MICHAEL HOCHBERG And LEONARD HOCHBERG
Every nation-state faces a fundamental choice: With whom will we engage in trade?
What goods and services will we sell, and to whom? What will we buy, and from whom? For authoritarian regimes, where the state subsumes commercial activities, the answer is simple: Only transactions that enhance the power of the state are permitted.
While this principle may be suspended temporarily (for example the New Economic Policy in Russia, or the temporary Chinese market reforms of the Deng-Jiang-Hu era), or may be incompetently implemented, authoritarian regimes exercise full control over trade, in order to enhance state power.
For commercial republics, by contrast, these choices present thorny dilemmas; commercial republics generate massive wealth from free trade. As a result, it is often difficult to differentiate between commercial and national interests; more trade generates more wealth, which in turn generates more power and security.
Except when it doesn’t.
The semiconductor industry, as the book Chip War by Chris Miller, assistant professor of international history at the Fletcher School of Law and Diplomacy, vividly illustrates, is one where free trade has provided enormous benefits to the national security of the United States and its allies.
This played out first during the Cold War, both through the economic and societal impact of the digital revolution and through the many direct impacts on military systems and capabilities. The wide-reaching implications of Moore’s law in the United States provided a crucial technological edge over our adversaries.
The emergence of precision targeting and smart weapons systems offset huge numbers of Soviet weapons with far fewer of America’s own. The drive to increase the complexity and decrease the cost of semiconductor devices involved national specialization, including both chip-building collaborations with Japan, Taiwan and South Korea, and partnerships throughout Southeast Asia and China focusing on packaging and testing these chips and assembling them into finished products.
These relationships strengthened ties among the alliance that the United States forged to contain the Soviet Union.
Although the last part of the book focuses on the emergence of an adversarial relationship with China around chips, a key distinction that is implied and illustrated, but never quite explicitly articulated, is between these international collaborations among friendly nations and the more recent interactions with China.
During the Cold War, semiconductor companies were free to collaborate with companies and nationally sponsored entities across the free world, but were not permitted to interact with the state-sponsored semiconductor development efforts within the Soviet Union. Selling semiconductors and transferring associated technology to China was in the national interest, because – in the aftermath of Richard Nixon’s and Henry Kissinger’s initiative in 1972 – China was part of the alliance against the Soviet Union.
But at the end of the Cold War, American policymakers mistook the benefits of national specialization and decided that anything that increased world trade must be an unalloyed good for the United States.
Even when free trade came at the cost of sending the secrets to building our technological “crown jewels” to China, the justification was that commercial engagement would produce wealth, wealth would result in a larger Chinese middle class, and, inevitably, a Chinese middle class would demand political liberalization and reform. This was wishful thinking of the worst kind.
The Communist Party of China (CPC) took full advantage of American naïveté, capturing a near-monopoly in the lower-value portions of the semiconductor assembly chain – most notably in consumer-scale board and system-level assembly. China has, however, struggled mightily to design and build the chips itself; the chips are the highest-value and highest-complexity components of nearly every modern consumer and military product.
China’s efforts have resulted in onshore capabilities that are now only about two generations (that is, about five years) back from the cutting edge at Intel, Samsung and Taiwan Semiconductor Manufacturing Company; 15 years ago they were much further behind. In 2008, when TSMC was making 40-nanometer transistors, China’s Semiconductor Manufacturing International Corporation (SMIC) – its national champion in this space – was at 300nm. This represents a five-to-six-generation difference.
If you want historical context for how this happened, and to understand the key personalities, decisions and incentives that drove what might seem from the outside to be a set of deeply perverse outcomes, Chip War is a great place to start. The book is, at its core, a history of the geopolitics of semiconductors.
Chip War: the book
Chip War achieves something that few books about the semiconductor industry manage: It provides enough context for those outside the field to understand the nuance and complexity of this important industry without drowning the reader in technical jargon and detail.
The book does this by telling the story of the semiconductor industry, more or less from its inception, through a series of short vignettes. Each vignette addresses a particular person and episode, moving chronologically through the industry’s development.
This is a lot of ground to cover, and it becomes clear that Chris Miller must have selected the most telling and entertaining anecdotes out of perhaps thousands gathered from conversations with leaders of the semiconductor industry.
