 |
$9.4 |
Democracies |
||
United States, 28.3 percent market share |
Revenues |
Market Cap |
Synopsys |
$1.314 billion |
$65 billion |
Cadence |
$0.357 |
$63.4 |
Ceva |
$0.134 |
$0.58 |
Rambus |
$0.087 |
$6.59 |
United Kingdom, 46.5 percent |
Revenues |
Market Cap |
ARM |
$2.74 billion |
est. $60 billion |
Alphawave |
$0.175 |
$1.16 |
Taiwan, 1.6 percent |
Revenues |
Market Cap |
eMemory Technology |
$0.105 |
$5.38 |
Autocracies |
||
China, 2 percent |
Revenues |
Market Cap |
VeriSilicon (with some investment backing by Intel and Samsung) |
$0.133 |
$4.9 |
Democracies |
||
United States, 64 percent market share |
Revenues |
Market Cap |
Nvidia |
$26.9 |
$1.04 trillion |
AMD (Advanced Micro Devices) |
$23.6 |
$184 billion |
Qualcomm |
$44.2 |
$127 billion |
Apple (SCs for own use) |
$394.3 (total) |
$3.01 trillion |
Amazon (SCs for own use) |
$514 (total) |
$1.32 trillion |
Microsoft (SCs for own use, esp. AI) |
$198.2 (total) |
$2.54 trillion |
Meta (SCs for own use, esp. AI) |
$116.6 (total) |
$753 billion |
Google (SCs for its own use) |
$279.8 (total) |
$1.53 trillion |
IBM (SCs for its own use and R&D) |
$60.5 (total) |
$120 billion |
Taiwan, 18 percent |
Revenues |
Market Cap |
MediaTek |
$18.7 |
$35.3 |
Realtek |
$3.6 |
$6.9 |
Autocracies |
||
China, 15 percent |
Revenues |
Market Cap |
HiSilicon (owned by Huawei) |
[private] |
[private] |
Unisoc (owned by Tsinghua Unigroup) |
[private] |
[private] |
Democracies, 90 percent of the global market |
||
Japan |
Revenues |
Market Cap |
Shin-Etsu Chemical |
$17.2 |
$65.7 |
SUMCO Corp. |
$3.1 |
$4.9 |
South Korea |
Revenues |
Market Cap |
SK Siltron css (part of SK, a very large South Korean conglomerate) |
$103 |
$5.9 |
Taiwan |
Revenues |
Market Cap |
GlobalWafers |
$2.2 |
$6.8 |
Autocracies |
||
China |
Revenues |
Market Cap |
Hua Hong Semiconductor |
$2.4 |
$4.3 |
Democracies |
||
United States |
Revenues |
Market Cap |
Applied Materials (90 percent of market for deposition and doping equipment) |
$26.6 |
$120 |
KLA |
$10.6 |
$64 |
Lam Research (etching machines) |
$18.8 |
$84 |
Japan |
Revenues |
Market Cap |
Tokyo Electron (90 percent of photoresist market) |
$16.5 |
$69 |
Canon |
$30.3 |
$26 |
Nikon |
$4.4 |
$4.5 |
Netherlands |
Revenues |
Market Cap |
ASML (100 percent of EUL [Extreme-Ultraviolet] lithography/etching equipment for SCs below 7 nm) |
$25.5 |
$282.3 |
Autocracies |
||
China, 8 percent of local requirements and 2 percent of global supply |
Revenues |
Market Cap |
SMEE (Shanghai Micro Electronics Equipment) (only at 90 nm process) |
[private] |
[private] |
Democracies |
||
United States |
Revenues |
Market Cap |
Intel |
$56.4 billion |
$136.6 billion |
Micron |
$23.0 |
$67.9 |
GlobalFoundries |
$8.1 |
$34.9 |
Taiwan |
Revenues |
Market Cap |
TSMC |
$74.5 |
$526.4 |
United Microelectronics |
$9.0 |
$19.2 |
South Korea |
Revenues |
Market Cap |
Samsung |
$65.6 ($233 entire) |
$367 |
SK Hynix |
$36.2 |
$63.3 |
Autocracies |
||
China |
Revenues |
Market Cap |
SMIC (Semiconductor Manufacturing Int’l Corp) |
$6.8 billion |
$28.6 |
YMTC (owned by Tsinghua Unigroup) |
[private] |
[private] |
GigaDevice |
$1.12 |
$10.1 |
Taiwan |
Revenues |
Market Cap |
ASE Technology Holding |
$12.5 |
$16.4 |
China |
Revenues |
Market Cap |
JCET Group Co. Ltd. |
$4.6 |
$7.9 |
NAURA Technology Group |
$2.0 |
$15.5 |
