Introduction: Blockades Create What They Seek to Prevent

On December 2, 2024, Intel CEO Pat Gelsinger was dismissed by the board of directors (CNBC, 2024-12-02). In the same month, the DeepSeek team utilized 2.788 million GPU hours of H800 compute power to train the V3 model at a cost of approximately $5.57 million, achieving performance comparable to GPT-4 across multiple benchmarks (CSIS Analysis and DeepSeek Technical Report, 2024). Washington faces a dual challenge: the export control network and industrial subsidy system, built at the cost of tens of billions of dollars, exposed a deep logical misalignment in the same month. The only domestic enterprise with the potential for advanced process wafer foundry work is sliding toward a collapse of both financial and technical roadmaps, while competitors cut off from the most advanced hardware by export controls have completed a deep restructuring of underlying algorithms under extreme compute constraints.

What 71% Means

In the second quarter of 2025, TSMC's share of the global wafer foundry market climbed to 71% (WCCFTech, Q2 2025). In the 3nm and 2nm advanced process sectors, this company's market share is near an absolute monopoly of 100%. Behind this figure lies the core contradiction in Washington's semiconductor strategy. Policymakers frequently cite technological sovereignty as the justification for the CHIPS Act. Sovereignty requires exclusive control, but when more than 70% of global chip manufacturing capacity and nearly all the world's most advanced silicon wafers depend on a single company located on a geopolitical fault line, the industrial sovereignty Washington claims is effectively suspended in mid-air.

TSMC plans to invest $52 to $56 billion in capital expenditures in 2026, with its Arizona expansion plan growing to four factories, supported by $6.6 billion in CHIPS Act subsidies and $5 billion in loans (SIA Official Data, 2025). American taxpayers are on a massive scale subsidizing a Taiwanese company to replicate capacity domestically, in hopes that North American supply chains will not instantly paralyze in the event of a physical conflict in the Western Pacific. However, insurance is not the same as ownership. The ultimate control over the EUV lithography machines operating inside the Arizona factories, the core engineering teams, and the process know-how that determines yield remains in Hsinchu.

Purchasing the physical transfer of an ally's capacity through fiscal subsidies masks the reality that domestic manufacturing capability in the U.S. is difficult to restore. What Washington has spent heavily to build is not a self-controlled semiconductor ecosystem, but a leasing system highly dependent on external technical inputs. As long as the most core process iterations and capacity scheduling instructions are still issued by a board of directors across the ocean, this supply chain security built on sand could collapse at any time due to external variables.

The Counter-Effect of Blockades

Scarcity often breeds innovation. The original intent of export control policies was to lock competitors' artificial intelligence industries below the "compute poverty line" by cutting off the supply of high-bandwidth memory compute clusters. However, reality has diverged from this vision. The DeepSeek V3 model reached performance levels that Silicon Valley peers require tens or even hundreds of millions of dollars to achieve, using only $5.57 million in training costs.

The threshold for acquiring physical hardware was forcibly raised, compelling companies to concentrate R&D resources on software architecture and algorithm optimization. Under the constraints of H800 compute limits, video memory optimization, deep adjustments to Mixture-of-Experts (MoE) architectures, and the application of communication masking technologies became inevitable choices to bridge the compute gap. A policy assessment by the Brookings Institution pointed out that restrictions instead accelerated China's breakthroughs in efficient algorithms. The lack of hardware did not stop the evolution of computational models; instead, it eliminated the redundancy of brute-force compute stacking and pushed the underlying code toward extreme simplification and efficiency.

Export controls created compute scarcity; scarcity forced improvements in algorithmic efficiency; and improved algorithmic efficiency, in turn, reduced dependence on absolute compute scale, ultimately weakening the effect of the blockade. Meanwhile, the physical blockade network itself is riddled with holes. By establishing complex matrices of shell companies and cross-border procurement networks, entities like Huawei successfully acquired over 2 million Ascend 910B chips. A CSIS analysis report (2025) tracked at least eight active smuggling networks. The overlap of hardware leakage and algorithmic breakthroughs is dismantling the technological ceiling Washington set through export controls.

The Hollow of Intel

Pat Gelsinger's departure at the end of 2024 marked a phased failure of Intel's foundry strategy. As the only U.S. domestic company that once possessed the most advanced silicon manufacturing capabilities, Intel's trajectory of decline stands in stark contrast to the expansion cycle of U.S. semiconductor industrial policy. Washington's efforts to reverse industrial cycles through administrative means have exposed significant limitations in the face of cold commercial laws.

The ledger of CHIPS Act fund allocation is a grim report on industrial strength. In the list of beneficiaries of the act, Intel received approximately $7.8 billion in direct subsidies, while foreign companies TSMC and Samsung took $6.6 billion and $6.4 billion, respectively (AIP Report, 2025). Policymakers attempted to use funds to compensate for the cost disadvantages of U.S. domestic manufacturing, but capital cannot buy yield. These subsidies cannot conjure a skilled workforce of semiconductor manufacturing engineers out of thin air, nor can they replace the process experience accumulated through countless trials and errors. During the Trump administration, there were even discussions about a policy to exchange CHIPS subsidies for 10% corporate equity, attempting to forcibly bind industrial interests using financial leverage.

The U.S. steel industry of the 1970s provides a similar precedent. Facing price shocks from Japanese and German steel, Washington provided protection for domestic firms through tariff barriers and import quotas. However, the protected steel giants did not use this breathing room to update blast furnace technology; instead, they distributed profits to shareholders and maintained obsolete production lines. When the protection period ended, the competitiveness of the U.S. steel industry had not only failed to recover but had accelerated its decline due to missing the window for technological iteration. Subsidies and protection can delay market elimination, but they cannot replace process experience accumulated through trial and error. The logic of the CHIPS Act is highly similar: administrative means can buy the physical presence of a factory, but they cannot buy the engineering capability that allows the factory to consistently produce yield.

There are rigid boundaries to the intervention of financial means in the laws of physical manufacturing. Intel's repeated delays and yield issues at advanced process nodes prove that semiconductor manufacturing is highly dependent on the non-linear characteristics of long-term engineering accumulation. Once a company falls behind in the rhythm of process iteration, the subsequent cost of catching up rises exponentially. Washington spent tens of billions of dollars in industrial subsidies attempting to rebuild a complete defensive line for advanced semiconductor manufacturing domestically, only to end up consolidating the dominant position of foreign foundry giants in the global supply chain. The policy documents speak of revitalizing American manufacturing, but the capital flows honestly toward the overseas branches of Asian enterprises.

Two Parallel Systems

The spotlight has long been focused on the microscopic carving of 3nm and 2nm advanced processes. Beyond the lens, in the mature process market, the structure of the global semiconductor supply chain is undergoing a violent tectonic reorganization. As of early 2025, China's mature process chip capacity is expected to account for 28% of total global production (Tom's Hardware, 2025-02-26). Western industry circles refer to this trend as the "China Shock" in the semiconductor sector.

Export controls have squeezed massive capital and engineering resources into the unrestricted realm of mature processes. This rapid expansion of capacity has directly altered the pricing logic for global low-to-mid-range chips. Through aggressive price wars, Chinese companies are reshaping the commercial rules of basic chip markets such as power management, microcontrollers, and automotive electronics. Industrial observations from Reuters (2025-02-10) show that mature process companies in Taiwan are facing severe survival pressure. When traditional suppliers cannot compete on the cost curve with emerging capacity, the profit distribution mechanism of the entire industry chain flips accordingly.

A parallel system is taking shape. Under restrictions preventing the acquisition of the most advanced EUV lithography machines, the self-sufficiency rate of equipment in China's semiconductor industry at mature nodes climbed from 4.91% in 2018 to 19.92% in 2023. Controlling nearly 30% of global mature process capacity means establishing deep supply chain control over massive downstream industries such as consumer electronics, industrial control, and new energy vehicles. Even without achieving parity in advanced processes, dominating mature process capacity to cut off the supply of underlying silicon wafers for the global industrial machine constitutes a highly deterrent strategic lever.

How Long Can the Time Gap Last?

The temporal logic of Washington's semiconductor strategy is built on the assumption of gaining a "time gap." The CEO of ASML once publicly assessed that the embargo on EUV lithography machines would leave China 10 to 15 years behind the West in advanced processes. A prediction model from Goldman Sachs (via Korea Times, 2026-02-21) pointed out that China's semiconductor self-sufficiency rate was only about 14% in 2024 and would only reach 37% by 2030. These calculations, based on linear extrapolation, constitute the source of confidence for export control policies.

The moat of the time gap is being eroded by two distinct forces. On the algorithm side, the DeepSeek effect proves that compute disadvantages can be compensated for through architectural innovation. On the manufacturing side, SMIC achieved a breakthrough at the 5nm node in 2025. Data from TrendForce (2025-03-28) shows that because it cannot obtain EUV equipment, SMIC must rely on DUV multi-patterning technology for difficult yield improvements; the manufacturing cost of its 5nm chips is about 50% higher than TSMC's, with a yield only one-third of the latter's.

From a commercial logic perspective, using triple the resources and a high scrap rate to produce silicon wafers of the same generation is an unsustainable war of capital attrition. However, the security ledger has a different algorithm. For missile guidance systems, radar arrays, or military-grade compute clusters, as long as chips meeting performance specifications can be physically produced, commercial losses can be fully absorbed by strategic gains. The effectiveness of export controls depends not only on whether a technological generational gap can be opened, but also on whether this gap can maintain a sufficiently long safety window in the face of an opponent's asymmetric breakthroughs. When deep algorithmic optimization and cost-insensitive manufacturing catch-up occur simultaneously, the sand in the hourglass of the time gap is flowing at a speed that has completely exceeded the expectations of policymakers.

The Transistor Moment

On September 18, 1957, in an office in San Francisco, Arthur Rock took ten one-dollar bills from his pocket and handed them to eight individuals as a contract between them. These eight men had just resigned from the company of Nobel laureate William Shockley, who called them the "Traitorous Eight." Ten years earlier, Shockley had co-invented the transistor at Bell Labs. That world-changing device was extremely crude—a plastic triangle wrapped in gold foil, pressed onto a piece of germanium, with two contact points less than a millimeter apart. Now, the finest engineers he had personally recruited were using a personal investment of $500 each, plus $1.4 million from a camera company, to build the goals he had failed to realize.

December 23, 1947

On December 23, 1947, John Bardeen and Walter Brattain demonstrated the first working solid-state amplifier at Bell Labs (PBS Transistor History, 1947). The invention process was full of serendipity; two laboratory accidents—condensation of water vapor and an accidentally formed oxide layer—became the foundation for the key breakthrough. The birth of the transistor was more an accidental product than the result of a meticulous plan.

As the project supervisor, William Shockley quickly fell into anxiety and jealousy upon realizing he had not personally participated in the specific operation of the device. He immediately attempted to claim the patent rights for the entire invention for himself, only to discover that as early as the 1920s, Julius Lilienfeld had already registered a field-effect patent based on similar principles, thereby blocking his legal path to monopolizing the invention. A power struggle erupted; the issue of technical attribution was fraught with tension from the very birth of the industry.

The Bell Labs Paradox

In 1956, an antitrust consent decree spanning dozens of pages was placed before AT&T executives. The Department of Justice forced this government-regulated monopoly to open 7,820 patents for free, accounting for approximately 1.3% of all active U.S. patents at the time (CEPR/VoxEU, 1956).

The "technical sovereignty" of 1947 was manifested in absolute closed control. As the research arm of AT&T, Bell Labs relied on monopoly profits to incubate the transistor, the laser, the Unix operating system, and the C programming language. The government established technology ownership through the regulation of monopolies; such ownership was indeed effective, but the price of closure was stagnation. Forcibly stripping that ownership instead ignited the industry. Economic research estimates that the 1956 antitrust ruling increased subsequent innovation by about 17%, with more than two-thirds of that increment coming from small businesses and individual inventors (AEA Papers and Proceedings, 2021). Gordon Moore later recalled that it was this lawsuit that allowed the commercial semiconductor industry to truly take off.

The mechanism was not complex: AT&T's patent portfolio covered the core pathways of solid-state electronics, and any company attempting to enter this field had to bypass them or pay. Forcing the opening of patents was equivalent to clearing entry barriers—not through subsidies, but by dismantling fences. Research by scholars such as Watzinger shows that the biggest beneficiaries were not large electronic firms (which already had bargaining power), but small inventors who previously could not enter the field at all. The vitality of the industry came from the periphery, not the center.

The government created an industry through antitrust intervention that forced a "letting go," which is completely different from the logic of today's attempts to lock in advantages through subsidies and control. The most successful industrial policy of the U.S. government was, ironically, an antitrust agreement signed almost haphazardly with the initial intent of dismantling a telecommunications giant.

Shockley's Paranoia

In a laboratory in Palo Alto, California, in 1957, the atmosphere dropped to freezing point. Because a female secretary had been pricked by a pin, William Shockley suspected intentional sabotage and demanded that all employees undergo polygraph tests. The polygraph did not find the saboteur, but it exposed the internal centrifugal forces.

Shockley did not come to California to create an industrial ecosystem; he was there simply because he grew up there, his mother lived nearby, and because of the lobbying of Stanford University's provost. He firmly believed that technology belonged only to a tiny minority of genius individuals and could not tolerate the independent research of his subordinates. This "inventor mindset" had already shown its signs in the patent disputes at Bell Labs, and in his own company, it evolved into disastrous micromanagement. In May 1957, eight core engineers jointly reported management issues to investor Arnold Beckman. After Beckman reneged on his support for them, they resigned en masse in September.

Shockley won the Nobel Prize but lost the company he founded, sowing the seeds of the semiconductor industry in California in the most humiliating way. When these eight people left, each contributed an additional $500 as startup capital. The ten one-dollar bills handed out by Arthur Rock were not just a pact of alliance, but the starting point of venture capital as an institution. Before this, Wall Street had no idea how to price technology companies that had nothing but brains and blueprints. This financing logic—exchanging a small amount of capital for equity in high-risk technology companies—later became one of the most difficult-to-replicate competitive advantages of the American technology industry. China can build wafer fabs, but it cannot use administrative orders to replicate an institutional environment where engineers are willing to bet their careers on $500.

Fairchild's Fission

Fairchild Semiconductor itself never achieved ultimate success in a commercial sense; its market value never exceeded $2.5 billion at its historical peak. Yet by 2014, the 92 listed spin-off companies traceable back to Fairchild had a total market value that had ballooned to approximately $2.1 trillion (Computer History Museum, 2014), a ratio as high as one to eight hundred.

In 1968, Gordon Moore and Robert Noyce left to found Intel, followed by Jerry Sanders establishing AMD, while Eugene Kleiner and Don Valentine built the venture capital landscapes of Kleiner Perkins and Sequoia Capital, respectively. On the 1986 SEMI genealogy chart, 126 companies were densely marked as directly originating from Fairchild. The company was no longer the ultimate reservoir of value but had become a talent incubator. This structure is highly similar to seed plants in biology: the commercial value of the parent itself is far less than the entire rainforest ecosystem propagated by the seeds it scatters to the wind. Fairchild's greatest success was precisely the disintegration of the parent company caused by its founders constantly leaving to start new ventures.

Early military orders served as the starting engine. In 1962, the Apollo program chose Fairchild semiconductor chips for its guidance computers, simply because a single-chip design could ensure reliability in extreme environments (Paul Ceruzzi, 1962). With technology spillovers and the explosion of the civilian market, by the late 1960s, the share of military procurement in the integrated circuit market had plummeted from nearly 100% in the early stages to less than a quarter (Employ America, 1960s). Once the engine was ignited, the government stepped out of the driver's seat.

Lesson One of Technical Sovereignty

From the birth of the transistor in 1947 to the formation of Silicon Valley in the 1980s, the early evolution of technical sovereignty presented a clear trajectory of losing control.

Every leap in the industry was the result of the collapse of a system that attempted to maintain closed control in the previous stage. Bell Labs' monopoly incubated the underlying technology, the 1956 antitrust ruling forcibly released it to the market, Shockley's management paranoia drove the top engineers out of the lab, and the centrifugal forces within Fairchild caused a fission into a complete ecosystem covering manufacturing, design, and capital in just eleven years (1957-1968).

This trajectory has one key characteristic: the government's role in each stage was to "exit" rather than to "deepen." The 1956 antitrust was a forced exit from monopoly control, the Apollo procurement was a gradual exit after starting the market, and the Fairchild fission was completed spontaneously in a situation where the government could not intervene. The source of industrial vitality has always been the part that the government cannot touch.

More than half a century later, auditors in Washington are attempting to rebuild that kind of absolute-control monopoly incubation period. The 2022 CHIPS Act allocated $52.7 billion for semiconductor manufacturing subsidies, while attaching strict guardrail clauses: subsidized companies must not expand advanced process capacity in China, must not export to specific countries, and must not buy back shares. This is a complete combination of "subsidy + control," diametrically opposed to the logic of "forced letting go" in 1956. The lesson of Bell Labs is: the price of closure is stagnation, the price of openness is loss of control, and loss of control is the source of industrial vitality. The transistor was not invented; it was released. When policymakers carefully calculate the flow of every EUV lithography machine and the foundry market share of every advanced process wafer, they are using physical fences to cage an organism that can only survive through diffusion.

Cold War Catalyst

$43.50 a Piece

In February 1962, a NASA procurement officer saw a figure on a quote from Fairchild Semiconductor: $43.50.

He signed it.

At the time, discrete components with equivalent functionality cost only a few cents. What the space agency needed was not low cost, but absolute reliability. The interior of the Apollo Guidance Computer was densely packed with 4,100 such integrated circuits. In the vacuum of deep space, 300,000 kilometers from Earth, the accidental failure of any micron-scale logic gate could bury three astronauts, along with the entire nation's Cold War dignity, in lunar orbit.

The core asset that government orders inject into an industry is never the funds on the balance sheet, but a set of rigorous reliability standards. To meet the high requirements for uptime in military and aerospace systems, Fairchild Semiconductor was forced to drastically improve its production processes, introducing epitaxial processes and raising wafer foundry yields from less than 10% to over 90% (Computer History Museum primary archives, 2010). Behind this figure lies a complete learning curve: every time cumulative production doubles, unit costs drop by approximately 20% to 30%. The improvement in yield was not a stroke of genius from engineers, but was forced out by extremely demanding customers with real orders. When the yield rises from 10% to 90%, it means the number of qualified products produced from the same batch of raw materials increases from 1 to 9, mathematically rewriting the cost structure.

