← Back to All Research
Commission a Report
← Back to All Research
Commission a Report
The emergence of single-company industrial monopolies in post-war US allied nations was not, at its origin, the result of explicit American industrial policy. It was an emergent consequence of a structural asymmetry that the United States deliberately created and then systematically failed to manage: the asymmetry between the security costs borne by the patron and the industrial opportunity available to the client.
The mechanism operates in five stages. First, the United States provides an unconditional security guarantee to a newly liberated or democratised nation — Japan after 1945, South Korea after 1953, Taiwan after 1949. This guarantee is backed by military presence, treaty obligations, and the credible threat of American force projection. Second, because the security threat is effectively outsourced, the protected nation can redirect defence budget to industrial capital formation. Japan's defence spending was constitutionally capped at 1% of GDP for decades; South Korea spent less on defence than comparable non-protected nations for much of the post-war period; Taiwan similarly underspent on military capability relative to the scale of the threat it nominally faced. Third, with capital freed from defence expenditure and often supported by state-directed credit, the protected nation's industrial champions pursue radical specialisation in export-oriented manufacturing categories. Fourth, over one to two decades, the combination of state direction, accumulated learning, cost advantage, and network effects produces global monopoly in one or two critical categories. Fifth — and this is the phase that matters for the present analysis — the monopoly becomes so deeply embedded in US supply chains that the security guarantee is no longer merely geopolitical in motivation. It is now industrial. The US is trapped.
The security guarantee that was supposed to create a strategic asset has, over thirty to fifty years, become a strategic liability. The patron cannot withdraw protection without destroying its own supply chains.
This report examines three canonical cases — Japan, Taiwan, and South Korea — with a partial examination of Germany, before turning to the investment implications and the question of whether the pattern is repeating in artificial intelligence.
Japan's semiconductor ascent in the 1970s and 1980s was the first and most consequential instance of the One Company Country pattern in the technology sector. By the late 1980s, Japanese firms — NEC, Hitachi, Toshiba, Fujitsu, and Mitsubishi Electric — collectively held approximately 80% of global DRAM market share. This was not a market outcome; it was the product of deliberate state-directed industrial policy coordinated through MITI (the Ministry of International Trade and Industry), the VLSI Technology Research Association (a government-funded consortium running from 1976 to 1980), and Japan Development Bank financing directed at semiconductor capital equipment investment.
The VLSI programme is worth dwelling on because it provides the clearest model of how protected-nation industrial policy works in practice. Between 1976 and 1979, MITI contributed approximately $130 million to a joint research project involving NEC, Hitachi, Fujitsu, Mitsubishi, and Toshiba. The programme produced over 1,000 patents in lithography, etching, and chip design, and the knowledge was shared across all participant firms. This was a state-subsidised technology commons that allowed Japanese firms to close a five-to-seven-year technological gap with Intel and Texas Instruments within a single decade.
| Company | Peak DRAM Market Share (late 1980s) | Key Products |
|---|---|---|
| NEC | ~17% | 64K, 256K, 1M DRAM |
| Hitachi | ~15% | 256K, 1M DRAM; SRAM |
| Toshiba | ~15% | DRAM, NAND Flash (inventor) |
| Fujitsu | ~14% | DRAM; logic ICs |
| Mitsubishi Electric | ~10% | DRAM; microcontrollers |
| Japan total | ~80% | — |
The US-Japan Semiconductor Trade Agreement of 1986 is frequently mischaracterised as a straightforward market-access deal. In reality, it was a deliberately engineered constraint on Japanese firms combined with a floor price mechanism that served US interests at the cost of Japanese competitiveness.
The agreement had three operative components. First, Japan agreed to open its domestic semiconductor market to foreign (primarily American) producers, committing to a target of 20% foreign market share within five years. Second, Japan agreed to monitor production costs of its semiconductor firms and to ensure that exports were not priced below cost — a provision the US framed as anti-dumping protection but which functioned in practice as a floor-price mechanism that raised prices for Japanese exports and reduced the cost advantage that had driven market share gains. Third, a side letter committed Japan to actively facilitating American market penetration.
The floor-price mechanism was particularly consequential. By preventing Japanese firms from competing on price, it removed the primary weapon that had enabled their market share gains. It also created a window for Korean entry: Samsung, which was ramping DRAM capacity in the mid-1980s, was not bound by the 1986 agreement and could price aggressively precisely where Japanese firms could not respond.
The 1986 agreement was renewed and strengthened in 1991. By the mid-1990s, Korean market share had risen from near zero to over 40%, primarily at the expense of Japanese firms. Japan's DRAM market share declined from approximately 80% in 1988 to under 10% by 2000.
The conventional narrative ends here: Japan lost semiconductors to Korea. The conventional narrative is wrong. Japan did not lose its monopoly positions. It changed categories.