In September 2022, Jake Sullivan, the national security advisor to President Joe Biden, gave a very important speech regarding the SC industry.4 He outlined what steps the United States was planning to take to deny Chinese companies, and thereby the Chinese military, access to the most powerful SCs. Sullivan indicated this had to be done for two reasons. First, China has a policy of civilian-military fusion, such that any SCs imported or developed by Chinese companies would invariably be made available to the People’s Liberation Army (the PLA), China’s armed forces. Second, China has instituted an oppressive system of citizen surveillance and oppression unmatched anywhere in the world, and this structure, which denies basic human rights to average Chinese citizens, is built upon technologies that have included components sourced from the democracies. Accordingly, Sullivan indicated that the US government would be taking steps to block high-performance SCs, and the equipment to make them, from finding their way into China.
For some time, China had requested TSMC build and operate in China a semiconductor fab for SCs in the 7 to 5 nm range, but the Americans have told TSMC and the Taiwanese government not to do so. More generally, following the Sullivan speech, on October 7, 2022, the Biden administration announced an embargo on American companies selling SCs to China that are 14 nm or more powerful (so, nothing in the 14 to 3 nm range for current SCs). Therefore, companies like Nvidia, AMD, and Intel cannot sell any more of their most advanced SCs to China, or to anyone that they know will pass along the SCs to Chinese companies or entities.
These sorts of restrictions have been in place for certain SCs for some time. In 2018, the Trump administration blocked the sale of American-made SCs to certain companies in China, essentially those involved in making devices for the Chinese military and those making devices used to conduct surveillance on Chinese citizens or the Uighur minority in Xinjiang. One company that was embargoed was Huawei, the giant Chinese telecom equipment maker and leading supplier in the world of telecom equipment for networks, with sales worldwide in 2017 of $92.5 billion. When this US embargo came into force, Huawei had to cancel billions of dollars of orders for equipment that used American SCs. Huawei’s sales plunged by 25 percent in 2018 and 2019. Huawei could find substitutes for some of the embargoed American SCs, but not all of them.
What’s new about the Biden administration’s embargo announced in October 2022 is that they prohibit the sale of not only high-end SCs, but also the equipment used to make those SCs. A number of US companies have lost material sales to Chinese companies as a result, and have suffered losses in the range of hundreds of millions of dollars in the six months following the announcement of the embargo. From the industry chart above, it is clear that there are a number of important SME makers outside of the United States. Therefore, an important question was whether the Japanese and the Dutch would go along with the embargo in respect to SC-making equipment they would otherwise sell (very profitably) to China. The answer so far is yes, and again the Chinese are none too happy about their high-end SC manufacturing base being crippled in this way. It is likely China will be able to continue making SCs in the 45 nm and older nm processes range, but will be stymied for some (and perhaps many) years in making higher-end SCs. Most critically, if China cannot buy ASML’s EUL lithography etching machine, it may take China about twenty years (or more?) to build a replacement for this marvel of high-tech manufacturing.