The cost of a single Fairchild chip soon dropped to the $20 to $30 range and fell further to $1.25 in 1971.

The payers completed the baptism of industrialization with near-obsessive requirements. The successful operation of the Apollo program provided a level of system-level credibility endorsement that no commercial marketing at the time could match, prompting the Burroughs Corporation to subsequently place a massive commercial order for 20 million integrated circuits from the semiconductor industry without hesitation (Computer History Museum, 2010).

60% of the Weight

In the mid-1960s, at the peak of the Apollo program, NASA alone consumed 60% of the total integrated circuit output in the United States (confirmed by dual sources: Computer History Museum and FedTech, 2016).

This was an abnormal market driven entirely by national will. In 1965, the Minuteman II Intercontinental Ballistic Missile (ICBM) project took over from Apollo as Silicon Valley's largest single consumer. Whether sending humans to the moon or nuclear warheads to Moscow, the underlying logic of military and aerospace procurement was consistent: size, weight, and power consumption enjoyed absolute priority, while price was almost at the bottom of the decision chain.

By the early 1970s, Transistor-Transistor Logic (TTL) circuits had become the primary source of revenue in the fields of minicomputers and peripherals (Computer History Museum, 2010). The military procurement share plummeted from 60% to single digits.

Washington politicians often interpret this history as a retreat of government support, but such a view is inaccurate. The exit itself was the highest level of success. When a precision manufacturing sector that once relied entirely on blood transfusions from the Pentagon can sustain itself by selling typewriter-sized microcomputers to businesses, it truly possesses the capacity for self-prosperity. The decline in the military share was not a signal of industrial contraction, but proof of industrial maturity—just as a child no longer needing monthly remittances from parents does not mean the family relationship has broken down, but rather that they are finally able to live independently. The departure of the government as the sole super-customer was the final mark of the semiconductor industry's completion of its commercial transformation.

The Paradox of Forced Openness

In 1952, Bell Labs took the initiative to share transistor technology with the outside world, charging a licensing fee of $25,000 to each of 40 companies.

This was merely a mild diffusion of technology. In 1956, an antitrust consent decree from the U.S. Department of Justice mandated that AT&T must license all existing patents to the entire world for free, explicitly including non-telecommunications fields.

Once the walls of secrecy and monopoly are pushed down by administrative power, the force of technological diffusion destroys the original closed ecosystem.

Econometric analysis of that period of history by academia has yielded precise conclusions. According to research published by Watzinger et al. in the American Economic Review (AEA) in 2021, this mandatory free licensing policy catalyzed decades of long-term innovation growth outside the telecommunications sector. Those foundational patents, originally locked in safes, transformed into the original code for countless startups.

Today, Washington's export control policies are replaying a mirror image of that history, as policymakers attempt to lock down the technological paths of adversaries. Faced with export control budgets totaling tens of billions of dollars, the extreme scarcity of computing power has forced Chinese companies to abandon crude brute-force compute models and instead pour all engineering resources into the underlying reconstruction of algorithmic architectures.

A miracle occurred: DeepSeek completed the full training of the V3 large model using only 2.788 million GPU hours of H800 computing power (approximately $5.57 million in training costs, 2024). This technological mutation, forced by external pressure, is precisely the modern echo of the 1956 mandatory diffusion logic. The direction of the causal chain is: export controls → compute scarcity → algorithmic breakthrough → control failure. Every step has verifiable data points, and every step is contrary to the expectations of policymakers; the blockade created exactly what it sought to prevent.

What Subsidies Cannot Buy

The industrial catalysis mechanism in history is clear and conforms to commercial common sense. The government throws out extremely demanding real orders, companies are forced to establish very high process standards to win those orders, process upgrades push up yields, the leap in yield smashes cost barriers, and finally, the broad civilian market is ignited.

The operational logic of the CHIPS and Science Act (CHIPS Act), however, is exactly the opposite.

Washington intervenes on the supply side; policymakers first use high subsidies to forcibly prop up wafer foundries on domestic soil, and then can only passively pray that there will be enough market demand in the future to fill those expensive cleanrooms. According to data from the Center for Strategic and International Studies (CSIS, 2024), Intel received up to $8.5 billion in grants and $11 billion in loan support in this subsidy feast.

The figures on the balance sheet mask the real hemorrhaging. Intel Foundry, with total revenue of $18.9 billion in 2023, delivered a disastrous financial report with losses amounting to $7 billion. The new factories will not generate substantial revenue until 2027 at the earliest. To translate: $8.5 billion in subsidies corresponds to $7 billion in losses, a net benefit of about $1.5 billion. Whether this $1.5 billion can be converted into real technological capability depends on a question that remains unanswered—who will pay the bill?

Fiscal subsidies can easily bridge the funding gap for capital-intensive plant construction, but they can never replace the real market in verifying the survival of a technological route.

As TSMC firmly controls the rhythm of the entire industry chain with a global wafer foundry market share of over 71% (Q2 2025 forecast), domestic capacity expansion in the U.S. looks more like an expensive political performance detached from commercial laws. Intel's predicament proves one thing: the Act has bought only the excess capacity of allies, not the technological sovereignty that Washington claims to defend.

The logic of government procurement in the Apollo era was: first, there was extremely demanding real demand; then, there was the forced upgrade of process capabilities; industrial independence was the natural endpoint of this chain. The logic of the CHIPS Act is: first, throw money at building factories; then, look for demand; then, pray that the process can keep up. The sequence is reversed, and the result is naturally different.

The semiconductor industry's first pot of gold from the Cold War was never a check issued by the White House, but rather the silhouette of that figure who, after signing for a high price of $43.50, decisively exited the stage once the industry was able to walk on its own.

In January 2025, an investigative report by the Financial Times mentioned that Chinese venture capital institutions are pursuing the personal assets of failed founders through the courts (Financial Times, 2025-01-07). "Redemption Right" clauses grant capital providers the legal power to seize real estate and freeze bank accounts. The "dishonest person list" has cut off entrepreneurs' access to high-speed rail and flights. Data subsequently released by China's Ministry of Industry and Information Technology revealed the transmission path of this "chill": between 2022 and 2023, the number of registered technology startups dropped by more than 20% (TechCabal, 2025-01-13). Meanwhile, Washington is calculating the disbursement progress of the $52 billion CHIPS and Science Act (CHIPS Act), attempting to reshape dominance in semiconductor manufacturing through fiscal subsidies. These two seemingly parallel sets of events point to the same issue: the policy blind spot of technological sovereignty—money cannot buy an ecosystem.

Section 1: The Law of 1872

Section 16600 of the California Business and Professions Code has abolished non-compete agreements since 1872 (California Business and Professions Code § 16600). In that year, telephone networks were not even widespread in the continental United States, yet this land had already laid the most fundamental legal framework for the birth of Silicon Valley. This local law grants engineers the freedom to resign at any time and immediately start a competing enterprise, allowing knowledge to flow.

Route 128 in Massachusetts once possessed the top intellectual resources of MIT and military contracts from the Department of Defense of an equivalent scale. However, the local judicial system allowed companies to enforce strict non-compete agreements. Under the deterrent of non-compete litigation lasting several years, East Coast engineers were forced to deeply bind their careers to a single employer, such as Digital Equipment Corporation (DEC), and tacit knowledge was firmly locked within the bureaucratic systems of large tech companies. AnnaLee Saxenian precisely measured the differences in the development trajectories of the two regions in her 1994 regional economics study (AnnaLee Saxenian, 1994). The legal blocking of talent mobility led to the defeat of Route 128 in the era of microcomputers.

Silicon Valley's earliest competitive advantage stemmed from a local law that predated the invention of the transistor by 75 years. Washington's $52 billion act can procure Extreme Ultraviolet (EUV) lithography machines and advanced process capacity, but it cannot replicate the legal immunity in Ohio or Arizona that allows technical personnel to jump ship at any time. When the "Traitorous Eight" of Fairchild Semiconductor took the technical blueprints and trial-and-error experience from their previous company to assemble the next generation of industry standards in a garage a few miles away, they relied precisely on institutional protection from being sued by their former employer.

Section 2: A One-Page Business Plan

In July 1968, the financing document submitted by Gordon Moore and Robert Noyce to financier Arthur Rock was only a page and a half (chiphistory.org). Noyce proposed over the phone that they needed $2.5 million in startup capital. Rock's reply was brief, calling it the most certain investment of his life. This startup, later named Intel, completed its second round of financing within a day and a half.

A year later, in May 1969, Jerry Sanders founded AMD with seven former colleagues from Fairchild Semiconductor. The eight founders scouted around and scraped together about $100,000 in seed capital (dcfmodeling.com). At the time, Innovation Magazine rated the company as the least likely to survive in the field of technology entrepreneurship for the 1968-1969 period. The business plans for the two companies that laid the foundation of the modern semiconductor industry totaled less than two pages.

Venture capital bets on people and tolerates extremely high technical uncertainty and commercial ambiguity; this is the core function of this new financial instrument. Capital providers know that in the early stages when transistor density is growing exponentially, any long-term financial forecast is meaningless.

The accountability logic of Washington's subsidies is completely different. The prerequisite for Intel to receive $8.5 billion in CHIPS Act subsidies is the submission of detailed quarterly capital expenditure plans, excess profit-sharing agreements, supply chain security audit reports, and commitments for childcare quotas in local communities to the U.S. Department of Commerce. Government capital pursues certain capacity indicators and politically correct employment reports. Requiring certainty from a company that is in a period of technical route transition and facing a yield crisis in advanced processes is equivalent to stifling its strategic space for trial and error. What Washington is demanding from Intel is a government procurement contract containing redundant social goals, with no room for the fault tolerance of venture capital.

Section 3: The 75x Trigger

In 1977, total venture capital commitments in the United States were only $68.2 million (NBER Working Paper No. 2832, James Poterba). At that time, Kleiner Perkins' first fund of $8 million raised in 1972 was already respected in the industry as the world's largest venture capital fund (venturevoice.substack.com). Today, that figure is equivalent only to an ordinary early-stage angel investment.

Fine-tuning of policy ignited a nuclear fusion of capital. In 1978, the U.S. Congress passed a bill cutting the capital gains tax from 49.5% to 28%. The following year, the Department of Labor revised the "Prudent Man" rule in the Employee Retirement Income Security Act (ERISA), officially removing the long-standing red line of fiduciary responsibility and allowing massive pension assets to be allocated to high-risk venture capital.

A capital tsunami quickly took shape. In 1978 alone, venture capital commitments jumped to $978.1 million, a 14-fold increase. By 1983, annual commitments had climbed to $5.0977 billion, completing a 75-fold expansion in absolute scale within six years. Sequoia Capital, founded in 1972 by former Fairchild Semiconductor executive Don Valentine, completed the historical process of capitalizing technical alumni networks precisely within the capital wave triggered by tax rates and pension decoupling.

A prosperous ecosystem stems from two precise policy adjustments. Other economies can completely replicate the trigger conditions for capital accumulation by modifying tax laws. However, the due diligence experience, technical evaluation networks, and industry norms of extreme tolerance for failure accumulated over thirty years through countless business bankruptcies cannot be achieved overnight through legislative provisions. The framework of a system can be fast-tracked, but the growth of flesh and blood takes time.

Section 4: The MITI Experiment

The industrial policy of Japan's Ministry of International Trade and Industry (MITI) in the semiconductor field was once seen by Washington as an invincible benchmark of mercantilism. In 1986, Japanese companies held more than 50% of the global semiconductor market, and three zaibatsus—NEC, Hitachi, and Toshiba—accounted for the top three in global revenue. By 2024, the number of Japanese companies in the list of the world's top ten semiconductor enterprises is zero (Gartner, 2024).

The model of concentrating national resources for breakthroughs showed amazing efficiency during the catch-up phase. The Very Large Scale Integration (VLSI) project led by MITI forcibly poured capital and talent into the research and development of Dynamic Random Access Memory (DRAM), destroying the cost advantage of American companies with extremely high yields. When the focus of industrial evolution shifted to design-intensive logic chips and microprocessors, the rigid resource allocation network immediately became paralyzed; the centralized model could not diversify the risk of huge technical divergence in the early stages of innovation.

After the experiment ended, the careers of Japanese engineers were firmly bound within the zaibatsus by the lifetime employment system. Independent venture capital institutions were nowhere to be found, and the lifeblood of innovation relied entirely on internal transfers from commercial banks and parent companies. When Samsung surpassed them in process iteration by relying on the efficient decision-making of a family zaibatsu's dictatorship, the talent flow in the Japanese semiconductor industry showed a pathological one-way loss. Top engineers did not choose to resign in Tokyo to start new chip design companies; South Korean competitors simply poached them to Seoul with salaries several times higher.

The path of Europe has similarly verified this rule. The three European semiconductor giants—Infineon, STMicroelectronics, and NXP—each have achievements in the field of mature processes but are collectively absent from the field of advanced logic chips. The European Chips Act passed by the EU in 2023 promises to invest 43 billion euros, with the goal of increasing Europe's share of the wafer foundry market from less than 10% to 20% by 2030. The prerequisite for achieving this goal is that Europe can establish a venture capital culture and talent mobility mechanism that has never existed within ten years. Laws can be passed within a legislative cycle, but culture requires generations of failure to accumulate.

The CHIPS Act attempts to use fiscal subsidies to replicate the industrial glory of "concentrating resources to accomplish great things." History has already completed the ultimate test of government-led technological innovation in Tokyo Bay. Elpida Memory, as the final crystallization of the Japanese government's forced merger of the remnants of the memory industry, eventually ended in bankruptcy and acquisition by Micron.

Section 5: The Cost of Failure

The ultimate achievement of Silicon Valley's institutional innovation is to make the cost of commercial trial and error approach zero. Engineers from bankrupt companies only need to clear their desks and walk across two blocks to bring the engineering lessons accumulated in their previous failed project into the laboratory of the next startup. Knowledge completes the most efficient market-based diffusion through repeated bankruptcies and reorganizations.

The redemption right clauses commonly found in Chinese venture capital agreements are dismantling this fault-tolerance mechanism. Court summonses and asset freezing orders forcibly transform commercial setbacks at the corporate level into the personal financial destruction of the founder. Data from 2019 shows that per capita venture capital in the United States reached $282, while in China it was only $20 (wuab.org). The gap in total capital masks a fundamental mutation in the nature of the funds. When investment agreements carry unlimited joint and several liability, so-called venture capital has degenerated into usury in the guise of equity.

When the cost of failure escalates from losing a venture investment to the complete social bankruptcy of the founder, the underlying infrastructure of the innovation network is uprooted. The more than 20% slump in technology startup registrations between 2022 and 2023 is merely the preliminary physical manifestation of this harsh recovery system. The cost of failure is priced by the founder's social bankruptcy.

The rational choice for engineers is to stay within large tech giants to maintain a stable salary or to completely withdraw from high-risk explorations at the technological frontier. As policymakers in Washington watch the judicial practices across the ocean that actively lock the valves of talent mobility, they should perhaps re-evaluate the focus of their own technological sovereignty strategies. In the $52 billion subsidy list, one cannot buy the right to be free from fear.

Price Umbrella

On April 17, 1987, the Reagan administration imposed 100% retaliatory tariffs on Japanese exports valued at approximately $300 million. That same year, Morris Chang founded TSMC in Hsinchu. Although these two events occurred in the same year, not a single line in Washington's policy documents mentioned the profound significance of this coincidence. The policy text of the Reagan tariffs was filled with calculations to force Japanese chips to maintain high prices to save American companies; its actual effect was akin to opening a massive price umbrella over the Pacific. The "U.S.-Japan Semiconductor Agreement" signed in 1986 mandated so-called "Fair Market Values" (FMV). According to audit data from the Heritage Foundation, in 1986, the minimum selling price for Japanese 256K DRAM chips in the United States was anchored at $2.60, while the same chips were sold for only $1.70 in Japan.

The United States won the first semiconductor defense war, but the spoils were pocketed by South Korea and Taiwan. The agreement restricted Japan but imposed no constraints on South Korean companies. While Japanese manufacturers were forced by bilateral treaties to maintain high prices in third-country markets, Samsung—beyond Washington's jurisdiction—swept the European and Asian markets with extremely low costs, completing its early accumulation of capital and market share by undercutting Japanese pricing. Meanwhile, the birth of TSMC precisely hit another blind spot. Japanese vertically integrated device manufacturers (IDMs) struggled under the pressure of the agreement, fiercely resisting the divestment of their manufacturing departments, while Morris Chang promoted a pure-play foundry model that separated design from manufacturing. According to industry retrospective data from SamMobile, Samsung's 33-year dominance in the DRAM market began around 1992, which was exactly the sixth year that Japan was bound by Washington's price shackles. Washington precisely struck its number one competitor through a bilateral agreement, but inadvertently gave rise to two even more formidable industrial giants.

Who Paid for the Agreement

The global semiconductor procurement list of 1990 was full of absurdities. According to statistics from the Heritage Foundation, the price of a 1Mb DRAM chip in Europe that year was $3.90, while it was as high as $5.00 in the United States. When the capacity was upgraded to 4Mb, the procurement price in Europe was $18, while the U.S. price soared to $30. The original intention of the agreement was to protect Silicon Valley's dwindling chip manufacturing lines, but the actual flow of funds forced downstream American personal computer manufacturers to pay for the artificially inflated profits of Japanese semiconductor companies, even indirectly subsidizing the next-generation process R&D of their competitors.

Japanese companies dumped products at low prices domestically and harvested American customers at compliant high prices overseas. European and Asian computer manufacturers, being unrestricted by the agreement, gained a significant hardware cost advantage. American hardware giants like IBM and Compaq were forced to pay 30% to 40% more for core components than their European counterparts, their profit margins squeezed dry by Washington's protective umbrella. The price of protecting the upstream was the bloodletting of downstream enterprises. This backlash mechanism of managed trade has reappeared like a ghost in current policy debates regarding restrictions on the export of advanced process chips. Cutting off the supply of high-performance computing power certainly slows the R&D progress of specific targets, but it also forces domestic downstream application developers in the U.S. to bear higher hardware acquisition costs and regulatory friction. Washington's policy loop has precisely punished its own industry.

