This exclusive Cogs of War interview is with Rakesh Kumar, the author of Chip Age: How Semiconductors Became the World’s Most Powerful Resource. Rakesh has worked as a co-founder, professor, researcher, and advisor in the chip industry for more than two decades. Chipmaking is an evergreen technology with ever-increasing geopolitical implications, and so we asked him to share his thoughts on his book and the field’s future.
You’ve worked with and around the chip industry for decades. If you could clear up one misconception about chips, what would it be?
It is worth clarifying that chip wars, chip shortages, tariffs, and export controls on chips are not new phenomena; they are rooted in fundamental industry realities. Ever since chips were invented, a scramble began among nations to access state-of-the-art chips because of their ability to power the key strategic technologies of the day. This included intercontinental and ballistic missiles and space missions in the 60s, fighter aircraft in the 70s, computers in the 80s, and GPS in the 90s. Nations also wanted to build their own chips to profit from this lucrative industry, secure their supply, and reduce import bills.
Chips being used for trade and geopolitical leverage are also not new. The United States famously used a combination of trade and geopolitical interventions in the mid-1980s to kill Japanese chip competition and trigger a decline of the Japanese chip industry. America’s successful denial of chip technology to the Soviet Union crippled its ability to develop modern precision-guided weapons and manage a competitive economy, costing it the Cold War.
Similarly, chip shortages have occurred periodically. Because chip demand is cyclical, it is inherently risky for manufacturers to invest significant capital to increase supply. They tend to be conservative in how they grow supply, causing shortages during high demand. The chip supply chain is also optimized for efficiency, so any disruption — natural disaster, conflict, or trade tension — can cause shortages. There were shortages in 2020–2023 (the Covid-19 pandemic), 2011 (Japan earthquake and tsunami), 2000 (poor demand forecasting by Intel), 1998 (American restrictions on Japanese memory chips), and 1994–1995 (personal computer boom), to name a few.
The details have merely changed today (e.g., U.S.-Chinese competition, AI, Trump’s tariffs). And while the stakes are different, they’re not necessarily higher.
Roughly two-thirds of chips today are manufactured using tech commercialized before 2005. What are the advantages and disadvantages of phasing these legacy chips out? How quickly should it be done?
Chip manufacturers such as TSMC want to phase out legacy chips to reallocate resources to advanced chip manufacturing, which is much more profitable. Legacy chips are built on smaller wafers that yield fewer dies per production cycle, making their production cost inefficient. Cheaper Chinese legacy chips are squeezing profits further.
However, replacement of legacy chips with advanced chips won’t be easy. Higher cost of advanced chip manufacturing will be passed down the supply chain, making consumer goods, cars, and appliances more expensive. Replacement will require hardware and software re-engineering, and recertification for highly regulated industries such as aerospace, medical devices, and automotive, further increasing costs. Replacement may not even be possible for several aerospace and defense applications, since legacy chips withstand harsh conditions much better than advanced chips.
We will likely see foundries such as TSMC and Samsung exit legacy chip manufacturing within the next two to three years. Existing legacy chip manufacturing nations, such as Singapore, will fill the resulting vacuum. New entrants such as India will also attempt to compete but will likely struggle with scale and cost efficiency in the short- to medium-term due to a late start.
You argue that there is an innovation problem in the semiconductor industry, with gains in efficiency slowing. This slowdown, you argue, is driven in part by high costs of experimentation and slow trials. In the U.S. and allied countries, who should be investing in this research and development, and where should it be targeted?
While it sounds heretical to point out innovation challenges for an industry that packs billions of microscopic switches onto a thumbnail-sized device, here are some concerning numbers: a microprocessor’s speed now increases only 10–15 percent each year (software speeds up even less); the computational power of an AI chip fall behind AI models by an additional 200 times every two years; and modern-day GPUs consume hundreds of millions of times more energy than the brain for an equivalent amount of computation. No obvious replacement technology is in sight that can mitigate the performance and energy challenges of today’s transistors.
Leading chip companies mostly focus on incremental (though still hard to achieve) progress due to time-to-market and cost pressures, and to minimize risks, while startups and universities find it hard to raise patient capital required to experiment with high-risk, high-reward chip ideas.