The effect of the SC and SC-making equipment embargo on China and Russia implemented by the US and some other democracies will be significant. Not only will China and Russia not be able to procure high-end SCs for some years—likely between fifteen and twenty years—from the democracies, but this embargo will stunt the growth of the AI industry in China and Russia because increasingly the leading AI programs need very powerful computers to run on. If those computers don’t have AI-specific SCs (such as the AMD V100, or Nvidia A100), then this will block China’s plan to become a world leader in AI by 2026. In effect, the democracies likely have about ten years of breathing room before the Chinese push to create a homegrown substitute SC etching machine bears fruit, but even then, replacing ASML’s EUL machine might well take double that time.
For a number of months China did not retaliate generally against the Biden administration and American companies in respect of the US equipment embargo, but in June 2023 Beijing announced that state-owned utility companies, especially in the telecoms domain, were henceforth prohibited from using SCs sourced from the American company Micron Technology because they had security shortcomings. This is widely seen as a tit-for-tat response to the US embargo on SCs (and SC-making equipment) implemented in October 2022. It is likely that additional retaliatory actions by the Chinese will be forthcoming. In the meantime, China will likely make an even greater push to build its own indigenous supply chain of Chinese businesses and technologies that can make higher-end SCs than is the case today. If China fails in this effort, it will fall behind seriously in AI development, as well as the production of military-related AI products and other weapons or industrial products that require the most powerful SCs. Given the massive importance of AI to China (and to the US and every other country on earth), an outcome where high-end computing power—including in the AI domain—was denied China would pose a very serious problem for China.
The challenges posed to China by having to work with a long-term technology deficiency will likely become even more acute should the Biden administration enact further technology-related controls on China. Given the relative success of the export controls on high-performance SCs, the additional measures being discussed in Washington include restrictions on US investment in Chinese AI, QC, and biotechnology companies; blocking the export to China of all SCs that are intended to process AI applications; and prohibiting Chinese AI companies from accessing computers through cloud arrangements with American suppliers.
Beijing might retaliate against these (and even just the existing) measures by limiting the export from China to the US of certain critical minerals that China has a virtual monopoly on at this point, such as dysprosium and terbium. The question for Xi will be: Does he want to start a full-scale trade war with the United States? It began when the Trump administration launched its own restrictions on SC sales to Chinese entities that ended in a draw. Would Xi do as well now, particularly since the Biden White House is livid that China is assisting Russia as much as it is in respect of the war in Ukraine? China has stepped in to buy Russian oil and gas when the Europeans ceased buying these vital exports from Russia, as the Europeans applied sanctions against Russia over its unprovoked attack on Ukraine. As well, China’s exports to Russia are up about 34 percent since Russia began the war in Ukraine—this is the figure that particularly riles the White House, as it includes the sales of reconnaissance drones from China to Russia.
For his part, President Biden is also worried that China might invade Taiwan. If that were to happen, TSMC’s most advanced SC fabs would certainly be adversely affected by the military conflict, if not completely or partially destroyed. If fighting cut off the supply of high-end SCs to the United States, that could be devastating for the economies of the democracies. For this reason, the Biden administration is pursuing two important initiatives in SC industrial policy.
In terms of reliance on foreign SC manufacturing expertise, the US government is unhappy with America’s heavy dependence on TSMC to make 90 percent of the high-end SCs that US companies require. Accordingly, the Biden administration has negotiated with TSMC to build two fabs in the United States over the next few years. (Fabs take a long time to build, and they cost about $15–20 billion to construct, including installing all the SC manufacturing equipment noted above.) One fab is intended to make SCs in the 7 to 5 nm range for the US military and its prime defense contractors. The other fab will make SCs in the range of 5 to 3 nm, as advanced as the SCs that TSMC currently makes in Taiwan. TSMC has also agreed to put up a fab in Japan to make SCs in the 7 to 5 nm range. TSMC will also establish a fab in Europe. If all these fabs get built in the next number of years, they will go a long way to usefully diversifying the risk of a Chinese military move against Taiwan.
In a similar vein, in the summer of 2022 the US president proposed, and the American Congress approved, the Chips Act, which, among other things, makes available about $52 billion in subsidies and tax breaks to American and foreign SC makers (like TSMC) that build new fabs in the United States that are able to produce the highest-end SCs. Intel and Micron (both American companies) and TSMC (from Taiwan) have taken advantage of this US government program, and each have announced they will build fabs in the US in the next couple of years. Many other companies have expressed interest in participating in the Chips Act program.