Three Causes of Death

Washington's political pressure has long been regarded as the primary reason for the collapse of the Japanese semiconductor industry, but historical data does not support a simple victim narrative. According to a joint audit by the Center for Strategic and International Studies (CSIS) and the Heritage Foundation, in 1988, two years after the agreement was signed, Japan still accounted for 51% of global semiconductor sales. Archives from the Semiconductor History Museum of Japan (SHMJ) show that in 1987, Japan's overall share of the DRAM market was as high as 80%. The collapse of this vast empire stemmed from the breaking of three internal pillars: the vertically integrated IDM model, the state-led "whole-of-nation" system for concentrated research, and overly stringent quality standards.

When TSMC promoted the pure-play foundry model in 1987, Japanese companies still insisted on a closed-loop system covering everything from design and manufacturing to packaging and testing. As the global semiconductor industry transitioned toward a specialized division of labor in the foundry model, Japanese giants viewed divesting manufacturing departments as an unacceptable regression, missing the optimal window to cooperate with global fabless design companies while clinging to control over the entire industry chain. Bureaucrats in Tokyo's Kasumigaseki district bear direct responsibility for this. The Very Large Scale Integration (VLSI) project launched in 1976 was once seen as a miracle of industrial policy, with the government investing $300 million to pack six core companies into the Kanagawa joint laboratory. This government-led technical coordination did work during the catch-up phase, solving DRAM yield bottlenecks through unified standards. However, when the industry crossed the technical threshold into the innovation stage, the advantage of standardization rapidly turned into a disadvantage. The technical routes of the six companies became highly convergent, and internal differentiated competition was completely eliminated. This situation is similar to the inbreeding effect in biology: a convergent gene pool performs well in a stable environment but exposes fatal vulnerabilities during environmental mutations. When the semiconductor industry transitioned from memory chips to logic chips, the entire Japanese industry lost the ability to pivot flexibly.

Market changes delivered the final blow. Japanese DRAM products were designed for mainframes, with engineers obsessed with pursuing a 25-year lifespan and zero-defect high reliability. In the personal computer market that experienced explosive growth in the 1990s, Silicon Valley hardware assembly plants only needed "good enough" and cheap consumables. Samsung seized the opportunity, sweeping the market with a highly competitive cost advantage; its core competitiveness was not technical leadership, but precisely meeting the actual demand of end customers for quality degradation. Japan's high-end capacity eventually became an expensive exercise in futility. In February 2012, Elpida announced bankruptcy with a massive debt of 448 billion yen, setting a record for the largest bankruptcy in post-war Japanese manufacturing. According to archives from Nippon.com, the company had previously received an injection of 30 billion yen from the Development Bank of Japan and nearly 100 billion yen in loans from commercial banks. The birth of Elpida itself was the result of the Japanese government leading the merger of the DRAM departments of NEC and Hitachi. Two losers retreating in the market were forcibly bound together by administrative order, ultimately failing to reverse the decline. That same year, the American company Micron Technology acquired these expensive experimental remains for the low price of $2.5 billion. Elpida's bankruptcy was not the starting point of the decline of Japanese semiconductors, but the end point of the failure of government-led integration.

China is Not Japan, However...

The script of 1986 easily leads to modern misjudgments. Equating Washington's export controls today with the bilateral agreement of that era ignores the most fundamental geopolitical differences. Japan back then was a military ally dependent on the U.S. nuclear umbrella, lacking the strategic space to refuse Washington's demands. Today's China is a strategic competitor with nuclear deterrence capabilities, holding supply chain leverage over critical minerals and rare earths. Japan's domestic market in 1986 had only 120 million people, while today's China is the world's largest semiconductor consumer market, backed by the massive demand of 1.4 billion people. The General Agreement on Tariffs and Trade (GATT) lacked substantive constraints on bilateral trade bullying, whereas China, as a member of the World Trade Organization (WTO), has already formally initiated multilateral dispute settlement procedures regarding U.S. export control measures. China possesses bargaining chips that Japan could not reach back then.

However, in the underlying logic of industrial structure, Beijing seems to be retracing some of Tokyo's old paths. In an industrial ecosystem that has already been proven by TSMC's foundry model and ASML's EUV lithography supply chain to require extreme specialized global division of labor, attempting to rely on massive state capital injections to maintain or even rebuild a vertically integrated industry chain from design and manufacturing to packaging and testing is equivalent to embedding a piece of code at the system level that has already been proven inefficient by history. The operational logic of the China National Integrated Circuit Industry Investment Fund (the "Big Fund") shows a high degree of similarity to the VLSI project led by the Ministry of International Trade and Industry (MITI) back then: both rely on concentrated resources, both are led by government power, and both pursue technical autonomy across the entire chain. The Kanagawa joint laboratory in 1976 tried to achieve breakthroughs in electron beam lithography and 64K DRAM; today's hundred-billion-level Big Fund is trying to establish security backups at every node from equipment and materials to foundries and advanced packaging.

Japan's industrial history provides a complete test report. The logic of "concentrating power to accomplish big things" is indeed efficient during the technical catch-up phase, relying on administrative power to break through early yield barriers and capacity bottlenecks. However, once technical evolution enters an innovation stage that requires high-frequency trial and error and high differentiation, administrative-led resource allocation will manifest the drawbacks of rigidity and convergence. Whether the Chinese semiconductor industry is still in a catch-up period requiring concentrated power to break through yields, or has already touched the innovation threshold requiring decentralized trial and error, will be the key to evaluating the effectiveness of current industrial policies.

The Taiwan Miracle

In 1987, when TSMC was founded, both Intel and Texas Instruments rejected investment invitations from the Taiwan government. The view of both companies at the time was that a manufacturer that does not produce its own chip products would find it difficult to form a competitive advantage. Thirty-eight years later, in the second quarter of 2025, TSMC's share of the global foundry market reached 71%, with annual revenue approaching $75 billion and gross margins stabilizing above 40%. Meanwhile, Intel's market capitalization fell back to 1990s levels in 2024 and was removed from the Dow Jones Industrial Average. Those two "no's" of 1987 have become one of the most costly errors of judgment in the history of the semiconductor industry.

Morris Chang’s Tuition

Morris Chang spent 25 years at Texas Instruments, exchanging millions of dollars in R&D budgets for a profound realization: the Integrated Device Manufacturer (IDM) model has a structural barrier that is difficult to repair. The problem is not technology, but trust. Fabless design companies, such as Nvidia, have always found it difficult to comfortably hand over their most core GPU designs to a direct competitor like Intel for production, fearing that designs might be "borrowed" or that their production capacity would be deprioritized. The conflict of interest is deeply rooted, and the foundry business of IDMs has always struggled to gain broad market trust.

Chang’s approach was to eliminate the competitive relationship rather than relying solely on confidentiality promises. The pure-play foundry model he created was essentially an innovation in business structure, not just an improvement in technical processes. The core of this model is a commitment; TSMC's moat is not its most advanced process nodes, but the fact that "we will never compete with our customers' products."

This commitment cannot be easily replicated. Intel can invest tens of billions of dollars to catch up with the 3nm process, but it cannot guarantee to the market that it will not use its own CPUs to compete with AMD. In 1999, the value of this commitment was validated. At that time, IBM invited TSMC to jointly develop advanced copper process technology. This was a highly attractive cooperation opportunity, but Chang refused. He knew clearly that getting too close to any one customer would undermine his neutrality toward others, and neutrality was the very foundation of TSMC's survival. This is similar to the logic of a central bank: once suspected of favoring one side, the effectiveness of monetary policy is compromised, no matter how precise the interest rates are. It was this counter-intuitive choice that won the trust of hundreds of companies like Qualcomm, Apple, and AMD, driving the prosperity of the fabless design industry.

3.75% Sovereignty

In October 2024, the yield of the N4 process at TSMC's Arizona factory exceeded that of similar domestic factories in Taiwan by four percentage points. This data, reported by Bloomberg, was widely cited by supporters of the CHIPS and Science Act as evidence of a resurgence in U.S. semiconductor manufacturing. However, this higher-yield factory, once both phases are fully completed, will have an annual capacity of approximately 600,000 12-inch wafers. TSMC's total global capacity in 2024 was approximately 16 million wafers; Arizona's share is 3.75%.

Both figures are accurate. Supporters of the CHIPS Act cite the former, while critics cite the latter, and neither side is lying. This situation has formed a rare policy phenomenon: the same factory is both a technological breakthrough and an illusion of sovereignty.

Deeper problems are hidden in the schedule and costs. TSMC's factory in Kumamoto, Japan, started about a year later than the Arizona project but began production as scheduled in February 2024. The first factory in Arizona was delayed from its original 2024 production date to 2025, and the second factory was pushed from 2026 to 2027-2028. The U.S. Department of Commerce provided $6.6 billion in subsidies and approximately $5 billion in loans to support this project with a total investment exceeding $400 billion, but capital cannot solve cultural conflicts and efficiency differences. Disagreements over 12-hour work shifts, anonymous reports of data fabrication, and TSMC's employee ratings on Glassdoor (3.2/5 vs. 4.1/5, far lower than Intel's) all point to a deeper challenge. Yields can be transferred, but the ecosystem and work culture that support large-scale, low-cost operations cannot be copied. The CHIPS Act bought a technologically advanced factory, but it did not acquire a manufacturing ecosystem.

The Self-Dissolution of the Silicon Shield

The "Silicon Shield" is Taiwan's interpretation of its own geopolitical value, not a strategic arrangement by the United States. This theory posits that the global economy is so highly dependent on Taiwan's chip manufacturing that any attempt to destabilize Taiwan by force would incur unbearable economic costs, thereby creating protection. TSMC itself carefully maintains its distance from this label. Former R&D executive Konrad Young’s question clarified the corporate stance: "Why should I stand on the front lines of a war and be used as a shield?"

The deeper paradox is that the "de-risking" strategy pushed by the U.S. to have TSMC build factories in Arizona and Japan is gradually weakening Taiwan's unique leverage—the very leverage upon which the Silicon Shield theory is based. The logic of the Silicon Shield is that Taiwan is irreplaceable; the logic of U.S. policy is that Taiwan should not be the only option. These two logics cannot coexist.

As TSMC's production bases gradually distribute overseas, the global concept of security is also shifting from "interdependence equals protection" to "supply chain resilience equals security." Every time TSMC builds a new advanced node factory overseas, Taiwan's status as the world's sole source of advanced chips is weakened by a degree. On the surface, the U.S. is strengthening the supply chain security of its allies, but in reality, it is weakening the deterrent power of the Silicon Shield. The more successful TSMC is, the more urgent the U.S. feels to diversify its capacity; the more TSMC diversifies, the more fragile the Silicon Shield becomes. This is a process of self-dissolution, and no one is prepared to stop.

The 35 Years That Cannot Be Bought

The Boston Consulting Group (BCG) learning curve theory, which Morris Chang studied deeply during his time at Texas Instruments, is particularly evident in semiconductor manufacturing: for every doubling of cumulative production, the unit cost drops by approximately 20% to 30%. Over thirty-five years, TSMC has cumulatively produced billions of chips, and its cost structure is separated from any new entrant by a chasm that capital alone can hardly cross.

This is similar to surgery: a hospital can purchase the most advanced equipment, but it cannot buy the accumulated surgical volume of a surgeon, and surgical volume is the key factor determining the success rate of an operation. The learning curve in semiconductor manufacturing is highly path-dependent: one must produce enough chips to learn how to reduce costs, but one must first have sufficiently low costs to obtain orders and thus begin learning. The $52.7 billion in CHIPS Act subsidies can cover capital expenditures, but it cannot buy 35 years of accumulated knowledge.

Knowledge is not only reflected in process parameters but is also permeated in supply chain management, equipment maintenance, and a manufacturing culture where "if a problem occurs at 1:00 AM, engineers are on-site to solve it by 2:00 AM." It is also reflected in a massive ecosystem: EDA toolchains optimized for TSMC's design flows, thousands of verified IP core libraries, and a supply chain cluster within the Hsinchu Science Park where any problem can be solved within a half-hour drive. These are not things that subsidies can replicate; they can only be accumulated over time.

This forms a historical contrast with the 1986 U.S.-Japan Semiconductor Agreement. In 1986, U.S. protectionist trade policies set price floors for Japanese competitors, unintentionally creating survival space for the then-fledgling TSMC. In 2022, the U.S. attempted to use subsidies to rebuild domestic manufacturing under the giant shadow cast by TSMC, but this time, TSMC became the object of imitation. What the CHIPS Act bought was the capacity of an ally, not America's own sovereignty.

In 1987, Intel and Texas Instruments rejected the Taiwan government's investment invitation on the grounds that the pure foundry model had no competitive advantage. Thirty-eight years later, the U.S. government is using $52.7 billion to try to prove that judgment was wrong, but the way it is proving it is by buying TSMC's factories, not by rebuilding TSMC's ecosystem. The difference between the two is precisely the difference between capacity and sovereignty.

South Korea's Grand Gamble

On February 8, 1983, Lee Byung-chul announced in Tokyo that Samsung would enter the semiconductor business.

At that time, Samsung did not have a single engineer with wafer fab experience, no mass production lines, and nothing that could be called a "technological foundation." The judgment of critics was: Samsung cannot even make televisions well; pursuing cutting-edge technology is a reckless move that will inevitably fail within three years.

Nine years later, Samsung became number one in global DRAM market share, a position it held for thirty-three years.

The Confidence Behind the Tokyo Declaration

Lee Byung-chul's decision was not based on a technical feasibility assessment. In 1983, Samsung's accumulation in the semiconductor field was almost zero. It had acquired Korea Semiconductor in 1974 and began mass-producing integrated circuits for LED watches in 1975—this constituted the entirety of Samsung's semiconductor experience. Compared to Japan's NEC, Hitachi, and Toshiba, the technological gap exceeded ten years; compared to America's Intel and Texas Instruments, the gap was even larger.

On this basis, announcing an entry into DRAM yielded only one answer according to market logic: impossible.

Lee Byung-chul did not use the framework of market logic. The declaration was framed as "Serving the Country through Business" (Sa-eop Bo-guk). Semiconductors were a national strategic asset that South Korea had to master, regardless of whether market conditions supported it. The premise of this logic was: some bets cannot be measured by return on investment (ROI) because one cannot afford to lose.

The speed of execution proved that the declaration was not just a slogan. Development of 64K DRAM started in May 1983, and the first domestically produced 64K DRAM was successfully developed on December 1 of the same year. In May 1984, the first phase of the Giheung plant was completed, making South Korea the third country in the world capable of producing semiconductors, after the United States and Japan. It took less than two years from declaration to mass production, compressing the technological gap from over ten years to approximately four years.

Speed itself was a signal: a company making business decisions does not have this kind of execution rhythm; only a conglomerate executing a strategic command does.

The Structure of the Chaebol's Bet

Samsung was able to achieve this not because of technology, but because of structure.

The Samsung Group comprised about 62 companies, spanning textiles, insurance, food processing, construction, IT services, and theme parks. The core function of this structure was not diversified operations, but risk-bearing capacity. When the semiconductor division suffered continuous losses, profits from other divisions of the group could be reallocated internally to maintain investment. This type of cross-sector cross-subsidization is impossible under the governance framework of an independent listed company.

Japan's NEC, Hitachi, and Toshiba were independent listed companies that had to be accountable to shareholders. Quarterly losses would directly trigger shareholder pressure, limiting their ability to endure long-term losses. In 1983, Japanese companies already dominated the DRAM market and had no reason to accept long-term losses while holding a technological lead.

The logic of the Chaebol structure is similar to the logistics system of an army: frontline troops can continue fighting even when ammunition is exhausted because there is a supply line in the rear. The logic of an independent listed company is that of a mercenary: they retreat if there is no pay. Samsung's semiconductor division was the frontline unit, while textiles and insurance were the supply lines.

The path to technology acquisition also relied on the institutional conditions behind the structure. In 1983, Samsung obtained DRAM technology licenses from Micron Technology and Sharp. The South Korean government linked the entry of foreign companies into the Korean market with technology transfer; foreign companies had to transfer semiconductor technology in exchange for market access. Micron's licensing fee was estimated by the industry to be around $5 million, which was inexpensive given the technological value at the time.

Simultaneously, Samsung recruited Korean-American engineers from Silicon Valley. These engineers possessed practical wafer fab operational experience, filling Samsung's gaps in manufacturing processes.

The role of the South Korean government was to create institutional conditions, not to bear the risk. The government did not directly fund Samsung's DRAM R&D; instead, it coordinated joint research through the Electronics and Telecommunications Research Institute (ETRI), protected Samsung's cash flow sources by restricting imports of Japanese consumer electronics, and created institutional space for Samsung to acquire technology through market access conditions.

The entity that truly bore the risk of loss was the Samsung Group itself. Government support was a necessary condition, not a sufficient one.

The Gift of 1986

In September 1986, the United States and Japan signed the Semiconductor Trade Agreement. The core terms of the agreement stipulated that the export prices of Japanese chips must be jointly set by the US and Japanese governments to prevent sales below cost; price controls extended to third-country markets; and Japan committed to promoting the sale of US chips in the Japanese market, targeting a share of at least 10%.

The design goal of the agreement was to protect the US chip industry and stop Japanese companies from dumping at low prices.

One of the actual effects of the agreement was to open a market window for the South Korean DRAM industry.

The price distortion was obvious: 1Mb DRAM was priced at $5.00 in the US market and $3.90 in the European market; 4Mb DRAM was $30 in the US and $18 in Europe. Japanese companies were forced to maintain artificially high prices, while the agreement did not apply to South Korean companies. Samsung, Hyundai, and LG could sell at prices lower than the Japanese while maintaining profit margins.

This was one of the key turning points for the South Korean DRAM industry to shift from loss to profit.

But this "gift" had a prerequisite: Samsung had to be there waiting already.

When the agreement was signed in 1986, Samsung's Giheung plant had already been operating for two years, and capacity had been established. If Samsung had not persevered through the loss-making period of 1983-1986, the 1986 market window would have been meaningless to them; without capacity, they could not have captured the orders when the price window opened.

The more accurate causal chain is: Chaebol structure → Enduring losses → Establishing capacity → Catching the unexpected window.

The 1986 agreement was a trigger, not the cause. South Korea's success was not due to good luck, but because they had placed themselves in a position to catch that luck before it arrived.

In 1992, Samsung developed the world's first 64Mb DRAM and achieved the world's number one DRAM market share in the same year, surpassing NEC, Hitachi, and Toshiba. It held this position for thirty-three years until it was briefly surpassed by SK Hynix in the first quarter of 2025.

Can the Korean Model be Replicated?

