Opening: The Reckoning in Veldhoven
July 2024. Veldhoven, North Brabant, the southern Netherlands.
In this town of 45,000 people sits the company that builds the most expensive machines on Earth. ASML. The full name: Advanced Semiconductor Materials Lithography. A single machine costs $350 million. And there is only one company on the planet that can make it.
The day ASML's CEO Christophe Fouquet announced Q2 2024 results, one number unsettled investors. China revenue share: 49%. Nearly half of ASML's total revenue came from China. Before semiconductor export controls took full effect, Chinese customers had stockpiled every piece of equipment they could still buy. That year, DUV (Deep Ultraviolet) systems sold to China accounted for 70% of the total.
By 2025, the annual share had dropped to 33%. The 2026 forecast: 20%.
A 29-percentage-point evaporation in two years. The kind of trajectory where you can almost hear the impact each time a $350 million machine sale gets canceled. Fouquet addressed the investor conference with composure. "We operate our business within geopolitical realities." His words were neither a warning nor a protest. They were a statement of fact. ASML had not sold a single EUV (Extreme Ultraviolet) system to China since the first Trump administration in 2019. That was not ASML's choice. The Dutch government had refused to issue export licenses. It was the result of acceding to an American request.
What does this scene imply?
Semiconductors are the oil of the twenty-first century. But there is one decisive difference between oil and semiconductors. Oil comes from many places: Saudi Arabia, Russia, the United States, Canada, the UAE. Advanced semiconductors are different. Design happens mostly in California. Manufacturing happens mostly in Taiwan. And the critical equipment needed for that manufacturing comes from a single company in the Netherlands.
This extreme geographic concentration creates the bottleneck of technological competition. The U.S.-China tech rivalry is unfolding across multiple fronts simultaneously: AI model development, data collection, talent acquisition, regulatory design. But beneath all of these fronts lies a single foundation: semiconductors. This chapter tracks the present state of the war over that foundation.
Four dimensions define this chapter: the implementation status of the CHIPS Act, ASML as a global bottleneck, the evolution of export controls and the economics of smuggling, and the reality of China's semiconductor self-sufficiency. Where these four dimensions intersect, the economics of technological competition reveal themselves.
Section A. The CHIPS Act — Bringing Semiconductors Home
A Reversal 52 Years in the Making
August 9, 2022. President Joe Biden signed the bill into law. The CHIPS and Science Act — CHIPS standing for Creating Helpful Incentives to Produce Semiconductors. A total of $52.7 billion in federal funds would flow into semiconductor manufacturing, research, and workforce development, with an additional 25% investment tax credit.
The historical context of this legislation matters more than the dollar figures.
In 1970, the United States accounted for 37% of global semiconductor production. By 2022, that figure had fallen to 12%. What happened over those 52 years? Production migrated to Asia in pursuit of cost efficiency. Taiwan's TSMC and South Korea's Samsung came to dominate the foundry market. American companies focused on design and outsourced manufacturing to Asia. NVIDIA became the world's most valuable semiconductor company as a fabless operation, designing chips without owning a single factory. This division of labor functioned efficiently for more than two decades.
Then in 2022, two shocks arrived simultaneously. The COVID-19 supply chain crisis exposed the reality of semiconductor shortages. Automobile factories shut down. Electronics production halted. Millions of people worldwide were told their new car would take six months to deliver because there were no chips. And geopolitical tensions escalated around Taiwan. The fact that more than 70% of TSMC's advanced process capacity was concentrated in a single location suddenly registered with new urgency. What had been invisible became obvious.
The CHIPS Act was the product of this cognitive shift. A state intervening directly in the market to restructure its production base, a case of "institutional redesign" in real time.
Three Companies, Three Subsidy Deals
In 2024, the Biden administration finalized subsidy agreements with three companies.