Between this industrial policy requiring TSMC to build new fabs in the United States, providing financial incentives to US-based and non-Chinese foreign SC manufacturers, and the American embargo on sales to China of high-end SCs and SC-making equipment, Cold War 2.0 is certainly underway in the SC sector of the global economy. The Biden administration’s goal is nothing less than the commercial crippling of the Chinese high-end SC industry sector. Presumably at some point in the near future the Chinese will respond with their own actions against the US and the other democracies implicated in policies that could prove devastating to China’s previously ambitious plans for developing their SC and AI sectors. (Because they are key accelerator technologies, they would also have a deleterious impact on virtually all other innovation in China.) As noted above, the Chinese have taken a relatively small retaliatory step against the US company Micron, but it is unlikely this will be the end of China’s pushback.
China will invariably partly respond to the actions of the Americans by looking to build up China’s own SC manufacturing equipment supply chain. Chinese SC fabricators SMIC and YMTC could likely operate factories that produce 7 nm and 3 nm SCs, so long as they had SC manufacturing equipment that replaced the ASML, Japanese, and American equipment needed to do so. What is the likelihood of Chinese companies succeeding in producing that equipment and the thousands of components that go into these hyper-complex machines? Not very good, based on recent history.
In 2014, the Chinese government in Beijing created an investment fund specifically aimed at creating greater capability in the SC manufacturing domain. It is called the China Integrated Circuit Industry Investment Fund, but is known more colloquially as “Big Fund I.” The fund raised just under $22 billion. The way the fund worked is that bureaucrats in Beijing would select the recipient of financial support from Big Fund I, and then that company would approach a local government in China to negotiate tax breaks or other matching funds from the municipality in consideration for agreeing to site a new fab or other factory in the municipality. Only when deals were done with both levels of government would the Chinese company invite other investors from the private sector into the new venture.
There were two big problems with this mechanism for conducting industrial policy. First, by having the government lead the investment initiative, it was government bureaucrats who decided which Chinese companies received the money and how much. Not surprisingly, the bureaucrats went with the safe option and gave most of the money to well-established companies like SMIC and YMTC, in order for them to build more of what they were already doing. Each of these companies used the Big Fund I money to build more fab capacity (SMIC for processor SCs and YMTC for memory SCs), but that isn’t what China needed out of a SC industrial policy. In 2014 what was desperately missing in the Chinese supply chain for SCs was a viable Chinese player that can make world-class SC etching/lithography equipment, and that was hardly an afterthought in the first round of Big Fund I investments. In short, China crucially needed a “Made in China” ASML, but that was exactly what the bureaucrats responsible for Big Fund I failed to try to seed with their money, largely because they simply didn’t know how to go about such a daunting task. China also needed a world-class player in the EDA space, so that SC developers could have adequate tools with which to design the highest-performance SCs. Again, Big Fund I missed this market segment as well.
The second, related problem, is that at the end of the day, there was (and continues to be) insufficient input from private sector investors as to which companies the Big Fund I money should be invested in. Chapter 2 discussed how dysfunctional the Russian central planning, top-down approach to industrial strategy was in Cold War 1, and how it especially stunted the Russian technology sector, both for the computer industry and biotechnology sector. China repeated the same type of mistakes when it completely missed the opportunity to build a SC manufacturing equipment industry in China.
The result is a very serious disaster for China in the SC domain (and for the design, development, and deployment of AI software), now that the US is restricting the flow of high-end SCs and equipment necessary to make them. For many years Big Fund I was reluctant to fund the development of Chinese vendors of EDA software; now, the Chinese deficit in EDA software is proving to be a very serious hole in China’s domestic SC supply chain.5 Moreover, consider the daunting wall that China has to get over to make viable EDA tools. The leading Chinese player in EDA is Empyrean. It has tiny sales relative to the American giants of the EDA space. Empyrean has no meaningful sales outside of China, and less than 10 percent of the Chinese market. It has about 150 people working in R&D. Compare that to Cadence, the American global market leader, which has about 5,000 engineers working on EDA R&D. As well, Synopsys supplements its technical capacity through an aggressive program of acquiring talent; in the last 30+ years it has acquired some 80+ companies.6 In the field of EDA software China does not have a player that can substitute, even remotely, for the American leaders in the market.