Samsung's success required three conditions to be met simultaneously: the Chaebol structure (the ability to endure long-term losses), a technology acquisition path (the institutional conditions of Micron licensing plus engineer recruitment), and an external window (the market opportunity created by the 1986 agreement).

China's semiconductor strategy after 2015 has attempted to replace the Chaebol structure with state subsidies in the first condition. SMIC and YMTC received massive direct government capital injections, attempting to replicate Samsung's 1983 trajectory.

The alternative solution has profound differences in incentive mechanisms. The risk-bearing capacity of the Chaebol structure comes from internal capital allocation; the Samsung Group used profits from other divisions to subsidize semiconductor losses. Decision-makers faced direct financial pressure from losses while having sufficient capital buffers. Subsidies for state-owned enterprises come from the outside; the pressure structure for decision-makers regarding losses is different, as are the incentive mechanisms for technological breakthroughs.

The second condition has been severed. From 2022 to 2023, US export controls explicitly prohibited the transfer of advanced semiconductor technology to China, and there is no institutional space for "market access in exchange for technology." Samsung was able to obtain Micron's license for about $5 million in 1983 because the South Korean government created institutional conditions that made Micron willing to license based on business logic. Today, these institutional conditions do not exist.

As for the third condition, no external window similar to 1986 has appeared. The US-Japan agreement was a product of US pressure on Japan, and its unintended effect was to create market space for South Korea. In today's geopolitical landscape, the direction of US pressure is toward China; no similar unexpected gift exists.

The three conditions of the Korean model are all indispensable, and the fact that all three were met simultaneously between 1983 and 1992 is a historical accident of extremely low probability.

This judgment is uncomfortable for proponents of "omnipotent national will": the Korean model required the Chaebol structure, which is a product of historical path dependency, not something that can be designed by policy. It is equally uncomfortable for proponents of "omnipotent market logic": Samsung's success proves that under specific historical structures, national will can indeed override market logic and ultimately achieve commercial success.

The logic of the critics in 1983 was not wrong; within the framework of market logic, their judgment was correct.

The problem was that Lee Byung-chul was not using the framework of market logic. Whether this framework can be replicated in China today depends not on whether the will is strong enough, but on whether the historical structures supporting the framework exist. Samsung's Tokyo Declaration held true not just because Lee Byung-chul said "yes," but because the 62 companies of the Samsung Group behind him said, "We will bear the burden."

The Efficiency Trap

Within 72 hours of the Great East Japan Earthquake on March 11, 2011, procurement departments in the global semiconductor industry received a top-level alert: Shin-Etsu Chemical's Shirakawa plant had halted production. This plant accounted for 20% of the global supply of semiconductor silicon wafers. Without silicon wafers, there are no wafers; without wafers, chip production lines come to a standstill. Over long years of pursuing efficiency, the globalized supply chain concentrated the vast majority of its resources into a single basket—one that happened to be placed on the Pacific Ring of Fire.

The Supply Chain Map of a Single Machine

An ASML EUV lithography machine consists of 100,000 components and 2 kilometers of cabling. Transporting this equipment, priced at over $100 million, from Europe to an Asian wafer foundry requires 40 standard containers, 20 heavy-duty trucks, and three Boeing 747 freighters (Asia Times, 2021). This machine can be called the steel exoskeleton of globalization.

ASML only produces about 15% of the core components itself; the remaining 85% rely on a precision supply network covering the globe (Incremental Returns). Germany's Zeiss is responsible for the optical systems, America's Cymer provides the extreme ultraviolet light source, and Germany's Trumpf supplies the high-power lasers (Trumpf, 2025). In the field of chemical materials, Japanese companies firmly control about 80% of the global photoresist market, with EUV photoresist almost entirely monopolized by Japanese suppliers in its early commercialization phase (Fountyltech, 2020). Among them, JSR and Shin-Etsu Chemical together hold a 90% share of the photoresist market.

Once the machine stops, the entire industrial chain suffers a shock.

The distribution of the semiconductor manufacturing landscape is identical to the specialization trap in ecology. The more specialized a species evolves, the higher its efficiency in obtaining resources in a stable environment, but the weaker its ability to survive environmental mutations. Every supplier is the only company in the world capable of delivering a specific part at a specific yield. While extreme specialization creates engineering miracles, it also amplifies single-point-of-failure risks to the level of the entire industry. A break at any tiny node can instantly halt an assembly line worth hundreds of billions of dollars.

The Logic of Just-in-Time

In modern business contexts, supply chain efficiency is often equated with cost minimization, while resilience is ignored. The Just-in-Time (JIT) inventory management model creates incredible capital turnover miracles during stable periods; its core logic lies in pushing risk into an unknown future. Companies trade zero-inventory balance sheets for a bold assumption: upstream supply will always be stable, and logistics networks will never be interrupted. However, the market has never priced this fragile assumption.

The 2021 chip shortage was precisely the backlash against this assumption, causing economic losses of approximately $210 billion (AlixPartners, 2021). In 2020, the automotive industry cut orders due to pandemic panic, and wafer foundries quickly shifted idle capacity to the consumer electronics sector where demand was surging. When automotive demand unexpectedly rebounded, car companies found that chip lead times had stretched to over 26 weeks. That year, General Motors and Ford were forced to close several core production lines. The automotive industry's priority in semiconductor procurement was already lower than that of consumer electronics; it was the first to be squeezed during capacity constraints, as market pricing logic remained unchanged.

However, Toyota, the advocate of Just-in-Time, remained unscathed during the crisis. After experiencing multiple natural disasters, the company had quietly changed its strategy, stockpiling months of inventory for critical semiconductor components.

TSMC's Gravity

The high concentration of market share is an inevitable result of economies of scale; scale determines destiny. The capital threshold for developing a generation of advanced process nodes has become breathtakingly high, with TSMC's annual R&D expenditure consistently exceeding $4 billion. Only the largest wafer foundry can amortize these massive sunk costs through high-volume silicon wafer shipments and quickly reinvest the resulting cash flow into the competition for next-generation lithography equipment. Lower per-chip costs attract more customer orders, and ample cash flow feeds back into the R&D of the next-generation process, ultimately forming an exclusive positive feedback loop.

As of the second quarter of 2025, TSMC occupies about 70% of the global wafer foundry market, with advanced processes contributing nearly 70% of total revenue (Tom's Hardware, 2025). This precision collaboration system, built on sand and light, has irreversibly concentrated toward a single geographic coordinate. Economic model calculations by Bloomberg in February 2026 show that if a conflict between the US and China were to occur in the Taiwan Strait, the global economy would face a devastating loss of approximately $10 trillion. This phenomenon is not the malice of a monopoly but the natural result of capital chasing the highest rate of return; concentration is a law of physics, not merely a business choice.

The 2011 Rehearsal

The disaster script has already been performed, yet the system has not been repaired. While the 2011 Japan earthquake destroyed Shin-Etsu Chemical's Shirakawa plant, MEMC's silicon wafer production line in Utsunomiya also halted simultaneously; the combination caused 25% of the global semiconductor silicon wafer supply to dry up instantly (Bloomberg, 2011). Renesas Electronics lost 40% of its microcontroller capacity. The disaster in the areas where Mitsubishi Gas Chemical and Hitachi Chemical were located cut off the source of 70% of the world's copper-clad laminate materials.

This was an extreme stress test of geographic concentration. Although the system miraculously resumed operation after several months of violent shocks, it was entirely because factories could be rebuilt and damaged machines could be repurchased.

The 2021 chip shortage a decade later was a collapse on another level. It required no crustal movement to cause physical destruction; a mere slight deviation in demand forecasting models was enough to paralyze the entire industrial chain. The trigger mechanisms of the two crises were completely different, yet both pointed to the same deep dilemma. Every participant in the industrial chain is aware of the system's fatal weaknesses, but under the profit pressure of quarterly financial reports, the allure of efficiency always easily overrides any proposal to build redundant buffers.

Can Decoupling Solve the Problem?

Washington is attempting to use political means to forcibly change the laws of physics, at a high cost. Through the $52.7 billion CHIPS Act subsidies, the US government is requiring TSMC to invest up to $100 billion in production capacity in the US (data as of January 2026). From a static ledger perspective, decoupling seems to be repairing geographic fragility. Analysis by TechInsights based on the Scotten Jones strategic cost model shows that since equipment depreciation costs account for more than 70% of semiconductor manufacturing and labor costs less than 2%, the actual cost premium of TSMC's Arizona plant compared to its Taiwan plants is only about 10%.

However, this static calculation ignores the dynamic cruelty of technological evolution. Decoupling has not eliminated fragility; instead, it has transformed the physical risk of a broken chain into the systemic risk of technological stagnation.

Without the scale support of a unified global market, fragmented domestic capacity will struggle to generate sufficient profits and will be unable to support the procurement of next-generation EUV lithography machines and the massive R&D for processes below 2nm. Taiwanese officials explicitly told Reuters in February 2026 that transferring 40% of advanced process capacity is an absolutely untouchable bottom line. The US is attempting to defend technological sovereignty through export controls and reshoring capacity, but in doing so, it has personally dismantled the commercial foundation required to maintain technological leadership.

The deeper contradiction is that the logic of decoupling presupposes a stable technological endpoint, as if technological sovereignty can be guaranteed simply by moving existing capacity back home. The reality of semiconductor manufacturing is the opposite: process nodes iterate every two years; today's 3nm plant will become obsolete capacity in four years. Without scale, R&D funding is lost; without R&D funding, next-generation processes are impossible; with process stagnation, the strategic value of domestic plants will drop to zero within one technological cycle.

No one can accurately predict which fragility—the single point of failure on a seismic belt or the technological stagnation from losing iteration funds—will trigger destruction sooner.

China's Chip Dream

In 1997, Huajing Electronics held its production ceremony in Wuxi. This production line, which took seven years and cost 2 billion RMB, finally began producing wafers. At that time, the international mainstream process had reached 0.35 microns, while Huajing had only achieved 0.8 microns, lagging two generations behind technologically.

Semiconductor technology follows Moore's Law, iterating every two years. In other words, the moment a national project is approved, its target technology may already be an obsolete product in the global market. The logic of Project 908 seemed clear: capital for equipment, equipment for processes, and processes for competitiveness. This logic works in fields like high-speed rail and nuclear power, but in the semiconductor sector, every step has encountered obstacles.

The Bill for Three Charges

Project 909 continued a similar path, investing nearly 10 billion RMB, exceeding the sum of all national investments in the electronics industry since the founding of the People's Republic. Hua Hong NEC achieved a 94% yield rate in 1999 and generated 1.773 billion RMB in sales revenue in 2000. However, in 2001, global DRAM market prices plummeted, leading to a single-year loss of 1.384 billion RMB. Transitioning to foundry work became a pragmatic choice, but it also meant abandoning the route of independent process R&D.

Entering the era of the Big Fund, investment scales grew exponentially. Phase I in 2014 invested 98.7 billion RMB, Phase II in 2019 invested 204.1 billion RMB, and Phase III in 2024 is expected to reach 344 billion RMB, totaling approximately 646.8 billion RMB—a hundreds-fold increase in capital scale. However, for the Huawei Ascend 910B AI chip manufactured by SMIC, the yield rate still hovered around 20% after more than six months of mass production; for every five chips produced, four were defective.

The accumulation of capital cannot bridge the gap in craftsmanship. Money can buy equipment, but it cannot buy processes; even if processes are obtained, yield rates cannot be guaranteed; even if yield rates improve, the establishment of an ecosystem remains out of reach. Each layer depends on the one above it, and the very bottom of this dependency chain is now blockaded.

Why Semiconductors are Different

The catch-up strategy for high-speed rail and nuclear power provided a seemingly successful template. From importing entire train carriages to disassembly and reverse engineering, and finally to independent R&D, this path is feasible because they are system integration problems. Components can be disassembled, processes can be observed, and knowledge can be precisely encoded into engineering drawings for successors to reference.

But semiconductors are not like this. Manufacturing a 7nm chip using EUV lithography requires only 9 steps, but under export controls, relying entirely on DUV lithography with multiple exposures to manufacture the same process causes the number of steps to skyrocket to 34. These 25 additional steps are not simple physical additions. Improving yield rates depends on the experience of engineers accumulated through tens of thousands of chemical vapor deposition, etching, and lithography experiments. This knowledge cannot be written into operation manuals, nor can it be fully transferred by purchasing equipment or headhunting technical executives, because it is deeply embedded in the collective memory of a team of hundreds collaborating.

The R&D investment intensity of China's semiconductor industry is only 7.6%, far lower than the 18.8% in the United States. More capital flows toward capacity expansion rather than basic research. TSMC's moat has never been just the ASML equipment in its cleanrooms, but the failure data recorded on every batch of scrapped wafers over thirty years—data that is not on any procurement list.

Corruption in the Big Fund

From July to September 2022, a series of investigations swept through the semiconductor investment hub controlling over 120 billion RMB in capital. Lu Jun, former CEO of SMIC Capital; Ding Wenwu, former President of the Big Fund; as well as Wang Wenzhong, Ren Kai, and others were investigated in succession. Lu Jun was accused of bribery, embezzlement, abuse of power, and even using public funds to renovate real estate for personal use.

The state invested hundreds of billions in a massive effort to shorten the technological gap with global advanced processes, but part of those funds flowed into high-end property renovations for fund managers. This is not just a moral issue; it exposes the deep-seated problems arising from the forced combination of national will and market-oriented operations. When fund managers simultaneously hold absolute investment decision-making power and actual control over invested enterprises, the core metric of project evaluation quietly shifts from "which technical node most needs to be conquered" to "which project provides the greatest rent-seeking space." Administrative mandates to quickly spend huge budgets, combined with market-oriented operational authority lacking effective checks and balances, together spawned a massive arbitrage network. Capital indeed flowed out of the treasury as scheduled, but it bypassed the most difficult underlying technologies and flowed toward peripheral projects that could be quickly packaged for listing to provide short-term returns.

The reform direction for Phase III of the Big Fund (2024) is for the six major state-owned banks to participate directly in funding and strengthen supervision. This solves the rent-seeking problem to some extent, but it cannot solve the challenge of accumulating tacit knowledge.

Acquisitions: Closing Another Path

The slowness of the self-built path gave rise to a seemingly more efficient alternative: acquiring overseas companies with advanced technology. In July 2015, Tsinghua Unigroup, a subsidiary of Tsinghua University, made a $23 billion acquisition offer to Micron Technology, equivalent to $21 per share, a premium of about 19% (Reuters, NYT, July 2015). Micron did not haggle; the board's response cited only one reason: a CFIUS review was almost certain, and the probability of the deal passing was zero. The absolute amount of the bid was meaningless in the face of regulatory barriers.

In February 2016, Fairchild Semiconductor, facing a $2.6 billion joint bid from China Resources Microelectronics and Hua Capital, made the same choice as Micron (FT, Reuters, February 2016). This venerable company rejected a massive cash injection that would have pulled it out of financial distress simply because it knew the money could not pass through the labyrinth of Washington's scrutiny. The fact that target companies proactively reject high-premium acquisitions shows that external policy deterrence has been deeply embedded into the behavioral patterns of the capital market.

The logic of CFIUS interventions and export controls cuts off two completely different paths in the semiconductor supply chain. Export controls prevent the output of technology, while CFIUS blocks the transfer of technology ownership. This transfer of ownership involves the overall handover of R&D systems, organized teams of engineers, and underlying patent portfolios; its threat to technological sovereignty far exceeds the loss of a single piece of equipment. The Foreign Investment Risk Review Modernization Act (FIRRMA), signed into law in August 2018, completely codified this logic (Treasury Department FIRRMA Summary, August 2018): the review threshold was forcibly lowered from traditional "control acquisitions" to "any meaningful participation." Even if Chinese capital only attempts to purchase a 5% non-controlling minority stake, as long as the target involves critical technology, it will trigger mandatory filing procedures.

The rise and fall of Tsinghua Unigroup provides a clear financial audit of the cost of "buying a path." This enterprise, which attempted to achieve a "leapfrog" through high-leverage mergers and acquisitions, found itself with nowhere to place its massive capital costs after repeated setbacks in overseas acquisitions. The approximately $30 billion in debt accumulated through M&A eventually crushed the parent company's balance sheet, leading to its bankruptcy reorganization in 2021 due to a $197 million bond default (Asianometry, 2021).

Both Paths Blocked Simultaneously

The two paths of China's semiconductor catch-up strategy—self-building and acquisition—encountered distinctly different blockages between 2015 and 2019. The obstacles to the self-built path were endogenous: tacit knowledge cannot be purchased with capital, yield accumulation cannot be compressed by administrative orders, and corruption in the Big Fund further diverted limited resources in the wrong direction. The obstacles to the acquisition path were exogenous: CFIUS systematically closed the channels for technology ownership transfer, and FIRRMA completely welded that door shut.

The blocking of both paths was no accident; it was because U.S. policymakers deeply understood the two ways semiconductor knowledge propagates and designed targeted blockade mechanisms for each. Export controls cut off the flow of equipment, and CFIUS blocks the flow of capital. Together, these two systems constitute a complete defensive line against technology diffusion.

The investment focus of Phase II of the Big Fund (2019) shifted from chip design and manufacturing to semiconductor equipment, materials, and components, clearly reflecting this forced strategic adjustment. The problem is not a lack of funds to build factories, but that the underlying tools to equip these wafer fabs are still insufficient. Phase III of the Big Fund (2024) entered the field with 344 billion RMB, with the six major state-owned banks directly participating in funding and endorsement. Although the capital scale is large, the regulatory mechanism has tightened significantly following the anti-corruption purge.

Deep-seated obstacles remain; the ceiling of the fourth charge is determined by the blockade at the equipment layer, not the scale of capital. The most advanced domestic lithography machines currently provided by local manufacturers like Shanghai Micro Electronics Equipment (SMEE) still remain at the 28nm node. To leap across several generations of technological gaps to reach commercially viable levels for advanced processes before 2030 is a massive unknown. Without the introduction of new-generation EUV equipment, relying solely on existing DUV machines for extreme multiple exposures will mean that improvements in yield rates will proceed slowly in an inefficient and expensive manner—this is another life-and-death hurdle.

When Project 908 was established, Chinese engineers aimed for 0.8 microns. When it went into production seven years later, that goal had already become obsolete technology. The fourth charge still faces the same problem, only on a larger scale and at a higher cost, and this time, both paths have been sealed shut.