Intel receives $7.86 billion. At Fab 52 in Chandler, Arizona, it will mass-produce its Intel 18A (1.8nm) process — the first sub-2nm advanced node manufactured on American soil. Plans for two additional plants in New Albany, Ohio existed as well, but that project was pushed back to 2030. The direct cause: Intel posted its worst financial results in history in 2024, plunging into a management crisis. A company that once sat on the throne of global semiconductor manufacturing had ceded the lead to TSMC and now depended on its own government's subsidies to regain competitiveness.
TSMC receives $6.6 billion. Its first plant (Fab 1) in Phoenix, Arizona is already operational, producing AI chips on a 4nm process. The second plant's construction is complete, with 2nm/3nm equipment being installed. Mass production is targeted for 2027.
Samsung Electronics receives $4.75 billion. The plant under construction in Taylor, Texas was originally slated to launch at 4nm but has been upgraded to a 2nm GAA (Gate-All-Around) process. Mass production is targeted for late 2026. Tesla has signed a long-term supply contract worth $16.5 billion.
Together, these three trajectories tell a single story. TSMC's global foundry market share rose from 59% in 2023 to 66% in 2025. Samsung holds 12 to 13%. The two companies control 78% of the total. Within this structure, the United States is pulling both TSMC and Samsung onto American territory. The strategy: concentrating dispersed risk under American control.
Total subsidies: $19.2 billion, comparable to the annual capital expenditure of South Korea's entire semiconductor industry. That this amount is more than a simple industrial subsidy becomes clear from the terms of the agreements. Recipient companies are barred from building or expanding advanced semiconductor plants in China for ten years. The moment they accept the funds, the terrain of the technology competition shifts.
Culture Clash in the Arizona Desert
The TSMC Phoenix construction site was troubled from the start.
After groundbreaking in 2021, friction accumulated on site. Hundreds of engineers had come over from Taiwan. And there were American construction workers.
The Taiwanese engineers were accustomed to TSMC's culture: zero tolerance for schedule delays, weekends and night shifts taken for granted, problems resolved with immediate on-site decisions. The American construction site operated under a different logic. Labor laws applied. Overtime required premium pay. Safety regulations were not negotiable. The two teams looked at each other with expressions of mutual incomprehension. The "institutional rigidity" described in Volume 1 took on physical form at a construction site.
American media often reduced this friction to "a work ethic problem among U.S. workers." But that was the surface. The deeper issue was the transplantability of an ecosystem. TSMC was not relocating a single factory. It was attempting to transplant an entire organism — the supply chain, workforce training systems, subcontractor networks, and unspoken work culture that had formed naturally across Taiwan over decades. The wafer suppliers in Hsinchu. The chemical plants in Taoyuan. Thousands of subcontractors spread across the island. Replicating all of this in the Phoenix desert was not a matter of money.
The result: mass production at the first plant fell one to two years behind schedule. The cost of building semiconductor plants in the United States runs three to five times higher than in Asia (construction labor costs, regulatory compliance costs, and supply chain reconstruction costs compounding on top of one another). The $52.7 billion in CHIPS Act subsidies was designed to close this cost gap. But subsidies reduce costs. They do not create ecosystems.
Ecosystems cannot be bought. Ecosystems are built by time.
This is the hardest part of the CHIPS Act. The policy has been enacted. The funds have been deployed. Semiconductor manufacturing ecosystems function only on decades of accumulation, and accumulation does not materialize on command. Taiwan's cultivation of its semiconductor industry beginning in the 1970s was no accident. Compressing those fifty years into a short window is what the CHIPS Act actually demands.
The formula from Volume 1 repeats here. Technological innovation creates capital concentration (the Asian foundry monopoly). A supply chain crisis triggers social awakening. The state redesigns its institutions (the CHIPS Act). But this time, institutional redesign does not complete itself within a single nation. It functions only atop a multinational supply chain that includes Taiwan, the Netherlands, Japan, and South Korea.
Section B. ASML — The World's Most Critical Bottleneck
100,000 Parts
Building a single ASML EUV system requires more than 100,000 components. Transporting them takes 40 shipping containers, the cargo capacity of three Boeing 747 freighters. Installation and calibration take over a year.