In Chip War, Miller focuses primarily on the most advanced digital semiconductor chips; as a result there are entire swaths of the industry that are barely mentioned.
It would be possible for someone outside the semiconductor industry to get the impression that, at a geopolitical and geostrategic level, only cutting-edge processors and memory chips matter. This is far from the truth, since a chip without a package, or a packaged chip without a board to carry it, is useless.
And the semiconductor world is much more diverse than just the advanced digital chips on which this volume focuses. It is both deeper vertically, in terms of the tools, intellectual property (IP) and software needed to design, build, package and test these chips, and broader horizontally, in terms of the many other kinds of chips made in silicon and other materials.
Miller emphasizes the “bewildering” and “dizzying” complexity of semiconductor supply chains and international interconnections in the chip industry. And indeed, the products that the chip industry produces are incredibly complex – the most complex products, by many measures, built by humans.
The physics and engineering that go into building microchips is subtle, and requires extremely deep, specialized knowledge. But the industry itself is no more complex than any other large, highly cyclical, fast-growing technical industry. Conflating the complexity of the product with the complexity of the industry is a mistake.
Semiconductor companies are no more complex, as businesses, than any other. Think of any semiconductor industry company as a box that has inputs, outputs, and assets.
The inputs are capital, talent, tools, and the items that go onto the bill of materials, which is the list of input materials and components that the company transforms into their product. The outputs are IP (in the form of patents or licenses), products, trained people and (it is hoped) profits.
And then there are assets. These can be trade secrets, equipment, human capital and – often most important – a corporate culture that is well-adapted to efficiently converting the inputs into the outputs. Pretty much any company can be modeled this way, independent of the complexity of their actual product.
By thinking about the semiconductor industry in these simplified terms, it becomes possible to look at the interactions both among companies, and between companies and governments, in a systematic way.
One of the things that becomes clear from the vignettes in Chip War is that, even when chip companies have tried to remain largely independent of government influence, they’ve found that their interactions with governments – both friendly and hostile – are of existential importance. Anywhere that a government can touch any of the inputs, outputs, or assets of a chip company, that government has an opportunity to gain leverage. And the governments of the world have not been hesitant to use that leverage.
One of the benefits of telling the story of this industry through vignettes is that the importance of individual incentives and individual decisions comes through very clearly. The dynamic relationship between macroeconomic incentives and microeconomic behavior of the firms and leaders plays out again and again through the pages of Chip War.
Nation-states and the chip industry
The trajectory of the semiconductor industry depended in part on the interaction of that industry with various nation-states.
Since the formative days of the semiconductor industry in the United States, one dynamic was the high rate of exponential growth. This growth arose from rapid increases in committed capital, number of people involved, and consumer impact, all running in parallel.
With exponential growth, a very small industry can grow to be a significant chunk of the GDP seemingly overnight. Furthermore, some of the most important decisions revolved around how these firms chose to interact with the US government, and which governments they chose to work with.
Chip War makes it clear that, from the very beginning of this industry, the companies at the leading edge of silicon design and chip fabrication technology were subject to pressure to move their technology offshore (first to Japan, and later to China and elsewhere).
At the same time, the adversaries of the United States were devoting enormous resources to replicating the products of the semiconductor industry.
The Soviet Union was successful in maintaining their semiconductor capabilities at about two to three generations (that is, about five years) behind that of the West for almost the entirety of the Cold War. It did this largely through espionage and appropriation of Western intellectual property, rather than through internal innovation.
Because the Soviets did not have access to Western markets, the flow of information, knowledge, equipment, software, and personnel was dramatically limited. Furthermore, getting chips with billions of transistors to work requires a lot of data and very accurate statistics, and that requires building enormous numbers of chips.
The Soviets chose to be fast-followers and to only use silicon chips where they were absolutely required, rather than making them a pervasive pillar of their economy. As a result, they were never able to catch up.
Getting to massive scale has always been a key advantage in this industry; building more copies of the same chip makes it easier to collect the statistics needed to improve performance and yield, thus enabling higher and higher complexity for each subsequent generation. This dynamic has contributed to a dramatic concentration of the semiconductor fabrication business over time, into just a few giant players.