Huawei and HiSilicon: The Ceiling of the Spare Tire

September 14, 2020

September 14, 2020, marked the deadline for TSMC to supply Huawei with advanced process chips. In the weeks leading up to this, Huawei engineers exerted every effort to urge TSMC to deliver the batch of 5nm Kirin 9000 orders. The quantity of chips determined how long Huawei's flagship smartphone line could persist in the market; the answer was approximately two years.

According to the original letter reposted by the Tencent Cloud Developer Community, in the early hours of May 17, 2019, He Tingbo, President of HiSilicon, issued an all-staff letter announcing that "all the spare tires we have ever built have been promoted to regular status overnight." This sentence triggered a wave of emotion on the internet, while another sentence in the letter was rarely mentioned: "The buffer zone has disappeared, and every new product, from the moment it is born, must simultaneously adopt a 'technological self-reliance' solution." The spare tire plan was a one-time solution; once activated, there would be no new batch of spare tires.

After TSMC's production lines were cut off, Huawei faced not the activation of spare tires, but their exhaustion. The inventory of Kirin 9000 supported the Mate 40 and parts of the Mate 50 series. Once the inventory was depleted, the Kirin 9010 could only restart from a lower technological starting point. Between the declaration of May 17, 2019, and the reality of September 14, 2020, a gap emerged.

Fifteen Years of Spare Tires

Going back to 2004, according to internal data cited by EET China, Ren Zhengfei allocated a research and development team of 20,000 people and an annual budget of 400 million USD for HiSilicon. At that time, Huawei's total annual R&D expenditure was less than one billion USD. Such a resource tilt clearly did not align with short-term financial return logic; it was a proactive challenge to market laws. Ren Zhengfei's judgment was that advanced American chips would sooner or later become difficult to obtain, and alternative solutions had to be prepared in advance.

Early products experienced many setbacks. The K3V1 was jokingly called a "hand warmer" due to severe overheating, and the K3V2 was phased out due to technical defects. The Kirin 960 (2016, 16nm) began the pace of catching up, and the Kirin 970 (2017, 10nm) integrated the world's first mobile-side NPU. GSMArena confirmed in a 2018 launch report that the Kirin 980, manufactured by TSMC, became the world's first 7nm mobile chip, approximately half a year ahead of the Qualcomm Snapdragon 855. The Kirin 990 (2019, 7nm+) integrated a 5G baseband, and the Kirin 9000 (2020, 5nm) completed the peak of process evolution before TSMC's supply was cut off.

The strategic prediction was correct, but the coverage of the spare tire plan had blind spots: it solved the problem of "existence," but it failed to make arrangements for the transition period after the supply cut. From 2021 to 2023, a chip vacuum of about two years appeared in the flagship model product line. This vacuum was an area that the fifteen-year spare tire plan failed to cover.

This structure bears similarities to the Soviet military-industrial backup system during the Cold War: a backup system can maintain operations under extreme conditions, but its upper limit is determined by the technological level of the original supply chain, rather than the completeness of the backup. HiSilicon's spare tires guaranteed Huawei's survival floor but could not break through the ceiling of technological development.

The Bill for Activating the Spare Tires

"All the spare tires we have ever built have been promoted to regular status overnight." This declaration, written on the eve of the supply cut, underwent the test of reality in the following years.

Successful activations did indeed exist: mobile SoCs (Kirin series) were sustained through SMIC, baseband chips were integrated into the Kirin 9010, and Ascend AI chips and Kunpeng server chips established a firm foothold in the domestic data center market. Rumors once circulated in the market about the comprehensive replacement of desktop processors by HiSilicon, which were subsequently refuted with detailed data by the Securities Times; a spare tire for PC chips never appeared.

The other side of the bill is more noteworthy. Activating the spare tires solved the survival crisis of moving from nothing to something, but it failed to cross the technological chasm from existence to excellence. Ascend and Kunpeng are constrained by process bottlenecks on the server side, and advanced process mobile chips have lost the foundation to compete with TSMC. The failures on the list are more prominent than the successes.

He Tingbo also wrote in the letter: "The road ahead will not have another ten years to build spare tires and then promote them to regular status." This means that the spare tire strategy is a one-time emergency measure, not a supply chain that can be relied upon long-term. After activation, Huawei faces a world without a buffer zone, where every generation of new products must be completed within the boundaries of existing manufacturing capabilities.

SMIC's 7nm

A hardware teardown report released by TechInsights in September 2023 showed that the Huawei Mate 60 Pro was equipped with the Kirin 9010 chip using SMIC's 7nm (N+2) process. This was a genuine technological breakthrough and a public demonstration of China's chip manufacturing capabilities three years after the implementation of U.S. export controls.

However, SMIC's 7nm is not at the same level as TSMC's 7nm. SMIC lacks Extreme Ultraviolet (EUV) lithography machines and must rely on Deep Ultraviolet (DUV) machines to achieve 7nm circuits through multiple exposures. This method caused the yield rate to drop below 50%, far below TSMC's benchmark of over 90% for the same node. Although both are called "7nm," the two differ significantly in cost structure, yield performance, and scalability.

According to a 2023 report by Jefferies analysts, yield constraints suppressed SMIC's 7nm quarterly shipments to between 2 million and 4 million units. The rollout of the Mate 60 Pro proved the feasibility of this technological path, but the yield data simultaneously indicated that it is difficult to scale economically.

The Kirin 9000 represents the pinnacle of Chinese chip capability, and it also illustrates the limited height of this pinnacle. The Kirin 9010 is not a continuation of the Kirin 9000, but a reconstruction from a lower starting point: 5nm regressed to 7nm, and TSMC's yield regressed to SMIC's yield. These two regressions are not temporary but are determined by the physical limitations of manufacturing equipment. Without EUV, SMIC's process evolution encountered a hard ceiling at the 7nm node.

Huawei Can Win, HiSilicon Cannot

The Chinese smartphone market in 2025 witnessed a movement to reclaim lost ground. According to full-year 2025 data released by IDC, Huawei returned to the top of the domestic market for the first time in five years with 46.7 million units shipped and a 16.4% market share. Canalys data cited by the South China Morning Post showed that in 2024, Huawei's shipments grew by 37% year-on-year, while Apple's fell by 17% during the same period, dropping from first to third in the Chinese market. The original logic chain of U.S. export controls was broken in the Chinese market: cutting off supply did not lead to the collapse of the target enterprise but instead stimulated replacement demand among domestic consumers.

However, in overseas markets, the situation remains grim. The 5G ban has not been lifted, and the penetration rate of the HMS ecosystem outside of China is far lower than that of Google services. Huawei's share of the European mobile phone market fell from about 20% before 2020 to single digits. HiSilicon's technological ceiling is locked by SMIC's process capability—that is, the DUV limit without EUV. By 2025, TSMC had already mass-produced 3nm, and the generational gap between SMIC and TSMC remains at three to four generations. Growth in terminal sales cannot change the lithographic precision on silicon wafers.

Ren Zhengfei's 2004 judgment was correct: chips would eventually have to be made by themselves. He Tingbo's team used fifteen years to prove that national will can forcibly say "yes" when market logic says "no." The supply cut also proved the other side: spare tires can be activated, but activated spare tires cannot run faster. Huawei won the survival battle in the domestic market, but HiSilicon cannot win the race for the world's top-tier processes. The reason is not a lack of will, but rather that without an EUV manufacturing system, the ceiling is right there.

This ceiling will not be raised by any policy declaration; it is determined by the laws of physics: the wavelength of EUV lithography is 13.5nm, while DUV is 193nm. Multiple exposures can bridge part of the gap but cannot eliminate it. The story of HiSilicon is the most complete allegory of the chip war and its most honest footnote.

The Candle Factory

On November 11, 2009, SMIC issued two major announcements: Richard Chang resigned as CEO, and the company reached a settlement with TSMC. According to the agreement, SMIC was required to pay $200 million in cash compensation and transfer approximately 10% of its equity. This was the second time Richard Chang had lost to TSMC in a legal battle.

The first time occurred in 2000. At that time, TSMC launched a hostile takeover of Worldwide Semiconductor Manufacturing Corp (WSMC), and Richard Chang lost control of his own company without any prior knowledge. The founder, who had accumulated 20 years of experience in factory construction at Texas Instruments, subsequently chose to head north to Shanghai, initiating SMIC's ten-year development journey. In 2018, he founded SiEn in Qingdao, starting his third venture. This time, he chose the 8-inch mature process, completely avoiding that difficult-to-cross technical barrier.

Across three ventures, Richard Chang has always been doing the same thing: bringing wafer fabrication capabilities from established locations to undeveloped ones.

The Logic of Three Startups

The logic of migration was even more intuitively reflected in Liang Mong-song. In 2017, Liang Mong-song joined SMIC from Samsung, bringing with him a complete process roadmap for 14nm FinFET. In just 298 days, SMIC's 14nm yield rate improved from 3% to over 95%. Within TSMC, such transferred processes were once referred to as BKM1 (Best Known Method One). This code name reveals the core of semiconductor manufacturing more precisely than any courtroom defense: there is only one optimal process method, and it can be carried away through the movement of talent.

TSMC paid the price for this twice. In 2003, TSMC sued SMIC in U.S. federal court, alleging the theft of trade secrets. Court evidence showed that SMIC's 0.18μm process flow was highly similar to TSMC's, involving 15,000 documents and 500,000 pages of data. In 2005, the two parties reached their first settlement, with SMIC paying $175 million. However, SMIC failed to return the relevant documents, and TSMC filed suit again in 2006. In 2009, a California jury ruled against SMIC, with potential damages as high as $1 billion, eventually resulting in a settlement of $200 million in cash plus 10% equity.

By the time Richard Chang resigned, SMIC had already proven one point: talent is not the bottleneck. As long as there is sufficient high-end talent mobility, process knowledge and management systems can be precisely replicated. Richard Chang brought process knowledge from Texas Instruments and factory construction experience from WSMC; Liang Mong-song brought the FinFET roadmap from TSMC and Samsung.

Machines cannot be packed up and taken away, but the parameters in the human brain can cross the Pacific. ASML's EUV lithography machines cannot be replicated through the movement of personnel, but process knowledge can migrate with talent.

The Self-Reinforcement of Equipment Bans

On December 18, 2020, the U.S. Department of Commerce officially added SMIC to the Entity List. For equipment export applications targeting 10nm and below processes, the review standard was set to a presumption of denial.

The ban cut off the only channel for obtaining EUV equipment. To advance its advanced processes, SMIC had to pivot toward DUV multi-patterning paths on multiple fronts. However, this choice brought a deeper dilemma: multi-patterning requires more etching machines and thin-film deposition equipment, most of which come from U.S. companies like Lam Research and Applied Materials, which are similarly restricted by export controls.

A blockade mechanism thus formed a closed loop: every technical alternative attempted to bypass a single point of blockade increases physical process dependence on other controlled equipment. The bypass paths consume more restricted resources. Starting in 2023, ASML further restricted the export of its most advanced DUV equipment to China, raising the barriers of the technical blockade once again.

The physical path difference between multi-patterning and EUV is like using multiple manual tracings to replace high-precision laser printing: theoretically, extremely high precision can be achieved, but with every step of refinement, the number of process steps grows exponentially. TSMC's 7nm process using EUV requires approximately 40 to 50 mask layers, while SMIC's use of DUV multi-patterning requires 80 to 100 layers. Every additional mask layer multiplies the risk of alignment errors, and processing time and costs skyrocket accordingly.

In a factory relying on candles for illumination, workers attempt to produce advanced LED bulbs by extending labor hours and increasing process steps. The workers' efforts are real, but the upper limit of luminous flux was already locked by the laws of physics the moment the candle was lit.

The Truth of 7nm: Breakthrough or Limit

In September 2023, a teardown report from TechInsights confirmed that the Kirin 9000s chip in the Huawei Mate 60 Pro utilized SMIC's N+2 (7nm-class) FinFET process. The market viewed this as a technical breakthrough, but this technical label masks the underlying commercial reality.

TSMC's N7 process achieved mass production in 2018, while SMIC's N+2 process landed in 2023, a time gap of five years. A more critical gap lies in the yield rate: TSMC's 7nm process, due to the introduction of EUV, has a stable yield rate of over 90%, while SMIC's N+2 yield is estimated at only 20% to 33%.

When the yield rate is only 20%, for every 5 wafers a fab produces, 4 become scrap.

Such a vast difference in yield means that even if both foundries have carved transistors of similar density onto silicon wafers, TSMC is conducting efficient industrial production, while SMIC is performing laboratory-level technical assaults regardless of cost. When advancing to the 5nm N+3 node, SMIC's production costs are approximately 50% higher than TSMC's EUV-equivalent process, with a yield of only about 33%. The fab's assembly line becomes a black hole devouring capital.

SMIC's 7nm process is not a mass-production technology poised to disrupt the market, but rather an expensive specimen that proves physical feasibility but is difficult to sustain economically. The massive chasm between breakthrough and mass production is the true manifestation of the equipment ban's effectiveness.

Asking "Can China catch up?" is inaccurate; the real question is: why is the cost structure of catching up destined to be asymmetrical?

SMIC's customer structure partially reveals the answer. On September 14, 2020, SMIC stopped supplying Huawei HiSilicon, the first day after Huawei was added to the Entity List. Since then, SMIC's advanced process capacity has served almost exclusively a single customer, Huawei, whose demand is far from sufficient to support an economically viable advanced process production line. TSMC's 7nm capacity simultaneously serves Apple, Qualcomm, AMD, and NVIDIA, with economies of scale compressing unit costs to levels SMIC cannot reach. SMIC's advanced process is, in effect, an expensive private line customized for a single client, rather than an industrial highway that shares costs.

The Ceiling of the Candle Factory

In 2024, SMIC's annual revenue reached 57.8 billion RMB (approximately $8.07 billion), a year-on-year increase of 27.7%. However, the annual gross margin was only 18.6%, a stark contrast to TSMC's gross margin, which consistently stays around 53%.

The massive gap in profit margins reveals the dilemma of the pursuer: to maintain technical iteration in the absence of the most advanced tools, the foundry must invest heavily to compensate for the scrap loss and lengthy processes brought by multi-patterning. SMIC's revenue growth relies heavily on increasing wafer output to make up for low yields, much like running on a treadmill—the faster one runs, the greater the exhaustion.

Without EUV, 3nm and below processes are no longer economically viable. This is not because they are technically impossible to achieve, but because the imbalance between cost and yield prevents the products from entering a commercial cycle. DUV multi-patterning can theoretically continue to advance, but with every generation of shrinkage, the exponential growth in mask layers will push costs into a range that no commercial logic can sustain; the laws of physics do not compromise on this.

SMIC's true ceiling is not the technical capability of Chinese engineers, but the red line where the equipment ban transforms technical possibility into economic impossibility. Richard Chang proved with three startups that talent and process knowledge can migrate, and Liang Mong-song proved in 298 days that yield can be improved through engineering effort, but no amount of talent mobility can replicate ASML's EUV lithography machines.

This is precisely the most accurate cut in the chip war: do not cut off the talent, only cut off the machines. Talent can flow, but equipment cannot be packed up. The effectiveness of the blockade is built upon this very asymmetry. Talent bans can be evaded, but equipment bans are difficult to bypass.

The Door Closes

On October 7, 2022, the Bureau of Industry and Security (BIS) of the U.S. Department of Commerce released a 139-page update to export control rules. The technical annexes of the document were so complex that most media coverage could only remain at the headline level. However, semiconductor industry analysts quickly captured its profound meaning: the blockade system completed its final closed loop on that day.

The construction of this system began seven years ago.

The First Layer: The Capital Channel (2015-2018)

In July 2015, Tsinghua Unigroup made a $23 billion takeover bid for Micron Technology, but the deal died before it was even formally filed due to the implicit veto power of the Committee on Foreign Investment in the United States (CFIUS). This was only the beginning. Washington then spent three years gradually transforming this implicit veto power into an institutionalized review mechanism.

In February 2016, CFIUS for the first time proactively intervened in Tsinghua Unigroup's investment in Western Digital, forcing Tsinghua to terminate its $3.775 billion investment plan (Reuters, Caixin, February 23, 2016). This proactive intervention demonstrated that even without a filing, regulation could not be bypassed. In December of the same year, President Obama signed an executive order prohibiting China's Fujian Grand Chip Investment Fund from acquiring the U.S. subsidiary of Aixtron, citing the military application value of Gallium Nitride (GaN) technology (Obama White House Archives, December 2, 2016). In September 2017, President Trump continued this path, signing an executive order to block Canyon Bridge Capital's $1.3 billion acquisition of Lattice Semiconductor (Treasury Department, Harvard Law School Forum, September 2017).

In August 2018, the Foreign Investment Risk Review Modernization Act (FIRRMA) formally codified previous case experience (Treasury Department FIRRMA Summary, August 2018). The review threshold was significantly lowered from "acquisition of control" to "any meaningful involvement." Even if Chinese capital only attempted to purchase a 5% non-controlling minority stake, a mandatory filing would be required as long as the target involved critical technology.

Within four years, the capital channel completed the journey from implicit deterrence to institutionalized blockage.

The Second Layer: The Equipment Channel (2018-2020)

After the capital path was blocked, China's strategic focus shifted toward independent research and development. The core bottleneck of this path lay in equipment, especially ASML's EUV (Extreme Ultraviolet) lithography machines.

In 2018, under prolonged pressure from the U.S. government, ASML stopped exporting EUV lithography machines to China. This decision did not follow the U.S. unilateral export control route but was instead executed through the Dutch government's export licensing system. Washington's influence was more subtle here: it did not directly regulate the Dutch company but instead used ally cooperation to extend the scope of control beyond the jurisdiction of U.S. law.

On December 18, 2020, the U.S. Department of Commerce added SMIC (Semiconductor Manufacturing International Corporation) to the Entity List. For equipment export applications targeting 10nm and below process nodes, the review standard was set to a "presumption of denial." This move cut off SMIC's last possibility of obtaining EUV while forcing it toward the alternative path of DUV (Deep Ultraviolet) multi-patterning. However, the price of the alternative path was that every technical solution attempting to bypass a single point of blockade increased physical process dependence on other controlled equipment. The bypass solutions consumed more restricted resources.

The Third Layer: The Talent and Software Channel (2020-2022)

Following the equipment blockade, the system's designers realized that two potential loopholes remained: talent mobility and software tools.