What this machine does is use light to etch circuits onto silicon wafers. Lithography. The shorter the wavelength, the finer the circuits that can be etched. EUV uses extreme ultraviolet light at a wavelength of 13.5 nanometers. To generate this light, a laser fires at tiny droplets of tin 50,000 times per second, creating plasma. The light emitted from that plasma is focused by mirrors onto the wafer. Those mirrors must be flat to the level of the Earth's surface — with tolerances measured in single-digit nanometers. Imagine shrinking the Earth to the size of the Netherlands: the height difference between the tallest mountain and the deepest valley would need to be a few millimeters. That is the level of precision required.
The inputs that went into creating this single technology: decades of optical expertise at the Dutch headquarters. Ultra-precision mirror fabrication from Germany's Carl Zeiss SMT. Specialized multilayer coating technology from Japan's AGC. Laser light source technology from California's Cymer (now an ASML subsidiary). All of these integrate inside a single machine. Remove any one component and the EUV system does not function.
What would China need to develop EUV independently? It would have to reproduce every one of these technologies on its own. SMEE (Shanghai Micro Electronics Equipment), a Shanghai-based semiconductor equipment company, is working on it. But the assessment among industry analysts is unanimous: it is not physically impossible, but it is a project requiring ten to fifteen years and hundreds of billions of dollars.
What is lost without EUV? Sub-5nm processes are effectively blocked. The most advanced nodes that TSMC and Samsung currently mass-produce are 3nm and 2nm. AI accelerator chips (NVIDIA's H100, H200, and Blackwell series) are all manufactured on sub-5nm processes. Without EUV, these chips cannot be made. This is the source of ASML's power.
The Ownership Problem of the Bottleneck
One point demands attention. ASML is not an American company.
ASML is a Dutch company. Its shareholder structure shows U.S. institutional investors as the largest ownership group, but the corporate domicile is Veldhoven and its management is European. And ASML's technology belongs to no single nation. It is the crystallization of decades of collaboration among Dutch, German, Japanese, and American companies.
When the United States asked the Netherlands in 2019 not to grant ASML export licenses for EUV sales to China, the Dutch government complied. Additional restrictions on DUV equipment were subsequently imposed in phases. But this decision was not unilaterally enforced by the United States. It was the result of diplomatic negotiation with the Dutch government. For ASML, this meant billions of dollars in annual revenue losses. Persuading the Netherlands to absorb those losses was a matter of alliance management.
Rome maintained its alliance system even after Cannae. It had integrated the Italian allies into a network of military mobilization. The United States maintaining its technology alliance with the Netherlands, Germany, and Japan — the countries that make ASML possible — is the foundation that enables semiconductor dominance on every other front. ASML is not "an American weapon." It is "a bottleneck created by multinational technological cooperation." Sustained control of that bottleneck requires sustained alliance management.
What happens when alliances waver has already been demonstrated. In 2025, China tightened its rare earth export controls in retaliation for semiconductor equipment restrictions. Warnings emerged that European semiconductor equipment companies could be affected. Rare earths are raw materials for specialized components used in ASML's machines. When geopolitical pressure strikes multiple nodes of a technology supply chain simultaneously, the cohesion of the alliance network frays.
Between Stockpiling and Collapse
ASML's China revenue share for 2024 was 36.1% on an annual basis. But in individual quarters, it reached as high as 49%. Consider what this number means in concrete terms.
That year, Chinese semiconductor companies paid ASML more than a third of its total revenue. Why did they buy so much? Because they were trying to secure every available piece of equipment before regulations tightened further. DUV equipment in particular. EUV was already off the table. But using DUV multi-patterning techniques to achieve circuits approaching sub-5nm was not impossible. Disadvantageous, but not impossible.
So Chinese companies stacked DUV systems. This was not a matter of warehousing equipment. They built new fabs, expanded production lines, and purchased machines not yet installed for future use. Securing assets that would vanish from the market once controls tightened. A peculiar investment logic born of the U.S.-China technology war.