As Japanese firms exited DRAM manufacturing under margin pressure, they concentrated investment in the layers of the semiconductor supply chain that are less visible, harder to replicate, and subject to even more durable competitive barriers than fabrication itself: materials, chemicals, and equipment. The result is a set of monopoly positions in the semiconductor supply chain that are, in several respects, more strategically significant than DRAM ever was.
| Segment | Japanese Companies | Est. Global Market Share | Notes |
|---|---|---|---|
| Silicon Wafers (300mm) | Shin-Etsu Chemical, Sumco | ~60% | Shin-Etsu alone ~30%; Sumco ~28% |
| Photoresists (ArF immersion) | JSR, Tokyo Ohka Kogyo (TOK), Shin-Etsu | ~85–90% | JSR was acquired by Japan's INCJ (state fund) in 2023 |
| Photoresists (EUV) | JSR, TOK, Shin-Etsu, Fujifilm | ~90%+ | No viable non-Japanese EUV resist at commercial scale |
| Fluorinated Polyimide | Asahi Kasei, Ube Industries | ~90% | Flexible display substrate material |
| Hydrogen Fluoride (semiconductor grade) | Stella Chemifa, Morita Chemical | ~70% | Used for etching; ultra-high purity critical |
| Sputtering Targets | JX Nippon Mining & Metals, Mitsui Mining | ~50–60% | Cobalt, tungsten, and copper deposition materials |
| CMP Slurries | Fujimi, Resonac (formerly Showa Denko) | ~40% | Chemical-mechanical planarisation |
| Coater/Developer Equipment | Tokyo Electron (TEL) | ~85–90% | TEL has near-monopoly in track equipment |
| Etch Equipment | Tokyo Electron, Hitachi High-Tech | ~25–30% | Shared with Lam Research (US) |
| Thermal Processing Equipment | Tokyo Electron | ~35% | — |
In July 2019, Japan imposed export controls on three categories of semiconductor materials bound for South Korea: fluorinated polyimide, photoresists, and hydrogen fluoride. The official justification was national security concerns about the potential diversion of materials to sanctioned states. The actual context was a diplomatic dispute over Korean court rulings relating to wartime labour compensation.
The reaction from the global semiconductor industry was instructive. Samsung, SK Hynix, and LG Display immediately began emergency stockpiling. The Korean government described the restrictions as an existential threat to its semiconductor sector. TSMC, which relies on the same Japanese materials, was indirectly exposed. The US government, despite not being a party to the bilateral dispute, watched with evident concern as a materials-layer restriction threatened to cascade through the entire global semiconductor supply chain.
The dispute was partially resolved through diplomatic negotiations and through Korean efforts to diversify domestic materials sourcing — a process that demonstrated how difficult it is to replicate Japan's materials positions even with motivated state support. Five years later, Japan's dominant position in photoresists, wafers, and etchants remains largely intact. The 2019 episode was the first public demonstration that the materials layer of the semiconductor supply chain is as fragile and concentrated as the fabrication layer — and that a single nation retains the ability to switch off global chip production with targeted export controls.
Taiwan Semiconductor Manufacturing Company was not a market creation. It was an act of deliberate industrial policy by the Taiwanese state, with explicit intellectual and financial support from the United States.
Morris Chang, who would found TSMC in 1987, spent his formative career at Texas Instruments, where he rose to Group Vice President before leaving in 1983. He was recruited by Taiwan's Industrial Technology Research Institute (ITRI), a government-funded R&D institution that had spent the late 1970s licensing semiconductor technology from RCA and building Taiwan's first integrated circuit manufacturing capability. ITRI provided the initial capital, the manufacturing facility at Hsinchu Science Park, the government imprimatur, and the political protection that allowed TSMC to operate without the competitive pressures that would have faced a purely private venture.
The founding insight was architecturally revolutionary. Chang recognised that semiconductor design and semiconductor manufacturing were separable activities, and that a dedicated manufacturer — a "pure-play foundry" — serving independent fabless chip designers could achieve utilisation rates and learning-curve efficiencies that integrated device manufacturers (IDMs) could not match. This observation, combined with the labour cost advantage Taiwan offered in the late 1980s and the government's willingness to subsidise capital equipment investment, was the foundation of TSMC's competitive model.