EDA (Electronic Design Automation) software is the fundamental tool for chip design. Two companies, Synopsys and Cadence, together control approximately 70% of the global EDA market. Since 2020, exports from these two companies to China have been brought under the scope of control, and license applications for advanced process design tools similarly face a "presumption of denial" review standard. Without EDA tools, chip design cannot begin; without advanced versions, the research and development of advanced process chips becomes empty talk.

The blockage of the talent channel took another form. In the October 7, 2022 rule update, there was a clause that received almost no media attention: U.S. persons (including citizens and green card holders) are prohibited from supporting the development or production of advanced semiconductor facilities in China without a license. The practical effect of this clause was that Chinese-American engineers working at advanced Chinese semiconductor companies had to choose between their jobs and their citizenship. Dozens of engineers resigned as a result.

The Fourth Layer: Ally Coordination (2022-2023)

The October 7, 2022 rule update marked the entry of the blockade system into its fourth stage: unilateral controls giving way to multilateral coordination. The coverage of U.S. unilateral export controls is limited by its legal jurisdiction, while key nodes in the semiconductor supply chain are distributed across the Netherlands (ASML), Japan (Tokyo Electron, Nikon), and South Korea (Samsung, SK Hynix).

In January 2023, the United States, the Netherlands, and Japan reached an agreement in which the Netherlands and Japan agreed to implement advanced semiconductor equipment export controls on China similar to those of the United States (Reuters, January 2023). Subsequently, the Netherlands restricted the issuance of export licenses for ASML's most advanced DUV equipment (NXT:2000i and above models) to China. Japanese equipment manufacturers such as Tokyo Electron also simultaneously tightened their export policies toward China.

The geographical loopholes of unilateral control were mended through multilateral coordination. China's path to bypassing controls through non-U.S. suppliers was almost entirely blocked after 2023.

The Cost of the System

Over seven years, this blockade system achieved full-domain coverage across capital, equipment, software, talent, and unilateral/multilateral dimensions. Every time China found a potential loophole, the system upgraded accordingly. This speed of iteration itself illustrates a point: the designers of the blockade are constantly chasing the adaptive capacity of the blockaded.

The effectiveness of the system is built on a premise: the control of key nodes by the United States and its allies is so highly concentrated that any bypass path must pass through a controlled node. Only ASML can produce EUV lithography machines, only Synopsys and Cadence can provide EDA software, and core equipment for advanced processes is concentrated in a few U.S., Dutch, and Japanese companies. This concentration stems from historical accumulation and is by no means a product of policy design; Washington is merely exploiting the existing status quo.

The costs are equally obvious. The October 7, 2022 rule update included Nvidia's H100 chips in the scope of control, and the subsequent A800/H800 alternative solutions were again included in October 2023. Every tightening compresses the Chinese market revenue of U.S. chip companies. In fiscal year 2023, Nvidia's revenue from China accounted for approximately 20% to 25% of its total revenue, a proportion that continued to decline after the controls were implemented. The designers of the blockade system are also the injured parties. In immunology, this phenomenon is called an "autoimmune disease": the defense mechanism is over-activated and begins to attack its own tissues. Antibodies cannot distinguish between foreign invaders and host cells, thus causing damage to oneself while eliminating the threat. While the U.S. export control system blockades China's chip industry, it simultaneously cuts off the Chinese market revenue of U.S. chip companies in the same manner. The broader the scope of control, the deeper the degree of self-inflicted damage.

In 2015, CFIUS's implicit veto power caused Tsinghua Unigroup's $23 billion bid to fail before it was even formally filed. In 2022, a 139-page technical document completed the final closed loop of this system. The seven years spanning these two points in time witnessed the blockade system's complete construction cycle through case accumulation, institutional establishment, multilateral expansion, and the gradual inclusion of capital, equipment, and talent.

This door was by no means slammed shut suddenly one day; it was, in fact, walled up brick by brick by human hands.

The Huawei Ban

The Watershed of May 16

On September 14, 2020, TSMC stopped supplying Huawei. On that day, the final batch of Kirin 9000 chips was stored in Huawei's climate-controlled warehouses. These 5nm process silicon wafers, manufactured by TSMC, constituted the entire inventory supporting the Mate 40 series product line. Corporate executives knew these were the last of their kind; every single chip was strictly allocated to smartphones about to roll off the assembly line, with no margin for error. This precise allocation carried the urgency of a countdown.

The mainspring of this countdown had been wound tight more than a year earlier. On May 16, 2019, the Bureau of Industry and Security (BIS) of the U.S. Department of Commerce placed Huawei and 68 of its affiliates on the Entity List. This was no ordinary trade friction penalty. The "presumption of denial" principle of the Entity List legally severed the target company's routine commercial connections with the U.S. technology ecosystem. No components or software containing U.S. technology could be exported without explicit permission from Washington. Google suspended GMS (Google Mobile Services) authorization overnight, and shipping channels from Qualcomm, Intel, and Broadcom were immediately frozen.

A commercial giant with total revenue of 721.2 billion RMB in 2018, nearly half of which came from its consumer business, instantly lost the external lifeline required to maintain its operations. The logic of interaction between the Chinese and American semiconductor industries reversed from that point forward, sliding from a market-based division of labor rooted in comparative advantage toward a technology decoupling based on national security. Before May 16, 2019, Huawei was the world's second-largest smartphone manufacturer, with annual shipments of approximately 240 million units; after that date, Huawei became a company that had to redefine its own boundaries.

Two Fates of the "Spare Tire"

The public relations narrative of "promoting the spare tire to the front line" dominated public perception in the following years, yet it obscured a critical category error. The difficulty of reconstructing software systems versus hardware silicon has never been in the same dimension.

In the world of code, HMS (Huawei Mobile Services) and the HarmonyOS operating system demonstrated an extreme case of national will replacing market choice. 2.2 million developers and 96,000 applications mobilized rapidly in 2020, and by the end of 2021, over 220 million terminal devices were running on HarmonyOS. The cost of promoting a software "spare tire" is, ultimately, the time and capital required for ecosystem reconstruction. As long as the base of end-users remains, heavy investment can build a moat.

However, this conclusion has a geographical limitation: the Chinese market. In China, Google services were already unavailable, so the replacement cost of HMS was nearly zero. In Europe, Huawei phones held about a 20% market share before 2019; after the GMS cutoff, this figure plummeted to single digits within two years. The promotion of the software spare tire was a success in the domestic market but a failure in overseas markets. The term "promoting the spare tire" conflates two vastly different outcomes.

The logic of the hardware spare tire is entirely different. No matter how sophisticated the design blueprints for Kirin chips are, they cannot bypass the constraints of physical manufacturing. When TSMC's foundry doors closed completely in September 2020, SMIC (Semiconductor Manufacturing International Corporation), the most advanced foundry in China, could only stably mass-produce at the 14nm process. The manufacturing technology of the spare tire itself was highly dependent on the supply chain strictly blockaded by Washington. Lacking core equipment such as Extreme Ultraviolet (EUV) lithography machines, chip manufacturing lost the physical foundation to evolve toward advanced processes. The definition of a "spare tire" itself became invalid: a spare tire that depends on a blockaded supply chain to be manufactured is no longer a spare tire the moment the blockade takes effect.

Two Interpretations of the Mate 60 Pro

On August 29, 2023, the Mate 60 Pro quietly went on sale without any warning. There were no spotlights, no product launch events, and not even a single line of teaser copy. This occurred exactly during U.S. Commerce Secretary Gina Raimondo's visit to China. This silence was incredibly precise, using the existence of a physical consumer product to demonstrate a physical loophole in the export control network.

A report from the teardown firm TechInsights measured the specific dimensions of this loophole. The Kirin 9000s chip inside the device was manufactured by SMIC using an N+2 multi-patterning process, reaching an equivalent node level of 7nm. This was a process breakthrough achieved by China through the stacking of Deep Ultraviolet (DUV) multi-patterning processes in the absence of EUV lithography machines.

The technical ledger is cold. Data estimates from third-party analytical firms showed that the initial yield rate of the Kirin 9000s hovered only between 50% and 60%. This indicates that for every two chips produced on the line, one was a defect that had to be discarded. In contrast, TSMC's yield for the same process node had long stabilized at over 90%. The monthly capacity of SMIC's N+2 production line was only about 40,000 to 50,000 wafers, and the actual performance of the chip was only equivalent to 70% to 80% of the Qualcomm Snapdragon 8 Gen 2.

This is reminiscent of Germany's synthetic fuel industry during World War II. After losing overseas oil fields, German engineers relied on the Fischer-Tropsch process to liquefy coal, forcibly maintaining the fuel supply for the war machine. While coal-liquefied fuel could indeed power armored divisions, its production cost was five to eight times that of natural petroleum, and its capacity ceiling was destined the moment the factory was built. SMIC's multi-patterning DUV process relative to TSMC's EUV is exactly like synthetic fuel relative to natural petroleum: it can function, but the endpoint is already visible.

"Huawei is back" and "Huawei has found its ceiling" are not contradictory; they are two sides of the same reality. The significance of the Mate 60 Pro lies in proving the existence of a path, while the specifications of this phone simultaneously mark the end of that path.

The Bill for the Ban

The audit working papers are full of drama. The original intent of the Entity List was to precisely strip Huawei of its ability to obtain advanced process chips, and this goal was indeed achieved at the first-order effect level. Huawei's annual smartphone shipments plummeted from 240 million units in 2019 to 350 million units in 2021, a staggering 81.5% decrease. Consumer business revenue was halved to 243.4 billion RMB in 2021, down 49.6% from 482.9 billion RMB in 2020.

However, the operation of the sanction machine triggered second-order effects that exceeded design expectations. To survive in a vacuum, Huawei was forced to shift its R&D focus from product innovation at the application layer deep down to foundational technologies in materials and equipment. Its R&D investment climbed from 131.7 billion RMB in 2019 to 161.5 billion RMB in 2022, accounting for 25.1% of its total revenue that year. A consumer electronics company, under forced conditions, completed a transformation into a foundational technology company.

The desperate situation of one company evolved into a forced march for the entire Chinese semiconductor industry chain. The third phase of the National Integrated Circuit Industry Investment Fund entered the fray in 2023 with a massive 344 billion RMB. SMIC's capital expenditure surged from $4.5 billion in 2021 to $6.3 billion in 2022. Combined with YMTC's (Yangtze Memory Technologies Corp) all-or-nothing gamble on the 232-layer NAND architecture, these elements formed a capital-saturated rescue mission with no retreat. These massive injections of capital, which defy conventional commercial return cycles, would have been impossible in the greenhouse of free trade.

At the cost of destroying a commercial company's terminal business, Washington's sanctions list unexpectedly initiated a reset program for a national-level semiconductor industry. An assessment report by the Information Technology and Innovation Foundation (ITIF) in Washington in October 2025 reached a grim conclusion: export controls actually helped the sanctioned targets while damaging the global market revenue share of domestic U.S. companies.

This conclusion does not suggest that the sanctions failed. The sanctions were successful in their first-order effects: Huawei's mobile business was crippled, and the path to advanced processes for Kirin chips was blocked. But the second-order effects of the sanctions exceeded expectations: Huawei's crisis was transformed into a national mobilization order. The blockade created exactly what it sought to prevent: an industrial system forced to complete a foundational closed loop. The logic of policy encountered unforeseen feedback loops within a complex system.

Lithography Machine Blockade

The Most Complex Machine of Mankind

In July 2018, Dutch Prime Minister Mark Rutte met with senior officials of the U.S. National Security Council in a White House meeting room. The U.S. side presented a highly classified intelligence report detailing the potential military and strategic implications if Chinese semiconductor manufacturers were to acquire EUV (Extreme Ultraviolet) lithography machines. Shortly after the meeting concluded, the Dutch government decided not to renew ASML's license to export EUV equipment to China. Without the mandatory constraints of international law or public threats of economic sanctions, the global flow of this equipment—the most complex industrial device in human history—underwent a major turning point due to a single intelligence report and a closed-door meeting.

To understand the technical foundation behind the blockade, one must first understand the industrial parameters of this machine. A standard NXE:3600D EUV lithography machine weighs approximately 150 tons and contains about 100,000 precision components. International transport requires three Boeing 747 cargo planes and 40 standard containers (ASML Official Website, 2022). The equipment's price is equally staggering: a standard EUV costs between $150 million and $200 million, while the new generation High-NA EUV (EXE:5000) costs as much as $380 million (ASML Financial Report, 2024). Such high technical and financial barriers exclude nearly 99% of potential global competitors.

Complexity itself is the most powerful barrier. The operation of an EUV lithography machine relies on a multinational and extremely sophisticated collaborative network. Zeiss SMT in Oberkochen, Germany, is responsible for manufacturing high-precision optical lenses; Cymer in California, USA, provides high-frequency laser light sources; Trumpf in Ditzingen, Germany, supplies the drive laser systems; and finally, ASML in Veldhoven, Netherlands, completes the atomic-level assembly of the 100,000 components. If any single link stalls, this 150-ton device becomes a pile of expensive scrap metal.

The production of nearly all advanced-process chips globally depends almost entirely on the approximately 50 machines delivered by ASML each year. This supply chain has no redundancy; it is more like a single thread that could snap at any moment. In 2018, the U.S. government precisely seized the weakest link of this single thread, transforming export restrictions on a single piece of equipment into a multi-front blockade of the entire foundation of advanced manufacturing processes.

Why EUV is a Deadlock Rather Than a Bottleneck

Simply categorizing EUV export restrictions as a technological blockade actually obscures the core logic of the policy. Technological blockades usually assume that the target country lacks knowledge and only needs to obtain enough blueprints and code to bridge the gap. However, the essence of the EUV blockade is a blockade of supply chain concentration; the accumulation time of physical manufacturing capabilities cannot be shortened by capital investment alone.

The Republic of Venice in the Middle Ages once maintained an industrial monopoly for three hundred years by concentrating all glass artisans on the island of Murano and prohibiting the transfer of technology through harsh decrees. However, this type of monopoly eventually collapsed due to the defection of core artisans and technology diffusion. ASML's monopoly model is completely different. Even with a full set of design blueprints and core engineers, no single country could immediately replicate the processing capabilities of Zeiss SMT. The extreme ultraviolet optical system provided by Zeiss requires the surface roughness of the lens to be controlled below 0.1 nanometers—equivalent to magnifying the entire territory of Germany to the size of the Earth's surface, where the difference between the highest and lowest points cannot exceed one millimeter. This atomic-level processing precision relies on decades of process experimentation and accumulation in materials science.

The simultaneous concentration of multiple nodes makes EUV the only absolute deadlock in the global chip competition. In the field of mature processes, Shanghai Micro Electronics Equipment (SMEE) reportedly developed a domestic lithography machine with processing capabilities for approximately the 28nm node in December 2023 (Lianhe Zaobao, 2023). With the help of multi-patterning technology, ASML's existing high-end DUV equipment can even reach process nodes of approximately 5nm. However, when moving toward the physical limits below 7nm, EUV becomes an unavoidable technical gatekeeper.

The formation of this deadlock stems from the unique binding of three supply chain nodes: ASML, Zeiss SMT, and Cymer. ASML holds approximately a 24.9% stake in Zeiss SMT (ASML Annual Report) and fully acquired Cymer, firmly locking originally dispersed manufacturing capabilities through capital ties. When these three nodes are completely bound geographically and financially, the challenges faced by pursuers are no longer simple engineering problems, but the ultimate challenge of rebuilding the entire modern precision manufacturing system in a vacuum. Capital can buy computing cards, but it cannot buy a shortcut to reduce the time required for lens polishing.

The Netherlands' Dilemma Ledger

Washington's strategic layout collided unexpectedly with Amsterdam's financial reality. The Dutch government gradually tightened export licenses under external pressure, which triggered a chain reaction contrary to the policy's original intent: Chinese semiconductor manufacturers, in response to the imminent risk of supply disruption, launched cost-insensitive defensive procurement.

Panic bred demand. According to data from TrendForce and ASML's Q4 2023 financial report, mainland China accounted for only 14% of ASML's total revenue in 2022. However, by 2023, as rumors of tighter controls intensified, this proportion climbed rapidly to 29%, reaching a historical peak of 46% in the third quarter of that year. In 2023 alone, Chinese customers contributed approximately 9 billion euros in revenue to ASML. In the short term, the export control policy did not cut off the cash flow; instead, it became the strongest sales catalyst, concentrating potential orders from several years into the present.

However, the other side of the ledger is filled with long-term losses. The hoarding effect is essentially a one-time event. On January 1, 2024, the Dutch government officially revoked the export licenses for some of ASML's high-end DUV lithography machines (NXT:2050i and NXT:2100i) to China (ASML Official Statement, 2024). As policies tighten across multiple lines, the previously overdrawn market demand is rapidly drying up, and the precipitous drop in revenue share from the China region has become an irreversible financial reality.

Former ASML CEO Peter Wennink repeatedly warned that cutting off equipment supply to specific markets would not only weaken the company's R&D funding cycle but also prompt the blockaded party to accelerate the establishment of an independent ecosystem. Europe's most valuable technology company is trapped in an inescapable predicament: short-term financial losses were mitigated by panic buying from Chinese customers, but the long-term market shrinkage will be borne solely by Dutch shareholders. The United States formulates the strategy, while the balance sheets of its allies pay the bill.

The End of Technological Neutrality

The U.S. government's choice in 2018 to apply pressure through White House diplomacy rather than Commerce Department bans exposed a key weakness in the existing export control system. Compliance auditors from the U.S. Department of Commerce, after evaluating ASML's EUV system, confirmed that the value proportion of components in the machine subject to direct U.S. jurisdiction was consistently below the 25% minimum threshold (optics.org, 2020).

Legal jurisdiction thus failed. Washington could not invoke existing long-arm jurisdiction legal tools (such as the Foreign Direct Product Rule) to block equipment exports. When the U.S. technology content in a globalized supply chain is diluted below the legal standard, the superpower loses the basis for intervention through legal means. The U.S. could not legally stop the equipment from being shipped; it could only deliver an intelligence report that remains undisclosed to this day.

When a major power provides intelligence and exerts diplomatic pressure, allied nations at key technological nodes effectively have no choice. When the Dutch Trade Minister announced further restrictions on semiconductor equipment exports in 2023, she explicitly mentioned "concerns based on military end-use." This statement completely stripped away the lithography machine's disguise as a commercial product, redefining it as a strategic asset with geopolitical significance. The Netherlands eventually abandoned its long-held position of commercial neutrality, prioritizing national security alliances over free trade principles.