Then in 2025, as strengthened export controls began partially covering DUV systems, ASML's China revenue share fell to approximately 33%, with projections pointing to 20% by 2026. ASML's 2030 revenue forecast stands at $71 billion. What offsets the decline in Chinese revenue is the AI boom's demand for advanced semiconductors. The expansion of the global AI accelerator market is driving up EUV demand. As TSMC, Samsung, and Intel begin operating EUV equipment at their new plants in Arizona and Texas, that demand will grow further.
The world's most critical bottleneck is now having its value tested with maximum intensity.
The "dematerialization of leverage" described in Volume 1 operates here as well. Leverage that once migrated from land to factories, and from factories to machines, has now become the light itself — at a wavelength of 13.5 nanometers. The fact that only one place on Earth can produce that light demonstrates the extreme concentration of leverage.
Section C. The Economics of Export Controls — Sanctions, Smuggling, and Paradox
The Evolution of the Rules
October 7, 2022. There are industry insiders who remember the date the Biden administration announced its semiconductor export controls. They call it "10/7."
The core of the controls that took effect that day was blocking exports of NVIDIA's A100 and H100 GPUs to China. High-performance graphics processing units used for AI training. Without these chips, training large-scale AI models becomes extremely difficult. It was the moment semiconductor export controls expanded beyond conventional military technology restrictions into the realm of AI capability control.
What followed unfolded rapidly.
When China shifted from the A100 to the H800 as a workaround, the H800 was banned too, in October 2023. High Bandwidth Memory (HBM) was added to the restricted list in December 2024. Then in January 2025, on the final stretch of its term, the Biden administration issued the so-called "AI Diffusion Rule." It was a complex framework that classified countries into three tiers, differentiating access to AI chips above certain performance thresholds — in effect, further blocking sales of the H100 and H200 to China. But this rule was withdrawn after just five months. In May 2025, the Trump administration scrapped Biden's AI Diffusion Rule and announced a new export control framework. The direction of regulation itself had been reversed with a change of administration.
With each new regulation, China found alternatives below the restricted performance thresholds, and the United States responded with additional controls targeting those alternatives. Regulation and circumvention, in an endless loop. This is a pattern that appears with near-universal consistency in technology wars. When Britain banned the emigration of textile machine technicians in the 1800s, the United States smuggled British engineers across the Atlantic. The history of technology control is a history of regulation and circumvention without end.
Then in December 2025, a reversal. The Trump administration approved NVIDIA H200 exports to China. The H200 is an upgraded version of the H100. The chip Biden banned, Trump permitted.
Two pressures lay behind this retreat. First, NVIDIA's lobbying. NVIDIA's China revenue share is 13%. The percentage looks small. But given NVIDIA's annual revenue of $187 billion, that amounts to $24 billion, by no means a trivial sum in absolute terms. Second, the paradox of Huawei's rise. If the H200 could not be sold, that demand might flow to the Ascend 910C. The question became: is it strategically sound to hand Huawei the entire Chinese AI chip market?
This logic exposes the fundamental dilemma of semiconductor export controls. The stronger the controls, the stronger China's motivation to achieve self-sufficiency. The weaker the controls, the more advanced technology leaks through. Finding the optimal point between these two poles is extraordinarily difficult. And the fact that the optimal point shifts every time the U.S. administration changes is the single greatest structural vulnerability of the export control strategy.
The Smuggling Trial in South Texas
In 2025, an indictment was unsealed at the U.S. Attorney's Office for the Southern District of Texas.
The defendants were charged with operating a $160 million smuggling network for NVIDIA H100 and H200 chips between October 2024 and May 2025. The chips did not travel directly to China. They were routed through Singapore, Malaysia, and Hong Kong. On paper, the transactions looked like normal business-to-business deals. Shipping documents listed different goods.
The defendants' profiles revealed the structural characteristics of the smuggling network. There was an AI startup executive. There were trade intermediaries. There was a logistics company employee. What they shared was a single fact: AI laboratories and companies in mainland China wanted H100/H200 chips, and a route existed to supply them. Where demand meets supply, a market forms, even when that market is illegal.