US policy supported the TSMC model in ways that were less visible but no less consequential. American chip designers — first Altera, then Qualcomm, Broadcom, AMD, Nvidia, and Apple — adopted the fabless model precisely because TSMC existed to manufacture their chips. US policy encouraged this specialisation. The US maintained Taiwan's security through the Taiwan Relations Act of 1979 and regular arms sales, allowing Taiwan to underinvest in defence and redirect that capital to semiconductor infrastructure. And US export controls on technology to China, which accelerated through the 2010s and reached their current intensity after 2022, reinforced TSMC's monopoly by preventing China from developing a credible alternative.
| Node / Generation | TSMC Market Share | Viable Competitors | Key Customers |
|---|---|---|---|
| 3nm (N3/N3E) | ~100% | None at volume | Apple A17, Apple M3, Nvidia GB200 (partial) |
| 4nm (N4/N4P) | ~95%+ | Samsung (limited yield) | Apple A16, Qualcomm Snapdragon 8 Gen 2, Nvidia H100/H200 |
| 5nm (N5/N5P) | ~90% | Samsung (marginal share) | Apple A15, AMD Zen 4, Nvidia A100 |
| 7nm (N7/N7P) | ~80% | Samsung, SMIC (limited) | AMD Ryzen 5000, Apple A12, Nvidia V100 era |
| Sub-7nm combined | ~90% | — | Effectively all advanced AI and mobile SoCs |
The market share figures above understate the effective concentration because Samsung's yields at leading-edge nodes have been significantly below TSMC's, reducing its practical capacity. Qualcomm famously migrated its Snapdragon flagship back to TSMC after yield problems with Samsung's 4nm process. At present, TSMC is the only foundry capable of manufacturing at volume with competitive yields below 7nm.
The dependency is total across the categories that matter most for both AI and defence:
The Taiwan Relations Act of 1979 obligates the United States to provide Taiwan with the defensive arms necessary to maintain its security, and to regard any attempt by the People's Republic of China to determine the future of Taiwan by non-peaceful means as a matter of "grave concern." The Act was originally driven by geopolitical logic: Taiwan represented a democratic counterpoint to the PRC, held strategic position in the first island chain, and was the natural successor state of the Republic of China that the US had backed in the civil war.
By 2010, and certainly by 2020, a second and arguably more powerful driver had been added to the logic of the security guarantee: industrial dependency. The US cannot allow Taiwan to fall to the PRC not merely because of the 1979 commitments or the first island chain, but because TSMC is irreplaceable on any timeline that matters for American strategic and economic interests. A TSMC disruption — from invasion, from conflict, from deliberate Chinese sabotage of the Hsinchu fabs — would eliminate advanced chip production for approximately two to three years while alternative capacity was constructed elsewhere, at a cost that would be measured in trillions of dollars of lost economic output and in the collapse of American AI and defence electronics programmes.
This creates what is structurally a hostage situation, though the political vocabulary prefers not to describe it that way. TSMC's concentration is not merely a supply chain risk; it is a constraint on American foreign policy. The US cannot credibly threaten to disengage from Taiwan because doing so would destroy the supply chains that power the American economy. The security guarantee has become self-reinforcing in a way that was not true in 1979 and may not even have been fully appreciated when TSMC was at 50% market share in the 1990s.
In October 2023, Reuters reported that Nvidia A100 chips — manufactured by TSMC and subject to US export controls preventing their sale to China — had been found inside Huawei's Ascend 910B AI accelerator. The routing mechanism used Sophgo, a Chinese chip design firm incorporated in the United States, which had legitimately ordered chips from TSMC for ostensibly non-controlled applications. The chips were then diverted to Huawei.
The incident illustrated two structural problems simultaneously. First, the concentration of advanced logic manufacturing in Taiwan creates a single chokepoint through which all advanced chips must pass — and therefore a single point of failure for export control enforcement. Second, that chokepoint operates in a regime where beneficial ownership chains can be deliberately obscured to circumvent controls. TSMC subsequently tightened its customer verification procedures and stopped shipments to Sophgo. But the episode demonstrated that the very concentration that makes TSMC strategically critical also makes it the natural target for diversion schemes.
The United States, Japan, and the European Union have each committed to reducing the TSMC concentration through the CHIPS and Science Act (US, 2022), the Chip Act (EU, 2023), and Japan's extensive subsidies for TSMC fabs at Kumamoto and a planned second site. The headline numbers are large: TSMC's Arizona facility (fab 21) targets 4nm production in 2025, with a second fab targeting 2nm by 2028; the TSMC Kumamoto fab reached volume production in early 2024. But the timeline arithmetic is brutal.
The structural conclusion is that the advanced-node TSMC concentration in Taiwan is not meaningfully addressable before 2029–2031 under any realistic programme. Before then, every disruption scenario — military action, trade embargo, natural disaster, targeted infrastructure attack — remains a catastrophic risk to the global semiconductor supply chain with no viable alternative.
Samsung's entry into DRAM in the early 1980s was characterised by a level of state-directed risk-taking that is unusual even by the standards of East Asian industrial policy. Lee Byung-chul, Samsung's founder, authorised entry into DRAM in 1983 when Samsung had no relevant technology and the global market was dominated by well-capitalised Japanese firms. The Korean government supported the decision through directed financing from Korea Development Bank and through trade policies that protected domestic semiconductor demand.
