Chip War: The Fight for the World's Most Critical Technology

by Chris Miller

Today, the biggest analog chipmakers are American, European, or Japanese. Most of their

production occurs in these three regions, too, with only a sliver offshored to Taiwan and South Korea. The largest analog chipmaker today is Texas Instruments, which failed to establish an Intel-style monopoly in the PC, data center, or smartphone ecosystems but remains a medium-sized, highly profitable chipmaker with a vast catalog of analog chips and sensors. There are many other U.S.-based analog chipmakers now, like Onsemi, Skyworks, and Analog Devices, alongside comparable companies in Europe and Japan.

memory market, by contrast, has been dominated by a relentless push toward offshoring production to a handful of facilities, mostly in East Asia. Rather than a diffuse set of suppliers centered in advanced economies, the two main types of memory chip—DRAM and NAND—are produced by only a couple of firms. For DRAM memory chips, the type of semiconductor that defined Silicon Valley’s clash with Japan in the 1980s, an advanced fab can cost $20 billion. There used to be dozens of DRAM producers, but today there are only three major producers. In the late 1990s, several of Japan’s struggling DRAM producers were consolidated into a single company, called Elpida, which sought to compete with Idaho’s Micron and with Korea’s Samsung and SK Hynix. By the end of the 2000s, these four companies controlled around 85 percent of the market. Yet Elpida struggled to survive and in 2013 was bought by Micron. Unlike Samsung and Hynix, which produce most of their DRAM in South Korea, Micron’s long string of acquisitions left it with DRAM fabs in Japan, Taiwan, and Singapore...

market for NAND, the other main type of memory chip, is also Asia-centric. Samsung, the biggest player, supplies 35 percent of the market, with the rest produced by Korea’s Hynix, Japan’s Kioxia, and two American firms—Micron and Western Digital. The Korean firms produce chips almost exclusively in Korea or China, but only a portion of Micron and Western Digital’s NAND production is in the U.S., with other production in Singapore and Japan. As with DR...

exception of Intel, many key American logic chipmakers have given up their fabs and outsourced manufacturing.

Since the late 1980s, there’s been explosive growth in the number of fabless chip firms, which design semiconductors in-house but outsource their manufacturing, commonly relying on TSMC for this service.

The company that eventually came to dominate the market for graphics chips, Nvidia, had its humble beginnings not in a trendy Palo Alto coffeehouse but in a Denny’s in a rough part of San Jose.

Nvidia not only designed chips called graphics processor units (GPUs) capable of handling 3D graphics, it also devised a software ecosystem around them.

graphics requires use of programs called shaders, which tell all the pixels in an image how they should be portrayed in, say, a given shade of light. The shader is applied to each of the pixels in an image, a relatively straightforward calculation conducted over many thousands of pixels. Nvidia’s GPUs can render images quickly because, unlike Intel’s microprocessors or other general-purpose CPUs, they’re structured to conduct lots of simple calculations—like shading pixels—simultaneously.

2006, realizing that high-speed parallel computations could be used for purposes besides computer graphics, Nvidia released CUDA, software that lets GPUs be programmed in a standard programming...

By making the chips useful beyond the graphics industry, Nvidia discovered a vast new market for parallel processing, from computational chemistry to weather forecasting. At the time, Huang could only dimly perceive the potential growth in what would become the biggest use case for parallel processing: artificial intelligence. Today Nvidia’s chips, largely manufactured by TSMC, are found in most advanced data centers. It’s a good thing the company didn’t need to build its own fab. At the startup stage, it would probably have been impossible to raise the necessary sums.

Phone companies were trying to agree on a technology standard that would let their phones communicate with one other. Most companies wanted a system called “time-division multiple access,” whereby data from multiple phone calls would be transmitted on the same radio-wave frequency, with data from one call slotted into the radio-wave spectrum when there was a moment of silence in a different call.

It’s easy to lament the offshoring of semiconductor manufacturing. But companies like Qualcomm might not have survived if they’d had to invest billions of dollars each year building fabs.

Field-programmable gate arrays, chips that can be programmed for different uses, were pioneered by companies like Xilinx and Altera, both of which relied on outsourced manufacturing from their earliest days. The biggest change, however, wasn’t simply new types of chips. By making possible mobile phones, advanced graphics, and parallel processing, fabless firms enabled entirely new types of computing.

Bet-the-house gambles were replaced by calculated risk management.

The old model of integrating design and manufacture would struggle to compete when the rest of the industry was coalescing around TSMC. TSMC’s position at the center of the semiconductor universe required it to have capacity to produce chips for all its biggest customers.