This prosecution exposed the enforcement reality of export controls. Writing a law and enforcing a law are entirely different undertakings. A single NVIDIA H100 fits in the palm of a hand. Hidden at the bottom of a shipping container, it stands a chance of passing through customs inspection. Physically small, extraordinarily high in value per unit. From the customs enforcement perspective, this is the hardest category of contraband to intercept.
Some estimates suggest that 50% of the world's AI developers are based in China. If true, demand for American AI accelerator chips is structurally elevated in China. Demand finds a path. When legal channels are blocked, it creates illegal ones. This is the consistent law of embargo history. The contraband of the twenty-first century is no longer crude oil or weapons. It is silicon chips.
The Paradox of Innovation Under Scarcity
The most significant reaction to export controls was not smuggling. It was DeepSeek.
In January 2025, Hangzhou-based AI startup DeepSeek released its R1 model. Its foundation model, V3, was pre-trained on 2,048 H800 GPUs. R1 underwent reinforcement learning on top of that with 512 H800 GPUs. The H800 was a China-specific model that NVIDIA had created with reduced performance after the U.S. banned A100/H100 exports in 2022. But the H800 itself was banned in October 2023. DeepSeek trained on the H800 units it already possessed — wringing every drop of performance from chips it could no longer purchase.
The resulting model matched or, on some benchmarks, exceeded GPT-4. V3 pre-training cost $5.58 million. R1 reinforcement learning cost $290,000. Combined, the total was a fraction — on the order of one-fiftieth — of what comparable OpenAI models cost to train. NVIDIA's stock price briefly plunged 17% after the release. A question rippled through Silicon Valley: was the multi-billion-dollar investment in GPU infrastructure truly necessary?
How should this be interpreted?
Does it mean export controls failed? That is hard to conclude. The controls prevented China from acquiring H100 and H200 chips in bulk. Without that constraint, DeepSeek would have used more GPUs, and what the result would have been is unknowable. A more powerful model might have arrived even sooner.
But it does not mean the controls succeeded, either. The constraint redirected the vector of innovation. Instead of throwing more hardware at the problem, DeepSeek developed algorithms that achieved the same results with less hardware. This is "innovation under scarcity." MoE (Mixture of Experts) architecture, quantization techniques, inference optimization. These directions existed before DeepSeek, but DeepSeek pushed them to their limits because hardware constraints made it a necessity.
This paradox was something the architects of the sanctions had not anticipated. The intent was to slow China's AI development. Instead, a more efficient AI development methodology was born. And that methodology was released as open source and spread worldwide. American AI startups and European researchers began adopting it. The byproduct of sanctions paradoxically accelerated global AI progress.
Export controls demonstrably have an effect. China remains structurally disadvantaged in advanced sub-5nm processes. As long as EUV access is blocked, this constraint will not be resolved. But controls are not a tool that "ends the game." They are a tool that changes the rules of the game. And when the rules change, strategies change. Export control strategy must always operate on this premise. Institutions are slower than technology, a pattern that export controls have not escaped. By the time a new regulation is written, technology has already moved one step ahead — and when administrations change, the direction of regulation itself reverses.
Section D. China's Semiconductor Self-Sufficiency — Reality and Limits
The Gap Between Targets and Reality
The semiconductor self-sufficiency target that the Chinese government set in its 2015 Made in China 2025 initiative was 70%. The year 2025 has arrived. Counting all production within China's borders yields close to 50%, but the share of chips independently designed and manufactured by Chinese companies stands at just 19 to 23% (IC Insights/TechInsights).
A 20-percentage-point gap. On the surface, the numbers suggest progress. Disaggregate them and a different story emerges.
Where does the 50% self-sufficiency come from? Memory chips and legacy-node (28nm and above) logic chips have seen rising domestic production. Smartphone components, consumer electronics semiconductors, some automotive chips. In these segments, Chinese companies have genuinely gained ground. As China's share of the global legacy semiconductor market has grown, a paradox has emerged: Western semiconductor companies are the ones feeling the pressure.