One of the most dramatic stories the book relates is about how Taiwan rose to become the chip powerhouse that it is today.
The ingredients were simple: First, a brilliant engineer and businessman, Morris Chang, who had spent decades becoming the master of his trade in the United States before being persuaded to move to Taiwan to build up its semiconductor industry. Second, strong backing from the government, across decades, to support all of the needs of what became their most important company, TSMC.
Children in Taiwanese elementary schools today learn what a wafer stepper does and can explain how semiconductor wafers are made. Taiwan has engaged in a whole-of-society effort to own a big, critical piece of the semiconductor ecosystem. This has been wildly successful because, as the author points out, the Taiwanese have engaged very closely with Silicon Valley and the rest of the Western ecosystem.
Taiwan concentrated its efforts on a few very specialized niches where government intervention could be extremely effective, in support of aggressive commercial efforts. TSMC is, in effect, a direct result of business model innovation, relentless technical and commercial execution, and extremely effective industrial policy.
Chips: economic, strategic, and key products
Why is it so difficult for the leading companies to function without government intervention in this industry? In short, because the semiconductor industry is so important. Semiconductors are simultaneously an economic good, a strategic good, and a “key” good.
As an economic good, the chips are subject to market forces. The companies that make and sell them compete with one another on price, reputation, quality, execution, time-to-market and performance. Cutting-edge digital chips, which require the most expensive fabrication processes and design tools, tend to differentiate more on power consumption and performance, while trailing-edge ones tend to differentiate more often on price (though these are broad generalizations).
One thing that Miller conveys very clearly is that semiconductor chips are the digital “brains” behind every single “smart” product in the world. Any device that processes information, from toasters to cellphones, to cars and airplanes, remote controls to modern lightbulbs, depends on semiconductor chips.
If it has a display, or communicates by Wi-Fi or Bluetooth, or even if it just uses electricity, there are almost certainly silicon chips somewhere inside making it work. These chips can be as simple as power regulators or converters, or as complex as cutting-edge digital processors.
As a strategic good, semiconductor chips are the raw material of warfighting. Every modern military system depends on chips to function. Satellites, radar systems, missiles, tanks, planes, Global Positioning System (GPS) receivers, inertial and optical guidance systems – all of it relies on semiconductor chips.
Military radios rely on chips for encoding and decoding signals. The red-dot sights on rifles require LEDs, which are small chips that produce light. Rangefinders require lasers and chips for computation. Logistics are managed through computerized databases, and parts are tracked with barcode scanners and RFID (radio-frequency identification) systems, all of which rely on silicon circuits. Chips are the raw material for nearly every aspect of modern warfighting.
And furthermore, chips define the arena in which gray zone wars are fought. Any form of digital warfare, whether it is direct – that is, hacking an adversary’s infrastructure and systems (such as Stuxnet) – or indirect, via an influence operation through a social media platform.
As artificial-intelligence (AI) systems become more capable, conflicts between adversarial autonomous systems will become more common: The side with the better hardware will have a huge advantage in such conflicts, because their systems will be smarter and faster.
From a security perspective, relying on chips designed or fabricated by an adversary regime is a most uncomfortable position. Setting aside the obvious challenges that emerge from losing access to replacement hardware, there is a second, more insidious problem: Digital chips have become so complex that there is no way to know what they contain.
Unless you control the design and trust the fabrication facility, there is no way to know whether something has been inserted into the chip that creates a hardware vulnerability or a backdoor for an adversary. Detecting such an insertion as an end user is nearly impossible.
Regardless of what software is run on top of untrustworthy hardware, there is no real security. Furthermore, adding software layers to try to create some security on top of untrustworthy hardware will necessarily degrade performance.
In a modern conflict, the default winner is the one who designed and built the hardware and software that their adversary is using. This undoubtedly creates considerable uncertainty and discomfort in China today, given its near-100% dependence on the West (including Taiwan) for advanced semiconductor chips.
And as a “key” good – to deploy the term Halford Mackinder used in his Democratic Ideals and Reality – semiconductor chips “unlock” entire swaths of the economy. Without them, entire sectors of the economy would rapidly grind to a halt.