The United States should establish three to four time-bound, outcome-oriented industry-government consortia, each focused on a different innovation target — such as alternative computing paradigms, novel devices, novel materials, and breakthrough fabrication technologies. Consortia could maintain a shared infrastructure for testing and prototyping. Allies should be encouraged to participate, especially in the consortia where they have existing strengths, with appropriate safeguards on intellectual property. After all, extreme ultraviolet lithography technology came out in 2001 of a similar consortium set up in 1997.
The United States should require every federal investment chip beneficiary to participate in the consortia. Some Chips Act set programs (that have since been restructured) were set up to develop open-source tools and intellectual properties to reduce the cost and time of chip experimentation. Participation in such programs should become a precondition for federal chip funding.
When reading about how cloning inadvertently killed the Soviet chip industry, I was reminded of current debates over distillation. Are there any lessons from semiconductor history that you see repeating themselves in the AI race?
Several patterns from chip history are repeating in the AI race — namely, “pick and shovel” winners, the importance of software, consolidation of industry, and geopolitical weaponization.
During the personal computer boom in the 1980s and 90s, Intel, which provided the ultimate “pick and shovel” of the age, microprocessors, was the biggest winner, not the personal computer makers. NVIDIA, which provides graphics processing units — the pick and shovel of the AI age — similarly captures a large fraction of the AI industry’s profits today, while AI model and application companies compete bitterly for leftover revenue and dominance. This is likely going to be true in the future, where “pick and shovel” chip companies will be joined by companies delivering power and cooling, networking infrastructure, and chipmaking equipment.
Second, software, not hardware details, proved to be the ultimate moat for chips from the 1980s through the early 2010s. Many chip companies competed against Intel, but developers refused to rewrite their code for newer chips. The same story is playing out in the AI race, where developers have been reluctant to switch away from NVIDIA’s Compute Unified Device Architecture platform. As AI gets better at writing and porting code, however, this moat will weaken.
Third, the highly specialized, iterative, and capital-intensive nature of advanced chipmaking made it untenable for most chip manufacturers to compete, ultimately leaving only two or three advanced chip manufacturers; for the same reasons, only a small number of firms can now afford to build advanced AI models. The list may get further consolidated in the future if the AI scaling law holds.
Fourth, chips became a geopolitical lever during the Cold War, U.S.-Japanese competition in the 1980s, and more recently in U.S.-Chinese relations, leading to trade and technology restrictions and government subsidies. AI is triggering similar maneuvers to achieve and maintain technological superiority and to deny technological parity to geopolitical rivals. However, just as China is now a key player in the production of all but the most advanced chips despite restrictions, we will see China (and eventually Europe, India, South Korea, Singapore, etc., since they would prefer sovereign AI models) achieve near parity in all but the most advanced AI capabilities.
Corrupted, counterfeit, and faulty chips in military equipment have long posed a problem for the Defense Department — in fact, one almost started a nuclear war in 1980 — but you don’t seem optimistic about the Pentagon’s current methodology for identifying them. Where would you encourage supply chain experts to focus their efforts?
The key challenge is that the Pentagon accounts for a tiny fraction of the overall chip demand, which makes it hard for it to exercise meaningful control over the chip supply chain. Even though the Defense Department mandates its contractors and subcontractors to procure chips only from authorized sources, chip supply chains are complex and highly distributed, providing malicious actors and counterfeiters numerous entry points to meddle with the integrity and authenticity of a chip. While the federal government can use programs like the CHIPS Act to onshore some manufacturing (and other parts of the supply chain) and perhaps even have dedicated defense foundries to improve trust, high capital requirements for chipmaking mean that any onshore supply chain will need to be shared with commercial customers to manage costs, which creates new trust challenges. Also, any such onshoring will likely be partial — chipmaking equipment and input materials may still come from outside — leaving parts of the overall supply chain vulnerable. Furthermore, any dedicated defense foundry may be sustainable only for legacy chips, not for advanced chips, due to prohibitive costs.
Another challenge is that procurement mandates, such as enforcing authorized-only sourcing and maintaining an unbroken chain of custody, work well only during peacetime. These mandates are quickly abandoned for procurement speed during conflicts and for access security during chip shortages. Ironically, supply chains need to be most trustworthy during conflicts and shortages.