Chang thought the real risk was accepting the status quo.

Jobs didn’t have time to get all his ideas into the hardware of the first-generation iPhone, which used Apple’s own iOS operating system but outsourced design and production of its chips to Samsung. The revolutionary new phone had many other chips, too: an Intel memory chip, an audio processor designed by Wolfson, a modem to connect with the cell network produced by Germany’s Infineon, a Bluetooth chip designed by CSR, and a signal amplifier from Skyworks, among others. All were designed by other companies.

A year after launching the iPhone, Apple bought a small Silicon Valley chip design firm called PA Semi that had expertise in energy-efficient processing. Soon Apple began hiring some of the industry’s best chip designers. Two years later, the company announced it had designed its own application processor, the A4, which it used in the new iPad and the iPhone 4. Designing chips as complex as the processors that run smartphones is expensive, which is why most low- and midrange smartphone companies buy off-the-shelf chips from companies like Qualcomm. However, Apple has invested heavily in R&D and chip design facilities in Bavaria and Israel as well as Silicon Valley, where engineers design its newest chips.

This investment in specialized silicon explains why Apple’s products work so smoothly. Within four years of the iPhone’s launch, Apple was making over 60 percent of all the world’s profits from smartphone sales, crushing rivals like Nokia and BlackBerry and leaving East Asian smartphone makers to compete in the low-margin market for cheap phones.

By 2010, at the time Apple launched its first chip, there were just a handful of cutting-edge foundries: Taiwan’s TSMC, South Korea’s Samsung, and—perhaps—GlobalFoundries, depending on whether it could succeed in winning market share. Intel, still the world’s leader at shrinking transistors, remained focused on building its own chips for PCs and servers rather than processors for other companies’ phones. Chinese foundries like SMIC were trying to catch up but remained years behind. Because of this, the smartphone supply chain looks very different from the one associated with PCs. Smartphones and PCs are both assembled largely in China with high-value components mostly designed in the U.S., Europe, Japan, or Korea. For PCs, most processors come from Intel and are produced at one of the company’s fabs in the U.S., Ireland, or Israel. Smartphones are different. They’re stuffed full of chips, not only the main processor (which Apple designs itself), but modem and radio frequency chips for connecting with cellular networks, chips for WiFi and Bluetooth connections, an image sensor for the camera, at least two memory chips, chips that sense motion (so your phone knows when you turn it horizontal), as well as semiconductors that manage the battery, the audio, and wireless charging. These chips make up most of the bill of materials needed to build a smartphone.

The concept remained much the same as Jay Lathrop’s upside-down microscope: create a pattern of light waves by using a “mask” to block some of the light, then project the light onto photoresist chemicals applied to a silicon wafer. The light reacts with photoresists, making it possible to deposit material or etch it away in perfectly formed shapes, producing a working chip. Lathrop had used simple visible light and off-the-shelf photoresists produced by Kodak. Using more complex lenses and chemicals, it eventually became possible to print shapes as small as a couple hundred nanometers on silicon wafers. The wavelength of visible light is itself several hundred nanometers, depending on the color, so it eventually faced limits as transistors were made ever smaller. The industry later moved to different types of ultraviolet light with wavelengths of 248 and 193 nanometers. These wavelengths could carve shapes more precise than visible light, but they, too, had limits, so the industry placed its hope on extreme ultraviolet light with a wavelength of 13.5 nanometers.

Since the 1980s, Intel has specialized in a type of chip called a CPU, a central processing unit, of which a microprocessor in a PC is one example. These are the chips that serve as the “brain” in a computer or data center. They are general-purpose workhorses, equally capable of

opening a web browser or running Microsoft Excel. They can conduct many different types of calculations, which makes them versatile, but they do these calculations serially, one after another.

(To train a computer to recognize a cat, you have to show it a lot of cats and dogs so it learns to distinguish between the two. The more animals your algorithm requires, the more transistors you need.)

Chips optimized for AI can work faster, take up less data center space, and use less power than general-purpose Intel CPUs.

GPUs were designed to work differently from standard Intel or AMD CPUs, which are infinitely flexible but run all their calculations one after the other. GPUs, by contrast,

are designed to run multiple iterations of the same calculation at once. This type of “parallel processing,” it soon became clear, had uses beyond controlling pixels of images in computer games. It could also train AI systems efficiently. Where a CPU would feed an algorithm many pieces of data, one after the other, a GPU could process multiple pieces of data simultaneously. To learn to recognize images of cats, a CPU would process pixel after pixel, while a GPU could “look” at many pixels at once. So the time needed to train a computer to recognize cats decreased dramatically.