Cellphones, for example, have an average lifespan of just a few years; they have to be frequently replaced. We’ve recently been able to observe the dramatic consequences of losing access to chips in the automotive industry: For the lack of a few dollars’ worth of chips, it is impossible to complete the manufacture of a US$50,000 or $100,000 vehicle.
The same is true across multiple domains: Without cutting-edge central processing units (CPUs) and graphics processing units (GPUs) from AMD, Intel and Nvidia, the rollout of new services from Facebook, Amazon and other hyperscalers grinds to a halt.
Even in cases like Apple and Google, which design many of their own chips, access to the chip fabs at either TSMC or Samsung are essential to their businesses. Without comparatively inexpensive chips that might cost tens or hundreds of dollars, airplanes cannot fly, trains cannot move, and satellites cannot be launched.
The chips are a “key” good in each of these industries. Other parts can be replaced, or at least built in multiple places – the chips can, in many cases, only be built at one factory.
And very often that factory is a TSMC fabrication facility, located in Taiwan. Right now, TSMC is the default option for fabricating cutting-edge digital chips, because it is the best foundry in the world.
As a foundry, TSMC’s business model is to provide the capability to build extremely complex chips to other companies. With a design team and a pile of money, it is possible for any small company to design a custom silicon chip, and TSMC will make it and deliver it.
It is more informative to think of TSMC as a company that delivers the world’s most advanced processes for building semiconductors in volume, at scale, than it is to think of them as a company that builds chips.
TSMC’s core mission is to bring a new, more advanced fabrication process online every couple of years, defining the state of the art for building advanced digital chips. These fabs are operated as foundries, as a service for the world’s chip design companies – these design companies are known as “fabless” semiconductor companies.
Samsung competes, especially for giant customers, but TSMC gets the lion’s share of the business. Intel has historically been an “integrated device manufacturer,” running chip fabrication facilities solely to build its own products.
Intel is currently retooling its business to act as a foundry, but the transition from fab to foundry is likely to be extremely challenging. A foundry, while it uses very similar facilities to an integrated device manufacturer, is a fundamentally different business. Making the jump into a foundry business has historically been a long and painful process.
TSMC, Samsung and Intel make cutting-edge digital chips. These are the chips that lead Moore’s law, getting ever denser and more complex, with billions of transistors in a single chip today. Such chips require fabrication plants that can produce devices that are only a few nanometers wide.
Such fabs are hugely expensive – tens of billions of dollars for a single factory, and far more for the full R&D effort to bring up a process.
While most of Chip War is focused on the design and manufacturing ecosystem for the world’s most complex, advanced, digital semiconductors and memory chips, this is far from the only part of the industry that produces key, strategic goods of incredibly economic importance. Another area that deserves attention is the trailing edge.
Many other technologies, often lumped into “more than Moore” as a catch-all term, are in fact implemented in processes that don’t need few-nanometer linewidths or involve billions of transistors.
These chips are made in semiconductor fabs, and participate in the same ecosystem of tools, software, test equipment, chemicals, etc. Entire industries rely on these devices – light-emitting diodes (LEDs), lasers, microelectromechanical systems for position sensing, optical detectors, fluidics for health care and diagnostics, silicon photonics, high-power analog, high-speed and high-power-radio-frequency integrated circuits, radiation-hardened electronics and many others – that are made solely or primarily in trailing-edge foundries.
These foundries compete not by making linewidths narrower, but through process innovation, integrating novel materials, and making devices other than transistors.
These activities often advance faster than Moore’s law in their own critical performance metrics, generating goods that are simultaneously economic, strategic and “key,” but they are built in foundries that are orders of magnitude less expensive than those at the cutting edge of digital transistor fabrication.
They don’t need the ultra-fine linewidths of cutting-edge CMOS (complementary metal-oxide-semiconductor) electronics or memory circuits, making the fabs much more affordable.
These “trailing edge” fabs are far more important than the name implies: While their equipment may be trailing-edge for making digital transistors, these fabs build cutting-edge products for applications beyond digital, and their monumental importance is widely under-appreciated, especially in policy circles.
Intel’s chief executive officer, Patrick Gelsinger, has recently been explicit about the geo-strategic implications of semiconductor manufacturing, suggesting, “Where the oil reserves are defined geopolitics for the last five decades. Where the [chip] fabs are for the next five decades is more important.”