But in advanced logic semiconductors (AI accelerators, high-performance CPUs, latest-generation telecommunications chips) the picture is different. And the self-sufficiency rate for the equipment used to make semiconductors is far lower still. As of 2024: 13.6%.
This is the true bottleneck.
Manufacturing semiconductor chips requires dozens of types of equipment: lithography systems (ASML), chemical vapor deposition tools, etching equipment, ion implantation systems, and chemical-mechanical planarization tools. The global companies that make this equipment are based in the United States (Applied Materials, Lam Research, KLA), the Netherlands (ASML), and Japan (Tokyo Electron, Screen). If these machines can no longer be imported, semiconductor production itself becomes impossible once existing equipment reaches the end of its service life. The stockpiled DUV systems will eventually reach end of life too.
This is why China is pouring enormous resources into raising its equipment self-sufficiency rate. Achieving self-sufficiency in chip design and achieving self-sufficiency in equipment are challenges of entirely different orders of difficulty. Design can be accomplished with talent and software. Equipment is the accumulation of decades of manufacturing know-how and precision engineering.
The 40% in Shenzhen's Research Complex
At Huawei's research complex in Shenzhen, the Ascend 910C is being produced.
The Ascend 910C is an AI accelerator that Huawei manufactures in collaboration with SMIC (Semiconductor Manufacturing International Corporation), China's largest foundry. But SMIC operates without EUV, using DUV multi-patterning techniques to achieve a process approaching 7nm. This is the process on which the Ascend 910C is built.
The yield of this process is 40%.
A 40% yield means the following: if 100 chip dies can be stamped from a single wafer, only 40 function correctly. Sixty are discarded. At TSMC's advanced processes, yields run between 80 and 90%. From the same wafer area, TSMC uses 80 to 90 dies; SMIC uses 40. The efficiency gap is more than twofold.
Those 60 discarded dies — this is how the cost of sanctions takes physical form. Silicon wafers are expensive. The lithography process is energy-intensive. The materials, energy, and time invested in those 60 discarded dies are all losses. These costs are reflected in the final product price. The Ascend 910C is estimated to deliver 70 to 80% of the NVIDIA A100's performance, but its unit production cost is substantially higher.
SMIC is attempting to overcome this limitation through scale. Advanced-node wafer production is planned to rise from 45,000 per month in 2025 to 60,000 in 2026 and 80,000 in 2027. The Ascend 910C production target likewise climbs from 600,000 units in 2025 to 1.6 million dies in 2026. The strategy: expand volume to drive down unit costs.
What do these numbers say? That China can sustain its AI competition with "good enough" chips. Not perfect, but not surrendering. As DeepSeek demonstrated, algorithmic efficiency can compensate for hardware gaps to a meaningful degree. If Ascend 910C's 70 to 80% performance is supplemented through software optimization, Chinese AI companies can continue to train and run inference.
Then there is the wall called CUDA.
NVIDIA released CUDA (Compute Unified Device Architecture) in 2007, a software platform that enables GPUs to be used for AI computation. Over the eighteen years since, AI researchers and engineers worldwide have written code on CUDA. Libraries accumulated. Optimization techniques were refined. A developer ecosystem took shape. In a sense, this ecosystem is a deeper moat than ASML's EUV. Hardware is difficult to replicate but theoretically possible. A software ecosystem encapsulates eighteen years of developer habits and knowledge, making it paradoxically harder to duplicate.
Huawei's MindSpore and Baidu's PaddlePaddle exist as Chinese AI frameworks. But their share of the global AI developer ecosystem is negligible. This is the real limitation that extends beyond the Ascend 910C's hardware. The fact that a chip can be manufactured does not mean the software ecosystem to use that chip materializes alongside it.
Energy: The Hidden Variable
To complete the story of semiconductor self-sufficiency, one more variable must be examined: energy.