It gets worse for China. In the several months following the October 2022 announcement of the new US SC sanctions program, the Chinese government assessed how the investments made from Big Fund I have helped build an indigenous Chinese SC manufacturing equipment industry. When the bureaucrats in Beijing concluded (correctly) that Big Fund I did very little, if anything, to foster technology that could replace high-end equipment from the democracies, instead of calling Big Fund I a failure and trying something new, they did two things. First, in August 2022 they blamed the failure of Big Fund I on corruption by the three leaders of Big Fund I and one of the leaders of the SC industry in China, a tech entrepreneur who ran Tsinghua Unigroup, one of the stalwarts of China’s SC industry. These four people have been arrested and detained for corruption.7 The outcome of this proceeding is still unknown, but the arrests alone prompt a second observation.
The government in Beijing is trying yet again with another Big Fund (called “Big Fund II”). This one is aimed more squarely at building manufacturing capacity in the SC manufacturing equipment space. It’s too early to tell how this one will work out, but if Beijing uses the same model as Big Fund I, there is very little hope of it succeeding. Moreover, who from China’s private sector will be keen to take on the project of building the Chinese competitors to Cadence, Synopsys, and ASML, when the last fellow from the private sector who got involved in Big Fund I is now languishing in jail? Not many, or at least not many of quality, which is precisely the type of executive China needs for this ambitious project. The essential point, though, is that top-down politically managed funding of tech ventures, as practiced by autocracies, simply doesn’t work. It will be interesting to see how many more billions of dollars worth of public investment China wastes before it learns this simple lesson.
It should be noted that the failure of the Chinese model of investment strategy/funding for very sophisticated technology projects is not restricted to the SC sector. China has also been unable to design, develop, and manufacture jet engines for commercial aircraft. In May 2023 China announced with great fanfare that its project to bring to market a commercial aircraft that could compete with Boeing and Airbus had finally come to fruition, with the first successful flight of the new Comac C919 narrow-body short-haul jet. The bad news, though, for China at any rate, is that the engines for this plane are supplied by an American/French joint venture company. Moreover, the plane’s avionics (the complex software that is used to actually fly the plane and operate many of its features) is also sourced from the democracies. Once again, Chinese technology innovation is showing how far behind it is relative to the democracies. As well, if a Chinese invasion of Taiwan triggered an across-the-board sanctions policy from the democracies, the C919 would cease to fly.
A final point about SC industrial strategy is timing. It took the incumbent, ASML, about twenty years to crack the code for EUL etching. Assume it takes China only ten years, given that they have a number of ASML machines that they can reverse engineer. By the time China completes this exercise and is producing 5 nm or 3 nm SCs at SMIC, ASML will be well into the next generation of etching machine, and will be doing 1 nm and more powerful, so China’s inferior position relative to the democracies will not be alleviated. This is why China is putting so much energy into quantum computing, to which we now turn.
Could all of the above factors—the American SC and SC manufacturing equipment embargo, further US technology embargoes, and China’s weakness in making technology at the highest levels—lead to a grand bargain between the United States and China? Here’s what it might look like. China convinces Russia to pull out of Ukraine including Crimea; China allows Taiwan to become fully and legally independent; and China agrees to become a normal member of the international community and adhere to the rules-based international order—i.e., China drops its maximalist claims to the South China Sea and agrees to abide by the UNCLOS, and also agrees to enter into a nuclear arms limitation treaty. And in return the United States lifts all technology embargoes against China. Might such a grand bargain be the capstone of the second term of a Biden presidency?