The 1986 US-Japan Semiconductor Trade Agreement, as previously noted, was the inflection point. By constraining Japanese pricing flexibility while leaving Korean firms unrestricted, the agreement inadvertently handed Samsung the cost position it needed to compete. Samsung achieved 64K DRAM parity with Japanese producers by 1984, 256K by 1986, and 1M by 1988. By the early 1990s Samsung was competitive on cost with all Japanese producers. By the mid-1990s, Samsung and Hyundai Semiconductor (later renamed Hynix after the 2001 bailout) had overtaken most Japanese firms on market share.
The Korean ascent was also facilitated by a second phase of US protection. The US-Korea Mutual Defense Treaty of 1953 provided the same security umbrella logic as the Japan arrangement: Korea's defence was effectively subsidised by US military presence, allowing Korean conglomerates (chaebols) to direct retained earnings to semiconductor capital expenditure rather than defence procurement. Samsung's semiconductor capex trajectory through the late 1980s and 1990s was only possible in a security environment where Korea did not face the full cost of its own defence.
| Segment | Samsung Share | SK Hynix Share | Korea Combined | Notes |
|---|---|---|---|---|
| DRAM (volume) | ~40–42% | ~30–32% | ~72% | Micron holds ~20%; all others negligible |
| NAND Flash | ~30–33% | ~20% (KIOXIA combined) | ~50%+ | Samsung, SK Hynix/KIOXIA JV, Kioxia, WD, Micron |
| HBM2E | ~30% | ~60% | ~90% | Micron remainder; no other viable supplier |
| HBM3E | ~20–25% | ~50–55% | ~75% | SK Hynix shipped first; Samsung and Micron ramping |
| Advanced Logic Foundry | ~5–8% (3nm GAA) | N/A | — | Samsung foundry limited by yield issues |
High Bandwidth Memory (HBM) is the memory architecture that makes modern AI accelerators possible. Nvidia's H100 uses six HBM2E stacks; the H200 and B200 use HBM3 and HBM3E respectively. HBM is manufactured through a complex 3D stacking process in which multiple DRAM dies are vertically connected using through-silicon vias (TSVs) and bonded to a logic die using advanced packaging. The process requires equipment and process chemistry that only three companies in the world can currently execute at commercial scale: SK Hynix, Samsung, and Micron.
SK Hynix's position in HBM3E is, at present, approximately that of TSMC in advanced logic: a single company holds half or more of the global supply of a component that is critical to every advanced AI system. SK Hynix shipped the first commercial HBM3E to Nvidia in early 2024 and has maintained a six-to-nine-month lead over Samsung in yield and volume for each successive generation. The AI industry's dependency on SK Hynix HBM is a structural risk that receives a fraction of the attention paid to TSMC.
The Korean concentration is somewhat less geopolitically acute than the Taiwan situation for structural reasons. South Korea has a credible sovereign military (the largest non-nuclear military in the region after the US presence), a much larger and more diversified GDP base than Taiwan, and a Chinese threat that, while serious, is less direct than the Taiwan scenario. A conflict that disrupted Korean semiconductor production would be catastrophic; a conflict that permanently transferred Korean production to Chinese control is materially less likely than the Taiwan equivalent.
Samsung's ambition to challenge TSMC in advanced logic foundry is the most visible failure of the One Company Country model when the monopoly position is not clean. Samsung's 3nm Gate-All-Around (GAA) process, which it began shipping in 2022, has been characterised by persistent yield and performance issues that have prevented it from winning the flagship designs needed to achieve volume ramp. The Qualcomm migration back to TSMC and the failure to win meaningful Nvidia AI chip share are the most visible symptoms.
The Samsung foundry situation is instructive because it demonstrates that the protected-champion model does not guarantee indefinite monopoly extension. TSMC's monopoly in advanced logic is durable because it results from a specific combination of Morris Chang's original architectural insight, thirty years of accumulated process knowledge, a customer base that has co-developed processes with TSMC, and a yield advantage that has compounded over each successive generation. Samsung is a world-class manufacturer, but it is trying to replicate TSMC's moat from behind, without the customer lock-in, in a domain where process knowledge accumulation is the primary competitive variable. The results have been consistent with that structural position.
Germany's post-WWII reconstruction under the American security umbrella presents a partial but instructive parallel to the East Asian cases. The Federal Republic of Germany from 1949 onward was protected by NATO and by the direct US military presence under the NATO Status of Forces Agreement. German defence spending was constrained, both by the occupying powers initially and by political consensus subsequently, allowing German industrial capital to be directed toward export manufacturing rather than military procurement. The Wirtschaftswunder — the post-war economic miracle — was partly a function of this security subsidy.
The German champions that emerged from this process — BASF in chemicals, Volkswagen and Daimler in automotive, Siemens in industrial electronics — achieved global dominance in their categories. BASF became the world's largest chemical company by revenue; Volkswagen Group became the world's largest automotive manufacturer by volume; Siemens became a global leader in industrial automation and power generation equipment.