AI model training consumes enormous amounts of electricity. Data center power costs are a significant component of AI computation expenses. China's data center electricity rates are less than half those of the United States. In 2024, China added 430 GW of net new power capacity, more than fourteen times the 30 GW the United States added in the same year. One-third of total U.S. installed capacity (1,250 GW) was added in China in a single year.
By 2030, China is projected to have approximately 400 GW of surplus power capacity. AI data centers consume electricity voraciously. China's incremental data center power consumption is projected to reach 175 TWh by 2030. Because per-unit electricity prices are far lower than in the United States, the real burden is smaller still.
Where chip performance falls short, energy cost advantages can partially compensate. Even without the NVIDIA H100, more Ascend 910C units can be clustered together and run longer at lower cost. This is the asymmetric strategy through which China achieves a degree of balance against America's semiconductor advantage with its energy advantage.
This asymmetry demonstrates that the chip war is not simply a question of "who makes the better chip." The variables that determine AI computational capacity include chip performance, chip quantity, energy cost, software optimization, and talent — all of them. The structure in which the United States leads in chips while China leads in energy remains open as to which direction it will converge over time.
Through the lens of Volume 1, the United States holds the advantage at the design layer — chip architecture, software ecosystem — while China holds the advantage at the execution layer — applications, scale, cost. Which way this asymmetry tilts depends on how quickly each nation's institutions compensate for their respective weaknesses.
Connecting to Volume 1: From Steel to Semiconductors
The hegemonic transition from Britain to the United States had a decisive signal: steel. When American steel production overtook Britain's in the 1880s, it was not merely a change in an industrial metric. It was a shift in the base of productivity — a transfer of economic gravity.
Are we now reading the same signal in semiconductors?
Two structural differences exist.
First, steel could be measured by output. More, cheaper, faster — where these three metrics converged, that was where hegemony was migrating. Semiconductors are measured not by output but by process generation. China producing more semiconductors does not confer an advantage. Self-sufficiency in advanced sub-5nm processes is what matters. By this standard, China has not yet arrived, and as long as EUV access remains blocked, reaching it in the near term will be difficult.
Second, steel was a commodity a single nation could produce on its own. Iron ore, coal, and labor were sufficient. Semiconductors are different. Without a globally distributed supply chain, advanced semiconductors cannot be produced. Design in the United States. Equipment from the Netherlands and Japan. Materials from South Korea and Japan. Manufacturing in Taiwan. This network is too complex for any single country to internalize. As the CHIPS Act demonstrates, even the United States requires enormous cost and time to relocate a portion of this network onto its own territory.
Unlike steel, semiconductors resist hegemonic measurement by any single metric. This is the structural reason the chip war is more complex and longer-lasting. If hegemony shifted when Britain was overtaken in steel production by Germany, a hegemonic transition in semiconductors is possible only through simultaneous shifts across multiple dimensions. What form that transition will take is being written at this very moment — simultaneously in a factory in the Arizona desert and a research complex in Shenzhen.
Transition: The Individuals Behind the Great Collision
The chip war appears to be a war of numbers: $7.86 billion, 36%, 13.6%, 40%. ASML's earnings report. The reality of semiconductor self-sufficiency rates. The timeline of export controls. A ledger of nation versus nation.
But these numbers do not live inside ledgers.
There is a Taiwanese engineer working overtime at the TSMC Phoenix plant. The children of Hsinchu are handling wafers in the Arizona desert, thousands of kilometers from home. There is a Chinese engineer burning through the night at Huawei's research complex, trying to push the Ascend 910C's yield higher. Getting from 40% to 50% — that is one step toward his country's technological self-reliance.
And there are ordinary people living in the world this war has made. An American paralegal has lost twelve years of legal research expertise to AI. The chips born of semiconductor dominance, the models those chips power — they have taken her work. A Chinese customer service manager watched his thirty-person team shrink to six as an AI customer service system was deployed, then watched his own position disappear. An AI running on domestically produced chips built in defiance of sanctions has taken his work.
The chip war is a macroscopic structure. The path by which that structure reaches an individual life is invisible. But it reaches.
Behind the collision of great empires, two people face another morning.