Chokepoints in the chip ecosystem
Policy discussions about semiconductor issues often neglect to incorporate the fact that chips are not interchangeable. From an end-user perspective, a CPU chip from AMD or Apple provides the same functionality as one from Intel; they all support a software ecosystem that allows the end user to run Word or Gmail on their computer.
But these processors are not interchangeable; far from it. A computer that uses a chip from AMD requires different boards and supporting chips than the one for an Intel chip. Only at the system level can these chips be used interchangeably, and that system-level interchangeability is only true for a small subset of the systems that rely on chips.
This same dynamic is true across the ecosystem. While Cadence and Synopsys and Mentor Graphics all make software that does very similar things, these tools are not easily interchangeable. A chip design team that has made a multi-year investment in a Cadence-based design flow would lose years of effort transitioning to one from Synopsys, or vice versa.
The same is true for IP blocks, which are completed, tested, field-proven circuits that can be included in a chip design as a pre-completed block: Switching a circuit design from one IP block to another is often a nearly complete redesign.
For instance, ARM and RISC-V architectures both have commercially licensable IP blocks for CPUs that can be included in a chip design. While both blocks perform fundamentally the same function, replacing one with the other requires enormously disruptive changes to both hardware design and to the software and firmware that live higher in the engineering stack. Switching a product from one to the other could take multiple years and cost hundreds of millions of dollars.
Because of these dynamics, there are a lot of chokepoints in the semiconductor ecosystem, where, if a specific good or service is removed or disrupted, there is an amplifying cascade of downstream effects.
For instance, Miller writes in detail about how the entire advanced semiconductor ecosystem is dependent on the multi-hundred-million-dollar EUV (extreme ultraviolet) lithography systems from ASML, which exercises a monopoly on these tools.
This industry is rife with extremely sticky suppliers, where the cost in time and effort to switch from one to another is measured in years. Any disruption to the supply of machines, spare parts, software, design IP, masks, wafers, etc from a single supplier can disrupt an entire swath of the industry, jeopardizing the viability of multiple companies.
It gets worse. Even when there are multiple sources for standard commodity semiconductor chips, with a standard physical and logical interface, it is not always possible to use the parts interchangeably without a considerable engineering and qualification effort.
Because the specifications for chips are complex, even nominally identical products are often not perfect replacements for one another; the only way to know is to build systems with both sources of parts, and put them through a full reliability and qualification cycle. Doing so can easily take more than a year, and that’s if the tests get passed the first time, and no design changes are required to accommodate the alternative source of chips.
Large companies that are producing a product in volume will go to the trouble to qualify multiple sources for critical parts (when they can) in order to make their supply chains more robust. But this is utterly impractical for smaller companies, and it’s an enormous logistical and technical burden to qualify multiple suppliers for more than a tiny fraction of the chips that go into any complex system.
Another strategy to increase supply-chain robustness is to hold a buffer stock of critical components. But this is extremely expensive, and can have a dramatic impact on the economic competitiveness of the products. Fat supply chains are especially expensive in an industry where Moore’s law makes cutting-edge chips, and the systems built with them, obsolete every couple of years.
Ironically, the behavior required by the key and strategic nature of semiconductor products drives the industry toward developing robust, buffered supply chains, while the imperatives associated with economic goods drive the industry toward lean manufacturing and just-in-time delivery.
During times when geopolitical drama is absent, the companies have to run lean to survive. But when national interests come into conflict, lean supply chains collapse, and national survival demands fat, redundant supply chains. This represents a fundamental policy challenge.
This is especially true when we take into account the fact that economic survival is in and of itself a strategic good. A country whose economy collapses because of lack of a key good is in no shape to successfully prosecute a war. We’re seeing the beginning of this dynamic in Russia, as it loses access to Western chips and other high-tech goods.
Corporate behavior is shaped not only by supply chains, but also by demand, talent and capital chains.
One thing that comes through both in Miller’s book, and in decades of personal experience in this industry, is that autocratic governments have a rich history of seeking to insert themselves into these chains in any way that they can – especially at chokepoints – in order to exert leverage, exfiltrate IP, stand up their own chip design and manufacturing capabilities, and shape corporate behavior to their ends.
No comments:
Post a Comment