However, the German version of the pattern differs from the East Asian cases in one structurally important respect: it produced category dominance across a broad range of sectors rather than a single-company monopoly in a narrow critical category. Germany's industrial strength is distributed across hundreds of medium-sized specialist companies (the Mittelstand) and several large conglomerates, not concentrated in a single firm that controls 90% of a globally critical chokepoint.
This distributional difference matters for the dependency logic. The US did not become hostage to any single German company the way it became hostage to TSMC. German industrial exposure, while substantial — BASF's chemical supply lines, Volkswagen's global supply chain footprint — never created a scenario in which the collapse of one firm would disable a critical sector of the American economy.
Germany's semiconductor sector follows the same pattern as its broader industry: significant competence in automotive-grade chips and industrial semiconductors, but no leading-edge logic capability. Infineon Technologies (power semiconductors, automotive chips) and Bosch Semiconductor (MEMS, automotive sensors) are world-class in their segments, but both operate at process nodes in the 28nm-to-mature range that are nowhere near the leading edge where the critical dependencies are concentrated.
The revealing indicator of Germany's dependency position is the TSMC Dresden facility. The EU Chips Act's centrepiece German investment is a TSMC fab, not an indigenous alternative to TSMC. Germany's response to semiconductor concentration risk is to bring more TSMC capacity to Europe — which reduces geographic concentration somewhat but deepens the dependency on TSMC technology and process know-how. The Intel Magdeburg facility, originally planned to be Europe's answer to TSMC, has faced repeated delays and budget revisions that have pushed volume production timelines into the 2028–2030 range at best.
The One Company Country pattern can now be described with sufficient precision to map the current situation:
The question of intentionality is important for understanding whether the pattern can be managed differently going forward. The answer appears to be: the individual components were intentional; the systemic outcome was not.
US decision-makers in 1952 deliberately provided Japan's security guarantee. MITI officials in 1976 deliberately designed the VLSI programme to build Japanese semiconductor capability. The Taiwanese government deliberately recruited Morris Chang and funded ITRI's technology licensing. Each individual decision was made with explicit strategic intent. But no single decision-maker at any of these junctures appears to have worked through the full five-stage feedback loop to its conclusion.
This is consistent with the general pattern of emergent geopolitical liability: it accumulates through individually rational decisions that collectively produce systemic risk that is only visible in retrospect, or when a stress event forces attention. The 2019 Japan-Korea dispute was the first major stress event that forced attention to the materials layer. Russia's 2022 invasion of Ukraine was a comparable stress event for European energy dependency, a pattern with structural similarities to the semiconductor case.
The clearest evidence that the inversion is now structurally complete is the behaviour of US policy toward Taiwan since 2022. The CHIPS and Science Act, the export controls on advanced semiconductor technology to China, the classification of semiconductors as critical infrastructure, and the visible acceleration of TSMC's Arizona timelines all reflect a US government that has concluded that Taiwan's semiconductor dominance is too concentrated and too risky — but that also knows it cannot afford to let Taiwan fall in the interim. Every action the US is taking on semiconductor policy is a tacit admission that it is trapped by the dependency it allowed to form.
The United States is simultaneously trying to reduce its TSMC dependency (through CHIPS Act subsidies), protect TSMC's current capacity (through the Taiwan security guarantee), and prevent China from acquiring TSMC's technology (through export controls). These three objectives are, in the short run, consistent. In the medium run, they are in tension. In the long run, only one of them can succeed.
The forced diversification of advanced semiconductor manufacturing capacity is one of the most durable multi-decade capital allocation themes in global markets. The policy commitment is bipartisan in the US and cross-party in Europe; the industrial rationale is structural; and the timeline — a decade or more to meaningfully reduce TSMC concentration — ensures that the thematic exposure is not subject to a sudden resolution. The principal beneficiaries:
TSMC itself (TSM) is the paradoxical beneficiary of its own concentration risk: the subsidies flowing from the US CHIPS Act, Japanese METI, and EU Chips Act are effectively being paid to TSMC to build outside Taiwan. TSMC Arizona and TSMC Kumamoto are capital-light from TSMC's perspective precisely because host governments are funding the infrastructure. The incremental return on capital for TSMC's offshore fabs is structurally superior to its Taiwan operations because a substantial fraction of the capital cost is borne by taxpayers rather than shareholders.
ASML (ASML) remains the sole global supplier of extreme ultraviolet lithography systems, without which sub-7nm manufacturing is impossible. Every new fab worldwide requires ASML equipment. ASML is the one company in the world that benefits from every diversification attempt simultaneously, regardless of which government's fab succeeds.