Amazon, Microsoft, Facebook, Tencent, Alibaba, and others—have also begun designing their own chips, specialized to their processing needs, with a focus on artificial intelligence and machine learning.

Google has designed its own chips called Tensor processing units (TPUs), which are optimized for use with Google’s TensorFlow software library. You can rent the use of Google’s simplest TPU in its Iowa data center for $3,000 per month, but prices for more powerful TPUs can reach over $100,000 monthly.

All of China’s most important technology rests on a fragile foundation of imported silicon.

The Made in China 2025 plan didn’t advocate economic integration but the opposite. It called for slashing China’s dependence on imported chips. The primary target of the Made in China 2025 plan is to reduce the share of foreign chips used in China.

China’s import of chips—$260 billion in 2017, the year of Xi’s Davos debut—was far larger than Saudi Arabia’s export of oil or Germany’s export of cars. China spends more money buying chips each year than the entire global trade in aircraft. No product is more central to international trade than semiconductors.

If China’s drive for self-sufficiency in semiconductors succeeded, its neighbors, most of whom had export-dependent economies, would suffer even more. Integrated circuits made up 15 percent of South Korea’s exports in 2017; 17 percent of Singapore’s; 19 percent of Malaysia’s; 21 percent of the Philippines’; and 36 percent of Taiwan’s. Made in China 2025 called all this into question. At stake was the world’s most dense network of supply chains and trade flows, the electronics industries that had undergirded Asia’s economic growth and political stability over the past half century.

Governments often have plans that fail abjectly. China’s track record in spurring production of cutting-edge chips was far from impressive. Yet the tools China could bring to bear—vast government subsidies, state-backed theft of trade secrets, and the ability to use access to the world’s second-largest consumer market to force foreign firms to follow its writ—gave Beijing unparalleled power to shape the future of the chip industry. If any country could pull off s...

Whatever the architecture, China had virtually no domestic capability to produce competitive data center chips. China’s government set out to acquire this technology, strong-arming U.S. companies and pressuring them to transfer technology to Chinese partners.

IBM’s decision to trade technology for market access made business sense. The firm’s technology was seen as second-rate, and without Beijing’s imprimatur it was unlikely to reverse its post-Snowden market shrinkage. IBM was simultaneously trying to shift its global business from selling hardware to selling services, so sharing access to its chip designs seemed logical.

CFIUS, the U.S. government committee that reviews foreign purchases of American assets. AMD took the transaction to the relevant authorities in the Commerce Department, who don’t “know anything about microprocessors, or semiconductors, or China,” as one industry insider put it. Intel reportedly warned the government about the deal, implying that it harmed U.S. interests and that it would threaten Intel’s business. Yet the government lacked a straightforward way to stop it, so the deal was ultimately waved through, sparking anger in Congress and in the Pentagon.

The Chinese market was so enticing that companies found it nearly impossible to avoid transferring technology. Some companies were even induced to transfer control of their entire China subsidiaries. In 2018, Arm, the British company that designs the chip architecture, spun out its China division, selling 51 percent of Arm China to a group of investors, while retaining the other 49 percent itself. Two years earlier, Arm had been purchased by Softbank, a Japanese company that has invested billions in Chinese tech startups. Softbank was therefore dependent on favorable Chinese regulatory treatment for the success of its investments. It faced scrutiny from U.S. regulators, who worried that its exposure to China made it vulnerable to political pressure from Beijing. Softbank had purchased Arm in 2016 for $40 billion, but it sold a 51 percent stake in the China division—which according to Softbank accounted for a fifth of Arm’s global sales—for only $775 million.

Chip firms simply can’t ignore the world’s largest market for semiconductors. Chipmakers jealously guard their critical technologies, of course. But almost every chip firm has non-core technology, in subsectors that they don’t lead, that they’d be happy to share for a price. When companies are losing market share or in need of financing, moreover, they don’t have the luxury of focusing on the long term. This gives China powerful levers to induce foreign chip firms to transfer technology, open production facilities, or license intellectual property, even when foreign companies realize they’re helping develop competitors. For chip firms, its often easier to raise funds in China than on Wall Street. Accepting Chinese capital can be an implicit requirement for doing business in the country.