Therefore, onshore manufacturing and authorized-only sourcing efforts should be accompanied by multi-layered physical and electrical verification of these chips before they are used. This is akin to passing passengers through security checkpoints before they are allowed to board the plane. A cocktail of verification techniques, such as non-destructive imaging, advanced electrical and decapsulation testing, and analog and AI fingerprinting, can detect counterfeits and malicious chips with a high degree of accuracy and without needing to destroy many chips. Verification may increase the procurement-to-use time, so some sampling may be needed (again, think airport screening). Some malicious and counterfeit design modifications remain dormant during standard verification and are triggered during use under specific conditions. Online verification techniques will need to be built into the system architecture to detect these vulnerabilities.
In addition, critical defense chips should be required to have built-in security features that allow these chips to be authenticated physically or remotely. A combination of built-in hardware features and software techniques, such as blockchain, should be used to make these chips traceable through the supply chain.
You detail many semiconductor chokepoints familiar to our readers: lithography machines, critical mineral vulnerabilities, and single-company dominance of manufacturing, among others. If you were king for a day and given unlimited resources to apply to one of these problems, where would you begin your efforts?
Chipmaking is arguably mankind’s most complex engineering exercise, with supply chains spanning over 40 to 50 countries. Parts of the supply chain are dominated by one or a few countries, creating chokepoints. Not all chokepoints are created equal, however, and details matter greatly here. Some chokepoints can be addressed with time and money, while others currently look insurmountable.
Critical materials mining is geographically spread out (in Chile, Peru, Indonesia, the Democratic Republic of Congo, and more). Critical material availability is even more widespread. While the United States has large critical mineral deposits, processing and refining capacity is concentrated in China due to its early strategic investment and state support to achieve the necessary scale.
It will be hard for the United States to compete directly against China in critical minerals due to high costs and stringent environmental regulations. It should address its vulnerability primarily through friendshoring, including by directly strengthening the critical mineral industry of its allies to improve their competitiveness. It needs to also diversify its procurement while primarily sourcing critical materials from allies. Any domestic investment should primarily focus on material substitution, alternative extraction methods, and battery recycling.
As China found out over the last decade, national will and abundant resources will not be enough to build advanced lithography machines at home. ASML in the Netherlands took three decades and tens, if not hundreds, of billions of dollars to master the engineering to build an extreme ultraviolet lithography machine and build the extraordinarily complex and globally distributed supply chain behind these machines. This will be nearly impossible to replicate, with no guarantee of catching up (since ASML would have moved to the next generation by then). Instead of attempting to replicate ASML’s extreme ultraviolet lithography technology from scratch, the United States should focus on next-generation lithography alternatives, such as nanoimprint lithography and x-ray lithography, that have the potential to leapfrog ASML’s extreme ultraviolet lithography roadmap. Concurrently, the United States should strengthen its involvement in the current lithography supply chain (through direct investment, for example) in exchange for access security.
Single company dominance of manufacturing can be addressed by a king with sufficient resources. Intel can now manufacture chips as advanced as those from TSMC. However, for Intel’s advanced manufacturing to be economically viable, it needs large customers to share and offset costs, achieve economies of scale and yield improvement, and establish credibility. As a king, I would strongly lean into major U.S. chip companies (NVIDIA, Apple, Qualcomm, etc.) to use Intel for manufacturing and lean into major U.S. tech companies (Amazon, Microsoft, Google, etc.) to buy Intel-manufactured chips, until Intel advanced manufacturing is self-sustaining. I would also direct the government to buy chips exclusively manufactured by Intel. I would also provide direct federal funding and grants to Intel in exchange for equity or chips to accelerate its transition into a manufacturing leader. The same playbook can be used to develop other advanced chip manufacturers.
Rakesh Kumar is a professor and John Bardeen Faculty Scholar in the Electrical and Computer Engineering Department at the University of Illinois Urbana-Champaign and chip expert. He has received multiple awards in recognition of his influence on chip research and the chip industry. Some of the technologies he’s developed now power the chips in your phones, tablets, and computers. He has advised or consulted for chip and computer companies, startups, governments, and intelligence and advisory firms. He has written extensively about chips for a general audience.
Image: Midjourney