Japan's Shin-Etsu Chemical, Sumco, JSR, Tokyo Ohka Kogyo, Resonac, and Stella Chemifa hold monopoly or near-monopoly positions in semiconductor materials that are prerequisite for all advanced node manufacturing. These positions are geopolitically safer than TSMC Taiwan because (a) Japan is a formal US treaty ally, (b) the materials are shipped globally from Japan rather than manufactured in a single facility, and (c) the companies are less headline-visible and therefore less subject to the nationalisation-of-attention that drives political risk. The 2023 state acquisition of JSR by INCJ, Japan's government-owned industrial competitiveness fund, signals that the Japanese government understands the strategic value of these positions and intends to maintain them. For investors, the materials layer is structurally underpriced relative to the fabrication layer.
SK Hynix's HBM3E monopoly position in AI memory is the next TSMC-style concentration that policy will eventually be forced to address. Micron Technology is the most advanced western alternative, having achieved HBM3E production in late 2024. Micron's strategic value to the US government is analogous to Intel's: the last American manufacturer in a critical semiconductor category, supported by CHIPS Act subsidies and government procurement. The investment thesis for Micron is not DRAM cycle exposure; it is HBM strategic option value at a company that the US government is structurally motivated to ensure survives.
Intel Foundry Services (IFS) is the American government's preferred answer to TSMC concentration: a domestic, US-owned advanced foundry that can manufacture leading-edge logic chips without geographic concentration in Taiwan. Intel received approximately $8.5 billion in CHIPS Act grants and up to $11 billion in loans — the largest single CHIPS Act commitment — on the explicit expectation that IFS would provide a credible alternative to TSMC for advanced US defence and AI applications.
The investment thesis for IFS is difficult because it requires separating the strategic value to the US government (very high — a domestic foundry with leading-edge capability is worth almost any price to policymakers) from the commercial value as a business (uncertain — IFS has lost customer confidence from years of execution failure, has no established track record as a contract manufacturer, and faces TSMC on customers' preferred supplier lists after decades of relationship building). The most likely scenario is that IFS achieves a defensible position in US government workloads — where buyers are motivated by supply chain security rather than cost — while remaining a secondary option for commercial customers. That is strategically valuable but does not justify a TSMC-equivalent commercial multiple.
Samsung's 3nm GAA process represents the limits of what the One Company Country model can achieve when the monopoly is not clean. Samsung is a world-class manufacturer with deep process expertise, but it is pursuing TSMC's leading-edge customers from behind and against a company that has thirty additional years of customer relationship investment. The yield gap — estimated at 10–20 percentage points below TSMC at equivalent nodes — is the primary barrier, and closing it requires the kind of learning-by-doing that only comes from winning the customers that the yield gap is currently preventing Samsung from winning. This is a structural feedback loop that is difficult to break without a major TSMC disruption creating a forced diversion of orders.
GlobalFoundries (GFS) is not a leading-edge foundry and has explicitly exited the race to sub-7nm. Its strategic value is in a different and arguably more durable position: the 12nm-to-28nm range that serves automotive, aerospace, defence, and industrial applications. These are the nodes where supply chain security matters to the US government and where domestic US manufacturing capacity is most obviously absent. GFS received CHIPS Act support and has a long-term supply agreement with the US DoD for secure fabrication of defence-related chips. The investment thesis is not growth; it is strategic premium in a niche where the government is the effective demand anchor.
The question of whether the United States is currently creating a new generation of One Company Country dynamics — this time in artificial intelligence rather than semiconductor manufacturing — is uncomfortable but important to raise. Several indicators suggest the pattern may be emerging in a modified form.
OpenAI and Anthropic, the two leading American frontier AI labs, have received forms of implicit government support that bear structural resemblance to the early-stage protected champion incubation of Japan and Taiwan's semiconductor firms. This support has operated through several channels:
If the semiconductor analogy is instructive, the risks of the AI champion model are predictable from the pattern. The immediate effect of protection and implicit support is positive: American AI labs maintain their lead, develop dominant products, and accumulate market share globally. The medium-term risk is that the protected champion, shielded from the full discipline of competition, becomes complacent or moves in the direction that maximises its domestic market rather than global competitiveness. The long-term risk — the inversion risk that this report has mapped in semiconductor manufacturing — is that the champion becomes so critical to American economic and military capability that the government cannot afford to let it fail, even if it becomes uncompetitive, even if its interests diverge from the public interest, and even if maintaining it requires the kind of continued subsidy and protection that distorts the broader market.
The semiconductor case demonstrates that this inversion takes two to four decades to complete. The AI case may move faster, given the pace of the industry. But the structural logic is the same: protection enables monopoly; monopoly creates dependency; dependency inverts the original logic of protection.