Zhao’s real interest was in buying the island’s crown jewels—MediaTek, the leading chip designer outside the U.S., and TSMC, the foundry on which almost all the world’s fabless chip firms rely. He floated the idea of buying a 25 percent stake in TSMC and advocated merging MediaTek with Tsinghua Unigroup’s chip design businesses. Neither transaction was legal under Taiwan’s existing foreign investment rules, but when Zhao returned from Taiwan he took the stage at a public conference in Beijing and suggested China should ban imports of Taiwanese chips if Taipei didn’t change these restrictions.

The problem wasn’t simply that Chinese government-linked funds were buying up foreign chip firms. They were doing so in ways that violated laws about market manipulation and insider trading. While Canyon Bridge was maneuvering to purchase Lattice Semiconductor, for example, one of Canyon Bridge’s cofounders tipped off a colleague in Beijing, passing along details about the transaction via WeChat and at meetings in a Starbucks in Beijing. His colleague bought stock based on this knowledge; the Canyon Bridge executive was convicted of insider trading.

Tsinghua Unigroup’s activities were impossible to comprehend from the perspective of business logic. There were too many Chinese state-owned and state-financed “private equity” firms circling the world’s semiconductor companies to describe this as anything other than a government-led effort to seize foreign chip firms. “Call forth the assault,” Xi Jinping had demanded. Zhao, Tsinghua Unigroup, and other government-backed “investment” vehicles were simply following these publicly announced instructions. Amid this frenzied dealmaking, Tsinghua Unigroup announced in 2017 that it had received new “investment”: around $15 billion from the China Development Bank and $7 billion from the Integrated Circuit Industry Investment Fund—both owned and controlled by the Chinese state.

it’s more helpful to compare Huawei’s trajectory to a different tech-focused conglomerate, South Korea’s Samsung. Ren was born a generation after Samsung’s Lee Byung-Chul, but the two moguls have a similar operating model. Lee built Samsung from a trader of dried fish into a tech company churning out some of the world’s most advanced processor and memory chips by relying on three strategies. First, assiduously cultivate political relationships to garner favorable regulation and cheap capital. Second, identify products pioneered in the West and Japan and learn to build them at equivalent quality and lower cost. Third, globalize relentlessly, not only to seek new customers but also to learn by competing with the world’s best companies. Executing these strategies made Samsung one of the world’s biggest companies, achieving revenues equivalent to 10 percent of South Korea’s entire GDP. Could a Chinese firm execute a similar set of strategies?

For all the country’s export prowess, China’s internet firms make almost all their money inside of China’s domestic market, where they’re protected by regulation and censorship. Tencent, Alibaba, Pinduoduo, and Meituan would be minnows were it not for their home market dominance. When Chinese tech firms have gone abroad, they’ve often struggled to compete.

Huawei’s critics often allege that its success rests on a foundation of stolen intellectual property, though this is only partly true. The company has admitted to some prior intellectual property violations and has been accused of far more. In 2003, for example, Huawei acknowledged that 2 percent of the code in one of its routers was copied directly from Cisco, an American competitor. Canadian newspapers, meanwhile, have reported that the country’s spy agencies believe there was a Chinese-government-backed campaign of hacking and espionage against Canadian telecom giant Nortel in the 2000s, which allegedly benefitted Huawei. Theft of intellectual property may well have benefitted the company, but it can’t explain its success. No quantity of intellectual property or trade secrets is enough to build a business as big as Huawei. The company has developed efficient manufacturing processes that have driven down costs and built products that customers see as high-quality. Huawei’s spending on R&D, meanwhile, is world leading. The company spends several times more on R&D than other Chinese tech firms. Its roughly $15 billion annual R&D budget is paralleled by only a handful of firms, including tech companies like Google and Amazon, pharmaceutical companies like Merck, and carmakers like Daimler or Volkswagen. Even when weighing Huawei’s track record of intellectual property theft, the company’s multibillion-dollar R&D spending suggests a fundamentally different ethos than the “co...

The amount of space available in the relevant part of the radio-wave spectrum is limited. There are only so many radio-wave frequencies, many of which aren’t optimal for sending lots of data or transmitting over long distances. Telecom firms have therefore relied on semiconductors to pack ever more data into existing spectrum space.

at Analog Devices, which specializes in semiconductors that manage radio transmission.

Partly, 5G networks will send more data by using a new, empty radio frequency spectrum that was previously considered impractical to fill. Advanced semiconductors make it possible not only to pack more 1s and 0s into a given frequency of radio waves, but also to send radio waves farther and target them with unprecedented accuracy.

Tesla is also a leading chip designer. The company hired star semiconductor designers like Jim Keller to build a chip specialized for its automated driving needs, which is fabricated using leading-edge technology.