The shift from legal enforcement to political lobbying has become the most fragile link in the entire EUV blockade network. Dutch cooperation depends on a specific political climate and alliance commitments rather than the rigid constraints of international law. A blockade line maintained by intelligence sharing and high-level lobbying is only as strong as the political will of the allied governments. Once the interest assessment in The Hague changes, or if Europe's pursuit of "technological sovereignty" exceeds its dependence on the Transatlantic alliance, this blockade network could develop irreparable cracks from within.

AI Chip Battlefield

Computing Power Ladders and the Game of Rules

On August 26, 2022, the U.S. Department of Commerce cut off export channels for the NVIDIA A100 and H100 to China. NVIDIA subsequently disclosed a current risk exposure of approximately $400 million in its financial report (NVIDIA Financial Report, 2022). Regulation always lags behind technology. While policymakers attempted to limit hardware performance through total computing power thresholds, chipmakers achieved compliance by finely tuning interconnect bandwidth. These technical compromises forced Washington to introduce performance density metrics to close policy loopholes created by computing power clustering. On October 17, 2023, the strategy of multi-chip parallelism was completely blocked (BIS Regulations, 2023).

Every escalation of regulatory metrics closely followed NVIDIA's product updates. From the A800 to the H800, and then to the H20, the ingenuity of Silicon Valley engineers was repeatedly consumed by how to legally degrade chip performance. Until April 15, 2025, the export channel for the H20 was completely closed (GamerNexus Timeline, 2025). Regulators successfully blocked hardware exports, but the cost was a $5.5 billion loss on the balance sheets of domestic core enterprises (NVIDIA Financial Report, 2025).

The cost of this game is not only reflected in NVIDIA's financial statements. Each regulatory escalation prompted Chinese customers to rush-buy chips before bans took effect, creating a wave of stockpiling. Before the H800 ban in October 2023, Chinese data center operators and AI companies procured H800s on a massive scale; this inventory became the primary source of computing power for Chinese AI training over the following two years. Export controls created a shortage of computing power in the short term, but in the medium term, they bought buffer time for China's AI development.

The contest on the computing power ladder is not just a technological competition, but also a struggle for the power to define rules. Whoever can more quickly translate algorithms into regulatory metrics holds the initiative in compliance. Washington draws red lines between national security and economic interests, while Silicon Valley tests physical limits at the edge of those red lines. These passive policy responses ultimately turned export controls into a costly and self-consuming cat-and-mouse game.

CUDA's Moat

Competition in hardware parameters is explicit, but the true barriers are hidden within the software ecosystem. Confining the focus of AI chips to the number of floating-point operations in FP16 or INT8 is an over-fixation on parameters. The real advantage lies in those invisible software invocation logics. As of 2024, CUDA's share in the AI training market stabilized at approximately 90% (LinkedIn Analysis, 2024). This monopoly position does not stem from the performance of any single generation of GPU, but is the result of accumulated codebases by developers, engineering habits formed after countless debuggings, and a third-party library ecosystem jointly maintained by global researchers. These factors have turned the computing power market into a one-way revolving door.

Attributing the shortage of computing power simply to hardware backwardness, while ignoring the ecosystem inertia built by codebases and developer habits, is the greatest misunderstanding.

The platform lock-in achieved by Microsoft Windows in the PC era is now repeating in the field of AI computing power. Although Linux possesses replacement capabilities at the system kernel level, high software migration and employee training costs lead enterprises to still choose to stick with Windows. CUDA's lock-in mechanism is even more stringent than that of an operating system. Low-level optimizations for AI training are deeply embedded in the CUDA architecture. Even if Huawei's Ascend 910B can reach approximately 60% to 70% of the H100's peak throughput in specific matrix operations (Georgetown CSET Report, 2024), requiring an AI team to migrate hundreds of thousands of lines of code to the CANN or ROCm platforms is almost equivalent to tearing everything down and starting over.

Currently, AMD ROCm occupies only 5% to 8% of the market share, and Chinese domestic alternatives collectively account for less than 2% (Industry Report, 2024). Abandoning CUDA does not just mean a 20% to 30% performance loss; it means detaching from the global open-source AI ecosystem. The caution of commercial companies regarding computing power migration is the best proof that ecosystem appeal far exceeds the performance of a single chip.

HBM: The Overlooked Ceiling

The realization of computing power is inseparable from sufficient data support. While public attention is focused on GPU codenames like A100 or B200, the bottleneck that truly determines whether large models can run efficiently is actually those vertically stacked memories that sit close to the computing units and transmit data at extremely high throughput. High Bandwidth Memory (HBM) has become this hidden ceiling.

In a highly concentrated memory supply chain, the physical limits of the memory wall are more destructive than the policy thresholds of the computing power wall.

In 2024, the global HBM market exhibited an extremely concentrated oligopolistic structure. SK Hynix holds approximately 62% of the market share, Samsung controls 25% to 30%, and Micron takes the remaining 8% to 10% (AstuteGroup, 2026; SK Hynix Annual Report, 2024). In contrast, China's mass production capacity in the HBM field is almost zero. The production capacity of domestic enterprises such as CXMT remains stuck at the DDR4 and LPDDR5 stages (Industry Analysis, 2024). What is throttling the development of China's large AI models is not just Washington, but also the supply chain giants in Seoul.

The high concentration of the aforementioned supply chain significantly impacts hardware performance. The HBM bandwidth of Huawei's Ascend 910B is approximately 900 GB/s, which is only 27% of the NVIDIA H100's 3.35 TB/s (Georgetown CSET Report, 2024). The gap in computing power might only be 30% to 40%, but the disparity in data transmission channels is as high as over 70%. Without a sufficiently wide data channel, even the most powerful computing units can only idle. In December 2024, export controls on HBM were further expanded, directly targeting this hidden bottleneck. Washington realized that blocking GPUs without blocking HBM was equivalent to closing the front door while leaving the back door open.

The Boundaries of the DeepSeek Effect

Miracles also have their physical limits. In December 2024, the DeepSeek V3 model shocked Silicon Valley with a low training cost of $5.576 million. The development team rented 2,048 H800 GPUs and spent 55 days completing the training of a 671-billion-parameter MoE architecture (DeepSeek Official Technical Report, 2024). In comparison, the estimated cost for OpenAI to train GPT-4 is as high as $100 million. At 1/20th of the cost with comparable performance, this gap was interpreted as an illustration that "algorithmic efficiency can substitute for computing power."

Such interpretations have some merit. The DeepSeek team utilized the sparse activation characteristics of the Mixture-of-Experts (MoE) architecture to achieve activation of only 37 billion parameters at a time within a massive parameter pool. FP8 mixed-precision training achieved efficiency on the H800 close to that of the H100. Multi-head Latent Attention (MLA) reduced the memory footprint of the KV cache. These technical innovations truly exist and are not mere gimmicks.

Export controls force those restricted to perform extreme optimizations in algorithmic efficiency, but such innovations based on existing hardware will eventually hit a physical ceiling when inventories are exhausted.

Behind this set of impressive financial data lies a cold reality. The training of DeepSeek V3 relied on the H800, not the Ascend 910B. That batch of hardware consisted of existing assets rushed into inventory before the performance density ban took effect in October 2023. Improvements in algorithmic efficiency partially compensated for the computing power gap, but by no means did they completely eliminate it. When the lifespan of these H800s is exhausted, the training of next-generation V3-level models will inevitably face a true computing power vacuum.

The true significance of the DeepSeek effect is that Chinese AI teams have demonstrated algorithmic innovation capabilities that exceeded expectations under restricted conditions. However, the premise of "exceeding expectations" is "restricted conditions," and these conditions are gradually tightening as H800 inventories are consumed. The $5.57 million miracle is a miracle with an expiration date.

The three main threads of the AI chip battlefield—computing power regulation, the CUDA ecosystem, and the HBM supply chain—collectively point to one conclusion: hardware parameters are explicit, while ecosystem inertia and supply chain concentration are implicit, and implicit obstacles are often harder to overcome. China's innovation in algorithmic efficiency is a fact, but there are physical limits to the improvement of algorithmic efficiency. When H800 inventories are exhausted, the HBM supply chain remains difficult to break through, and the migration costs of the CUDA ecosystem remain high, the chasm between "partial compensation" and "eliminating the gap" will mark the true boundaries of this game more clearly than any export control document.

American Reshoring: Sovereignty Vacuums and the Subsidy Paradox

On December 2, 2024, the Intel Board of Directors announced the "retirement" of CEO Pat Gelsinger. On the day the news was released, the company's stock price rose against the market trend. The capital market made no secret of viewing the departure of this technical veteran as a significant positive. Just three weeks prior, the U.S. Department of Commerce had just allocated a $7.86 billion CHIPS Act subsidy to Intel (U.S. Department of Commerce, November 2024). This is the highest single manufacturing incentive grant in the entire $52.7 billion act.

The market's reaction itself sends a signal: America's largest chip company, at the moment the nation needs it most, is already difficult to save.

The Price of $52.7 Billion

In this allocation list, which includes $39 billion in manufacturing incentives and $13.2 billion in R&D funding, TSMC received $6.6 billion, Samsung received $6.6 billion, and Micron received $6.14 billion. The manufacturing capacity that the United States is attempting to rebuild through massive fiscal spending is precisely the part that Silicon Valley actively outsourced to Asia over the past thirty years based on market logic.

Washington is paying a high price for this reversal. The U.S. is using taxpayer money to subsidize the local capacity expansion of foreign companies without truly regaining technological dominance. Fiscal checks can pour the concrete foundations of cleanrooms, but they cannot rebuild the supply chain ecosystems and the underlying networks of engineers that were lost along with industrial transfer.

In this expensive investment promotion, Intel, which took the most taxpayer funds, became the only one among the four major beneficiaries whose core foundry business is deeply mired in losses. This result is by no means accidental; it is an inevitable manifestation of deep-seated industrial problems.

Intel's Dilemma and the Subsidy Paradox

If one looks at a longer timeline, Intel's trajectory of decline is clearly visible. Throughout 2024, Intel's stock price fell by more than 60%, and its total market capitalization dropped below $100 billion for the first time since 2012 (Fortune, December 2024). Its wafer foundry business lost as much as $7 billion in the year 2023 alone. Its Q2 2024 financial report fell below market expectations, triggering a single-day stock price crash of over 25%, the largest single-day drop since 1974 (Reuters, April 2024).

This $7.86 billion government grant came with a key provision: Washington required that Intel must not sell its wafer foundry business as a hard condition for receiving the subsidy. The original intention of the clause was to ensure the integrity of domestic manufacturing capabilities. However, the consequence of this requirement is that the massive grant effectively deprived the enterprise of the option to "cut off a limb to save the body."

An industrial policy aimed at reviving domestic manufacturing has ultimately locked a former chip giant into a long-term loss-making business line. The $7.86 billion has become a political justification for maintaining losses, and Gelsinger, who left with a severance package exceeding $10 million, leaves behind a financial black hole distorted by policy leverage (Reuters, December 2, 2024).

Government subsidies and commercial laws have come into sharp conflict here. The subsidies failed to stimulate the company's drive to turn losses into profits; they merely extended the survival time of a giant.

TSMC Arizona: Replicable Factories, Non-transplantable Culture

In the desert heartland of Phoenix, Arizona, the operation of TSMC's Fab 21 depends on a fragile demographic structure. As of the end of 2024, as many as 50% of the approximately 2,200 employees at this factory were dispatched from Taiwan across the ocean (9to5Mac, December 2024).

This data punctures the illusion of the omnipotence of capital expenditure. While equipment procurement accounts for more than 70% of the total cost, and wafer processing costs in Arizona are only about 10% higher than in Taiwan (TechInsights, 2025), the real difficulty arises in the daily operation of the cleanrooms. On the workplace review website Glassdoor, TSMC's U.S. factory has a rating of only 3.2, far lower than Intel's 4.1. In a country that values work-life balance, the militarized management, 24-hour shift system, and high demands for absolute obedience of this Asian foundry giant have encountered serious cultural adaptation problems.

Such adaptation difficulties are not rare in industrial history. In the 1980s, Toyota Motor established a factory in Georgetown, Kentucky, and it took a full decade to barely embed the "Toyota Production System" into the work habits of American workers. The fault tolerance of semiconductor manufacturing is several orders of magnitude more stringent than that of automobile assembly. At advanced process nodes on the nanometer scale, any communication error during a shift change or a minor deviation from standard operating procedures can lead to the scrapping of an entire batch of wafers. In Taiwan, TSMC relies on a team of engineers who can respond to production line alarms in the middle of the night at any time.

The 4nm process at Fab 21 was originally planned for mass production in 2024, but was ultimately delayed until January 2025 to barely launch (Reuters, January 2025). Toyota took ten years to cross the Pacific cultural divide; TSMC faces an even steeper wall.

Japan's Rapidus and Europe's ESMC: Different Ways of Betting

Policymakers on both sides of the Atlantic and the western shores of the Pacific are betting based on their respective industrial anxieties.

The Japanese government has invested more than 7 trillion yen (approximately $47 billion) into Rapidus, attempting to achieve mass production of 2nm advanced processes by 2027 through a generational leap (Nikkei Asia, 2025). Rapidus was established in 2022, jointly funded by eight Japanese companies including Toyota, Sony, and SoftBank, and is collaborating with the IBM Albany Nanotechnology Center to develop processes. This move can be described as extremely risky; building a 2nm production line directly without the accumulation of yield data from previous generations is equivalent to skipping the entire learning curve.

Europe has chosen a different strategy. The ESMC project located in Dresden, Germany, has a total investment of about 10 billion euros, including 5 billion euros in targeted subsidies approved by the EU, and broke ground in August 2024. This factory has completely abandoned the pursuit of advanced processes, clearly targeting the 28nm mature process. It is expected to go into production by the end of 2027, specifically to provide a stable local supply chain for Europe's massive automotive industry (Reuters, August 2024).

High-risk quantum leaps and low-risk defense of the bottom line are the two extremes of the global semiconductor landscape's reconstruction. An ironic time lag exists between these plans: TSMC's factory in Kumamoto, Japan, officially went into production in February 2024. This factory started nearly a year later than the Arizona project but delivered products earlier. Administrative efficiency, labor union power, and the cooperation of the local supply chain have shown to be more important in actual construction than the amount of subsidies.

The Real Cost of Re-industrialization

Washington must answer a core question: what is the true goal of this multi-billion dollar re-industrialization?

If the ultimate demand is "manufacturing semiconductors on U.S. soil," the massive fiscal expenditure is indeed slowly fulfilling that promise. If the goal is "American companies mastering the dominance of advanced processes," this act is accelerating toward failure. The planned annual capacity of 600,000 wafers for Arizona's Fab 21 accounts for only 3.75% of TSMC's global annual capacity of 16 million wafers. After these 3.75% of wafers are produced in the desert, the specialty chemicals, high-purity electronic gases, and precision equipment components behind them still rely on a vast global network spread across East Asia and Europe.

The confusion between "geographic sovereignty" and "corporate sovereignty" has become the core loophole in the U.S. semiconductor strategy. Forcing TSMC and Samsung into the country has indeed enhanced supply chain resilience under extreme geopolitical conflict. The price is that the U.S. has personally used taxpayer funds to support the strongest competitors of its domestic chip champions.

The engineer culture and high-density industrial clusters that TSMC spent forty years building on the west coast of Taiwan cannot be transplanted through subsidies. The atrophy of manufacturing capacity takes decades, and rebuilding it also takes decades. Subsidies can only compress financial costs; they cannot shorten time costs.

Intel's decline is the final footnote to this logic. America's largest chip company falls just when the nation needs it most, and it is TSMC and Samsung that fill the void. The CHIPS Act bought factories, but it could not buy the ecosystem that makes the factories run.

China's Limits and Counterattacks

On December 3, 2024, China's Ministry of Commerce announced a total ban on the export of gallium, germanium, and antimony to the United States. While the names of these three metals may be unfamiliar to most, they are the core materials for semiconductors, fiber optics, and military infrared detectors. China controls approximately 80% of global gallium production and 60% of germanium production (USGS, CSIS, 2023-2024). This ban is a response to the new round of US export controls in October 2024. A nano-scale technological blockade has gradually evolved into a ton-scale mineral counter-measure.

The 7nm Ceiling: Physical Limits and Yield Dilemmas

The physical chasm between 13.5 nanometers and 193 nanometers cannot be crossed by administrative decree. In the absence of EUV (Extreme Ultraviolet) lithography machines, SMIC (Semiconductor Manufacturing International Corporation) has attempted to forcibly break through the advanced process threshold using DUV (Deep Ultraviolet) multi-patterning technology. In engineering terms, this is akin to carving a strand of hair with a magnifying glass; it is theoretically possible but extremely costly in practice.

In February 2025, the yield rate of Huawei's Ascend 910C chip, which utilizes SMIC's N+2 (5nm) process, barely reached 40% and began to achieve slim profitability (Digitimes, February 2025). In contrast, TSMC's yield for the same node remains stable at over 80%. Behind these figures lies immense cost pressure: the manufacturing cost of a SMIC 5nm wafer is approximately 50% higher than that of TSMC (TrendForce, March 2025). The earlier generation 7nm (N+1) process is in a similar predicament, with a yield rate less than one-third of TSMC's equivalent node, yet priced 40% to 50% higher (Asia Times, February 2024).

A 30% yield rate means that for every ten expensive silicon wafers produced, seven become industrial waste. China's breakthroughs in advanced processes resemble a political mission more than a commercially sustainable market activity. The physical limitations of DUV quadruple patterning dictate that the mass production of processes below 7nm is destined to be a loss-making war of attrition.

The significance of this ceiling extends far beyond the technical level. It indicates that China's chip strategy must seek three flanking paths outside of a "frontal breakthrough": dumping mature processes, bypassing via the RISC-V architecture, and critical mineral counter-measures. Each of these three paths has its own logic and limitations.

Mature Process Dumping: The Strategic Logic of Quantity over Quality

The price of a $1,500 silicon carbide (SiC) wafer plummeted to $500 in just a few months. This was not a cost reduction brought about by technological progress, but rather the result of Chinese competitors forcibly changing pricing rules by relying on government subsidies. This move caused the market value of Wolfspeed, a veteran American wide-bandgap semiconductor manufacturer, to evaporate by 96% (TechPowerUp, 2024).

Market share is shifting rapidly. In 2024, China accounted for approximately 22% of global mature process (28nm and above) production, a figure expected to rise to 28% by the fourth quarter of 2025 and reach 39% by 2027 (TrendForce, TechPowerUp, October 2024).