The three historical cases examined in this report suggest that the One Company Country pattern, once established, is extremely difficult to unwind. The US-Japan Semiconductor Agreement of 1986 succeeded in constraining Japanese DRAM dominance — but it did so by creating the conditions for Korean dominance, not by genuinely diversifying the market. The result was a transfer of monopoly from Japan to Korea, not an elimination of concentration risk. The current CHIPS Act and EU Chips Act efforts are, structurally, attempting to do what the 1986 agreement failed to do: create genuine multi-country, multi-company diversification in a critical technology sector.
The obstacles are formidable. Advanced semiconductor manufacturing requires not just capital but accumulated process knowledge, workforce expertise, and supply chain density that takes decades to build. Taiwan's advantage in advanced logic is not primarily capital — it is thirty years of learning-by-doing compressed into an extraordinarily deep institutional capability. Replicating that capability in Arizona, Ohio, Dresden, or Osaka requires not just building fabs but building ecosystems: supplier networks, engineering training pipelines, materials supply chains, and the management know-how that can only be developed through sustained high-volume production.
The most underappreciated aspect of the One Company Country pattern is the asymmetry it creates in dependency relationships. Taiwan is existentially dependent on the US security guarantee for its own survival as a political entity; but the US is also critically dependent on Taiwan's semiconductor manufacturing for its own economic and military capability. This mutual dependency creates a stability of sorts — each party has strong incentives to maintain the relationship — but it also creates a stability that is fragile in a specific way: it is stable until a third party (China) decides to change the terms.
China's strategic calculus with respect to Taiwan is complicated by exactly this structure. A military campaign against Taiwan that physically destroyed TSMC's manufacturing infrastructure would impose enormous costs on China's own technology sector and on the global economy, which would rebound against China's export-dependent economy. This creates a deterrent to physical destruction. But a campaign that captured TSMC's infrastructure intact — and therefore acquired the manufacturing know-how, customer relationships, and equipment base — would be a different calculation entirely. The most dangerous Taiwan scenario is not the one that destroys TSMC; it is the one that transfers TSMC to Chinese control.
The moat created by the One Company Country structure is not primarily capital, scale, or patents. It is tacit knowledge — the category of expertise that cannot be transferred, purchased, reverse-engineered, or written into a specification document. This distinction is not semantic. It is the single most important reason why the concentration patterns described in this report are so durable and so difficult to displace.
Semiconductor manufacturing is unusual among capital-intensive industries in the degree to which its critical knowledge resists codification. Process recipes, yield improvement loops, defect characterisation procedures, and fab-specific tuning parameters are not primarily documented — they exist in the institutional memory of engineers who have spent entire careers inside one organisation, on one equipment set, refining one process generation. The gap between a written process specification and a functioning high-yield process is measured in years of learning-by-doing that cannot be compressed by money or intent.
This is not a theoretical claim. It is evidenced by every attempted fast-follower in advanced semiconductor manufacturing. The tacit knowledge embedded in Tainan's fabs is not in TSMC's internal documentation; it is in the accumulated intuition of thousands of process engineers who know, from experience, that a particular furnace in Fab 12P runs two degrees warm and needs compensating adjustment, or that a specific photoresist lot from a particular supplier produces a characteristic defect signature that requires a non-standard cleaning step. That knowledge is not transferable by hiring a few engineers or licensing a process node. It is the product of decades of compounding experience inside a specific institutional and physical environment.
Intel's failure to match TSMC's advanced node yields over two decades of sustained effort is the clearest available proof that the tacit knowledge gap is not a financial problem. Intel had the capital — it has spent more on capex in absolute terms than TSMC across multiple periods. Intel had patents, including foundational process patents that TSMC licences. Intel had engineers, including engineers hired from TSMC and from the broader industry. Intel had two decades of trying, with the full weight of US government concern about semiconductor sovereignty behind it from 2018 onwards.
Intel still cannot match TSMC's N3 yields. As of the mid-2020s, TSMC's N3 process is manufacturing at volume for Apple, Nvidia, AMD, and Qualcomm. Intel's equivalent node (Intel 3) has faced repeated delays and yield struggles that have caused multiple major customers to defect to TSMC for manufacturing. The gap is not financial. It is experiential accumulation that cannot be bought, and the Intel case is the most expensive real-world experiment in the history of industrial policy confirming that fact.
The structural mechanism behind tacit knowledge accumulation is compartmentalisation — the concentration of thousands of specialists in one geography, one culture, one organisation, and one supply chain. This concentration creates compounding knowledge density that distributed models cannot replicate. Knowledge begets knowledge inside the boundary; outside the boundary it dissipates. A process engineer in Tainan who troubleshoots a yield problem learns something. That learning propagates laterally to colleagues in adjacent fabs, upstream to equipment engineers, downstream to the materials suppliers who co-develop with TSMC's process teams. The same engineer in Phoenix acquires the same learning but propagates it into a thinner network, at lower velocity, with fewer feedback loops. Over five years the cumulative difference is measurable. Over thirty years it is the TSMC advantage.