The strategy of substituting quality with quantity bears similarities to the impact of the Japanese steel and automobile industries on US domestic manufacturing in the 1970s and 1980s: driving down end-product prices through asymmetric capital investment, weakening the profitability of Western counterparts, while providing low-cost foundry services for a large number of domestic fabless design companies. The only difference is that the national security attributes of semiconductors far exceed those of cold-rolled steel plates, and thus the speed of Western intervention is also faster. However, the efficiency of policy approvals consistently fails to catch up with the speed of wafers leaving the factory.

The strategic value of mature process dumping lies in the fact that this path satisfies approximately 80% of China's domestic chip demand (automotive, industrial, consumer electronics) while driving down global prices through overcapacity, thereby weakening the profitability of Western mature process manufacturers. This is a price war that can be fought without advanced processes.

RISC-V: The Real Boundaries of the Detour

Breaking free from the monopoly of ARM and x86 architectures has long been seen as a shortcut to achieving technological autonomy. As an open-source instruction set, the RISC-V standard itself is not subject to any form of export control; the US government cannot restrict an open-source standard. In 2023, the Chinese market contributed approximately 10 billion RISC-V chip shipments (RISC-V International, 2024).

The figure of ten billion seems highly disruptive. However, statistical breakdowns show that the vast majority of these chips are used in fields such as smart door locks, electric toothbrushes, and low-power sensors. In the fields of high-performance computing and AI servers, which truly determine the computing power landscape, the performance gap between RISC-V and ARM remains at two to three generations.

Alibaba's T-Head (Pingtouge) released the C930 RISC-V chip for AI applications in February 2025 and plans to spin it off for an IPO in 2026 (HPC Wire, February 2025; Caixin Global, January 2026). This is a signal of RISC-V's march into the high-performance field, but the weakness of the underlying ecosystem remains unavoidable: the adaptation of compilers, operating systems, and AI frameworks for RISC-V requires years of accumulation.

The detour path is indeed feasible, but its destination is the edge of the Internet of Things (IoT), not cloud data centers. RISC-V can solve the autonomy issue for embedded chips but cannot satisfy the core requirements of AI training chips.

The Double-Edged Sword of Critical Minerals

The comprehensive ban on December 3, 2024, was originally designed as a strategic weapon to precisely strike Western military-industrial and semiconductor supply chains. However, the supply disruption did not truly materialize. During the period the ban was in effect, global semiconductor production lines did not experience substantial shutdowns; raw materials continued to flow to Western wafer fabs through complex third-country entrepôt trade. Spot prices surged by more than 50% within weeks, yet supply remained operational.

A more critical second-order effect was that the artificially created scarcity activated dormant alternative supply chains. Abandoned mines in Canada and Australia received new capital injections, and new purification production lines were accelerated under government guarantees. Just 11 months later, on November 9, 2025, this comprehensive export ban was quietly suspended (Reuters, December 3, 2024; CNBC, November 9, 2025).

The lifespan of this weapon was shorter than the life cycle of a single generation of smartphones. While it pushed up competitors' procurement costs in the short term, in the long term, it provided the strongest commercial incentive for competitors to rid themselves of dependency. The gallium and germanium controls of July 2023 had already triggered the construction of alternative supply chains in the West; the comprehensive ban in December 2024 merely accelerated this process.

The Boundary of "Good Enough": A Moving Target in the AI Era

Compromises in technical routes are often hidden within the pragmatic narrative of being "good enough." In 2020, when smartphones and basic cloud computing dominated demand, the 7nm process was sufficient for 90% of commercial application scenarios.

The explosion of generative AI has shattered this static supply-demand balance. In 2024, the size of China's AI chip market was approximately $10 billion, and it is expected to exceed $30 billion by 2027. When the most cutting-edge AI training models must rely on the stacking of 3nm processes and HBM (High Bandwidth Memory), 7nm chips have been reduced to expensive waste-heat generators on the training side, barely able to support edge inference tasks. The Huawei Ascend 910C, using the N+2 process, can only reach about 60% to 70% of the peak computing power of the Nvidia H100.

"Good enough" has never been a fixed technical standard, but rather a target that changes rapidly with the demands of AI applications. Facing export controls, the real risk does not lie in the lack of a few EUV lithography machines today. In an era of computing power inflation, yesterday's "just right" quickly becomes today's "completely unusable."

The three flanking paths—mature process dumping, the RISC-V detour, and mineral counter-measures—are all realistically feasible strategic options. However, their common limitation is that they cannot satisfy the core demand for advanced processes in the AI era. The boundary of "good enough" is narrowing, and the speed of this narrowing is faster than the progress of any flanking path. China's counterattack is real, and the cost is also real: in an era where AI computing power has become a strategic resource, the end of the flanking maneuver remains that 7nm ceiling.

Three Futures

In 2024, TSMC's fabrication plant in Arizona finally began producing 4nm advanced process silicon wafers. In the same year, Huawei's Ascend 910C began large-scale supply to domestic AI developers at a price lower than the Nvidia H100. These two commercial events, occurring on opposite sides of the planet, outline a gradually hardening reality: the global semiconductor industry is splitting into two incompatible parallel ecosystems. This physical-level split is occurring faster than anticipated in any policy planning documents from Washington or Beijing. Consequently, three possibilities for 2035 are gradually emerging.

Parallel Systems: The Most Likely Future

Washington's export controls were originally intended as precise decoupling, but reality has driven the multi-line replication of infrastructure. The "50% Rule" introduced on December 30, 2025, requires domestic chip manufacturers to use more than half domestically produced equipment (SiliconAngle, 2025). From 2024 to 2026, driven by policy, China's self-sufficiency rate for semiconductor equipment has moved from a baseline level of 30-35% (Flanders-China, 2024) toward a 70% target for 2027 (TrendForce, February 2026). This 70% target has been postponed at least twice over the past five years, gradually shifting from an engineering plan into a macroeconomic forecast characterized by wishful thinking.

Both sides have sufficient reasons to maintain the current parallel state. The United States needs time to convert subsidies from the CHIPS and Science Act into domestic capacity; China needs time to absorb the high costs of yield improvement within its domestic supply chain. The parallel system is not a static endpoint, but a dynamic transition period masking technological compensation.

When Huawei's Ascend ecosystem and Nvidia's CUDA software stack become completely incompatible at the architectural level, the definition of "parallel" undergoes a qualitative change (Industry Analysis, 2024-2025). It evolves from "domestic alternatives existing for reliance on Western exports" to "achieving basic physical autonomy while facing a massive ecological divide." The true test is no longer whether domestic fabs can etch 5nm transistors using multi-patterning technology, but which ecosystem third-party countries in the Middle East, Southeast Asia, or Latin America will choose to build their national digital infrastructure upon once two parallel underlying computing architectures take shape. This is no longer a contest of individual chip performance, but a struggle between two completely independent supply chains for the remaining global market.

Tech Cold War: Taiwan is the Flashpoint

The export controls of October 2022 defined computing power thresholds, the 2023 controls expanded the scope of equipment embargoes, and the 2024 measures cut off cloud computing rental channels. By February 11, 2026, U.S. members of Congress pushed for further restrictions on China's access to all levels of chipmaking tools (Reuters, 2026). Policy tightening exhibits a mechanical ratchet effect.

This pattern of incremental escalation is highly similar to the trajectory of the U.S.-Soviet arms race in the 1970s. Both sides invest massive capital in a field that the other cannot completely blockade, and the final result is not the absolute victory of one side, but a massive loss of economic efficiency shared by both. The Soviet arms race eventually dragged down its domestic heavy industrial economic cycle, while the cost of the modern chip race is the disappearance of thirty years of specialization dividends in the global semiconductor value chain. The difference is that the chip sector lacks the physical deterrent floor of "Mutually Assured Destruction" found with nuclear weapons. The threshold for escalation is low, and the reasons to stop are even fewer.

Within this chain sliding toward a tech cold war, the situation in Taiwan is the only external variable capable of detonating the parallel system into a multi-front cold war in a short period. According to TrendForce statistics in 2025, TSMC occupies approximately 90% of the global foundry market share for advanced processes at 5nm and below. In other words, the world's most important digital infrastructure foundation is highly concentrated at the core of a geopolitical fault line. A conflict in Taiwan would by no means be a regional geopolitical friction, but a fragile single point of failure without redundant backup in the global semiconductor supply chain. Should such physical concentration encounter a geopolitical shock, the result would not be price fluctuations, but an instantaneous stagnation of global AI computing power expansion. Any prediction regarding the global technological landscape in 2035 that ignores the Taiwan risk premium is merely a meaningless numbers game.

Limited Cooperation: The Mathematics of Trust Deficits

The history of the Cold War era is not without precedents for compromise between hostile superpowers. The 1968 Nuclear Non-Proliferation Treaty and the 1975 Apollo-Soyuz docking in Earth orbit proved that even in an era where nuclear missiles were aimed at each other, technical cooperation remained possible (Historical Records).

However, applying Cold War optimism to the semiconductor field ignores a fatal mathematical logic. The prerequisite for cooperation is reciprocal deterrence or reciprocal benefit, and the technological gap directly destroys this reciprocity. According to technical assessments by CSIS and SemiAnalysis in 2024, China lags behind TSMC by approximately 5 to 7 years in advanced process nodes at 3nm and below. When such a massive technological gap exists, any substantial technology sharing or capacity swapping objectively means that the party in the catching-up position will receive disproportionate technological dividends. Washington's domestic political environment simply cannot accept such one-way transfers of interest.

This creates a reinforcing cycle: the larger the technological gap, the lower the political feasibility of establishing reciprocal cooperation; the more cooperation channels are blocked, the more the disadvantaged party must rely on costly domestic trial and error to narrow the gap. The list of external conditions required for limited cooperation sounds more like a mission statement destined for failure: it requires absolute stability in the Taiwan Strait, a simultaneous shift in leadership on both the U.S. and Chinese sides toward those obsessed with globalization, and perhaps even a global climate disaster that forces both sides to share computing power to reset the game framework.

Key Variables: Who Can Change the Probabilities

The probability distribution of these three scenarios is not irreversible. Behind each direction lie physical constraints that are not fully controlled by any single sovereign state. When the elasticity of political will is compressed to its limit, the rigid constraints of natural resources begin to dominate the situation.

Climate change and resource depletion are becoming unexpected entry points for breaking policy deadlocks. TSMC's 2023 Sustainability Report shows that its annual electricity consumption reached approximately 23 billion kWh, accounting for about 8% of Taiwan's total power supply, while its annual water consumption reached 170 million tons. With the exponential growth in demand for computing power for AI model training, global data center energy consumption is expected to double by 2030. The extreme consumption of physical resources is already showing hard boundaries within the United States: increasingly stringent water quotas in Arizona are pushing TSMC's overseas factory expansion plans into a corner.

No matter how large Washington's export control budget is, or how strict Beijing's equipment localization mandates are, neither side can manufacture silicon wafers in a vacuum. The staggering energy and water consumption of advanced process chips is a physics problem that transcends ideology. When the tipping point of resource depletion arrives before the tipping point of technological breakthrough, extreme environmental pressure may force the two parallel ecosystems to find small areas of convergence in decarbonization technology or energy management protocols. The semiconductor map of 2035 may ultimately be determined not by Pentagon bans or Ministry of Industry and Information Technology industrial funds, but by whose power grid and reservoirs can longer support this massive Tower of Babel built of silicon and light.

Conclusion

In 1986, Japanese semiconductor companies held more than 50% of the global market share, with six of the top ten companies hailing from Japan (Network World 2024). That same year, the U.S. government signed the "U.S.-Japan Semiconductor Agreement" with Japan, mandating that foreign chips occupy at least 20% of the market share. Technological superiority failed to withstand the power of political intervention. This was the earliest script of the chip wars. Thirty-eight years later, a similar logic is playing out once again.

The Ultimate Paradox of the Blockade

An export control budget costing tens of billions of dollars ultimately resulted in an open-source model report with a training cost of only $5.57 million. When Washington attempted to limit China's artificial intelligence development by cutting off the supply chain for H100 clusters, the decision actually altered the R&D cost structure of China's AI industry, forcing companies to pursue extreme algorithm optimization under limited computing power conditions. DeepSeek successfully launched the V3 model, which shook Silicon Valley, using only 2.788 million GPU hours of H800 computing power (DeepSeek Technical Report, January 2025).

The blockade, instead, catalyzed breakthroughs. The secondary effects of export controls offset the primary effects, inversely pushing toward the opposite of the policy objectives and facilitating the very achievements Washington sought to contain.

The contrast in policy effects is equally evident at the other end of the supply chain. In the second quarter of 2025, TSMC's share of the global foundry market reached 71%, and it held an absolute advantage of approximately 90% in the field of advanced process nodes below 5nm (TrendForce 2025). Such a highly concentrated market structure indicates that the United States has not truly mastered the "technological sovereignty" it claims to defend.

In December 2024, Intel CEO Pat Gelsinger was dismissed by the board. This company, which once symbolized the glory of American manufacturing, has completely lost momentum in the field of advanced process nodes. The $52.7 billion allocated by the CHIPS Act (U.S. Department of Commerce 2022), combined with the approximately 10% cost premium of TSMC's plant construction in Arizona (TechInsights 2025), has yielded only a partial transfer of production capacity from allies. Technological sovereignty remains an empty slogan.

The Dilemma of the Catch-up: Three Historical Cases

Technological advantage is not an amulet. Every catch-up player believes they have found a permanent equilibrium point until Washington's intervention or a competitor's dumping shatters the illusion.

In the 1980s, Japan pushed American companies to the brink in the DRAM market with extremely high yield rates and quality control. However, the price of reaching the summit was the "U.S.-Japan Semiconductor Agreement" and decades of suppression. By 2024, not a single Japanese company remained among the global top ten in semiconductors (Network World 2024).

South Korea's Samsung, which only entered the memory field in 1983, surpassed Japan across multiple lines by launching the world's first 64Mb DRAM in 1992 through massive counter-cyclical investments of state capital (SemiWiki, Samsung Official History). This leap was reflected in breakthroughs at technical nodes and also stemmed from a precise grasp of the downgrading of end-market demand—replacing "extreme excellence" with "good enough," successfully occupying the market space abandoned by Japanese companies after price protections vanished. This capital-intensive expansion model allowed South Korea to dominate the memory field for thirty years, but today it faces fierce price competition from China's mature process capacity.

Taiwan, meanwhile, concentrated all its efforts on developing the specialized foundry model in the 2000s. The strategy of abandoning branding and not competing with customers won the trust of global chip design companies. This highly specialized model eventually made Taiwan a key pawn in the geopolitical game.

The success of a catch-up player will always, at some point, trigger a political reaction from the one being caught. The inevitable laws of power politics dominate the situation, while the natural results of technological competition take a back seat. Every shift in the industrial center of gravity redistributes the global power landscape and inevitably triggers a forceful counterattack from the established powers.

Efficiency and Security: An Unsolvable Equation

The semiconductor industry is undergoing a forced adjustment that defies the laws of physics. The most efficient nuclear power plants are often centralized, large, and specialized, but this structure also brings the most fatal vulnerability—difficult to manage once a single point of failure occurs. Conversely, the safest nuclear power plant designs tend toward being decentralized, small, and redundant, a structure that comes at the cost of high construction expenses and low operational efficiency.

Physical constraints transcend policy choices. The logic of the semiconductor supply chain follows the same design principles as a nuclear power plant.

Major global economies are attempting to push an impossible goal through policy. The U.S. government allocated $52.7 billion in subsidies, the Japanese government injected over 7 trillion yen (approx. $47 billion) into Rapidus (Nikkei Asia 2025), and Europe launched a 10-billion-euro ESMC subsidy plan (Reuters 2024). These massive capital injections are, in fact, paying a "security premium" for "decentralization and redundancy."

According to the comparative advantage of global division of labor, semiconductor plant construction and operation costs could be reduced by 30% to 50%. However, forcibly dismantling production capacity originally concentrated in East Asia and dispersing it globally inevitably leads to a significant decline in overall industrial efficiency. The conclusion that "efficiency and security cannot coexist" is a heavy price to pay for the efforts of various governments spending tens of billions of dollars to prove the opposite. No engineering solution can simultaneously maximize both efficiency and security.

China's Fourth Path

Historical experience has become invalid. When efficiency and security are in complete opposition, the three paths of the past have all been blocked by geopolitical rifts.

Japan's reliance on high-quality manufacturing was predicated on an open global market. South Korea's foundation for counter-cyclical expansion through massive funding was the unimpeded acquisition of the most advanced equipment, such as EUV lithography machines. The success of Taiwan's extremely specialized foundry model relied on the absolute trust of global customers.

Export controls have cut off open markets. Sanctions lists have blocked the acquisition of advanced equipment. The rifts of great power competition have dissolved transnational trust.

As Chinese enterprises fight the uphill battle for advanced process nodes, facing initial yield rates of only 30% to 40% for SMIC's 5nm process (TrendForce 2025), they must overcome engineering challenges and rebuild an entire supply chain system that has been cut off externally, all within an extremely closed environment and at a high cost of trial and error.

There is no precedent in history. China faces technological challenges while simultaneously needing to explore an unprecedented path of breakthrough in an era where efficiency and security are in total opposition. This is an honest assessment of the difficulty of the problem, unrelated to optimism or pessimism.

The true lesson of the chip wars transcends the scope of national victory or defeat, pointing directly to the fact that the contradiction between efficiency and security cannot be resolved through technological means. Every success of a catch-up player stems from finding a temporary equilibrium point; and every equilibrium point will eventually be shattered by the next round of geopolitical shocks. This cycle happened to Japan in 1986 and South Korea in the 1990s, and it is now playing out in Taiwan. China's catch-up is the latest round of this cycle, and the difficulty of this round is unprecedented.

One detail is worth noting. When Samsung made counter-cyclical investments during the industry low in 1985, it purchased equipment that Japanese companies had abandoned due to conservatism. South Korea's catch-up was, in a sense, an opportunity created by Japan's retreat. Today, the opponent China faces has adopted an active blockade strategy, with no sign of passive retreat. Historical catch-up players have always found breakthroughs in the cracks of their opponents. When the cracks are deliberately filled, no one knows where the breakthrough point lies.

This is perhaps the most authentic conclusion of the chip war: both sides are paying an increasingly high price for a problem that cannot be solved individually, and the core of that problem is the fundamental contradiction between efficiency and security.