Japan's materials monopoly operates on the same logic. JSR's photoresist dominance is not a patent story — JSR's key photoresist patents have in many cases expired or been designed around. The dominance is forty years of process chemistry refinement, customer feedback loops operating at the level of individual fab teams, and co-development relationships with process engineers at TSMC, Samsung, and Intel that give JSR early visibility into next-generation requirements. The 2019 Korea dispute demonstrated the depth of this compartmentalisation with unusual clarity: Samsung and SK Hynix, despite being world-class manufacturers with massive R&D budgets, could not substitute Japanese fluorinated polyimide or hydrogen fluoride gas on short notice. Not because the chemistry was unknown but because the process-qualified supply — the accumulated trust, the characterised quality parameters, the certified process compatibility — existed only in the Japanese supply chain.
The US security guarantee plays a specific and underappreciated role in this dynamic. When a protected nation does not need to spend at the level of full sovereign defence, capital and engineering talent that would otherwise be absorbed by the defence sector flow into the industrial champion. Taiwan's defence spending has historically run at approximately 2% of GDP or below; the engineering talent that in an unsheltered nation would be distributed across defence contractors, military R&D programmes, and dual-use technology development has in Taiwan been concentrated in TSMC's supply chain. The compartmentalisation accelerates precisely because the protected environment is stable over decades, allowing the knowledge density to compound without the disruptions that geopolitical exposure would introduce.
This is not incidental. It is the core mechanism by which the security umbrella creates the monopoly: protection creates stability; stability enables concentration; concentration compounds tacit knowledge; tacit knowledge creates the moat that makes the monopoly self-reinforcing.
The implications for policy are uncomfortable. You cannot diversify tacit knowledge by building a new fab. TSMC Arizona will take a decade at minimum to accumulate the yield knowledge that Tainan has built over three decades, even with TSMC's own engineers on site attempting to transfer practice. The hardware can move — the EUV machines, the clean rooms, the process equipment — but the knowledge cannot move at the same pace, because the knowledge is not in the hardware. It is in the people, and in the institutional structures and feedback loops that those people operate within. An engineer who transfers from Tainan to Phoenix retains their individual knowledge but loses the network that made that knowledge productive. They are removed from the ecosystem of lateral propagation that is the actual source of the compounding advantage.
This is why every serious analysis of semiconductor diversification concludes with the same uncomfortable arithmetic: a decade of operation before yield parity, two decades before full process capability depth, at costs per wafer that will be structurally higher than Taiwan for the foreseeable future. The One Company Country pattern is hard to unwind not because of political will or capital constraints, but because the knowledge that makes it valuable is the one input that cannot be replicated on a faster schedule.
The same structural pattern that the United States created — intentionally and otherwise — in Taiwan, Korea, and Japan may now be emerging within the United States itself. The protected national champion, the compounding concentration of capability, the gradual accumulation of critical dependency: these dynamics are visible in the US AI frontier industry in 2025 in ways that closely parallel the early-stage semiconductor concentration of the 1970s and 1980s. The analogy is imperfect, but the structural logic is close enough to warrant serious examination.
The US security umbrella gave Taiwan the stability to concentrate industrial capital in TSMC by removing the primary source of external threat that would otherwise have diverted resources and attention. US regulatory policy toward frontier AI is performing a structurally analogous function for Anthropic, OpenAI, and Google DeepMind. There is no binding liability framework for AI-caused harm at the federal level. There is no mandatory safety certification equivalent to the FDA process that would slow deployment and impose compliance costs. Preferential government procurement — through defence contracts, intelligence community engagements, and GSA schedules — provides a protected revenue base that allows these organisations to compound capability investment faster than they could in a fully contested regulatory environment.
The result is a protected operating environment that accelerates concentration in exactly the way the security umbrella accelerated semiconductor concentration. The protection is not explicit — no treaty, no formal industrial policy designation — but its functional effect on the competitive dynamics of the industry is similar. Regulatory forbearance is a form of subsidy, and in a capital-intensive technology race, it functions as one.
The Biden administration's October 2022 chip export controls, extended and tightened through 2023 and 2024, and the Trump administration's continuation of the same policy framework, are explicitly designed to prevent Chinese AI developers from accessing the hardware required to compete at the frontier. Nvidia A100 and H100 export restrictions, entity list additions covering Chinese AI labs, and the licensing requirements for advanced chip exports to a broad range of countries all function to protect US frontier AI developers from their most serious potential competitor.
This is functionally equivalent to the trade protection the US gave Japanese and Korean industrial champions in their formative decades — except the protection now runs in reverse, applied against the challenger rather than in favour of the ally. In the 1970s and 1980s, US policy gave Japan and Korea protected access to the US market and limited competition from US firms in certain segments. In the 2020s, US policy is denying China the hardware inputs required to compete in AI frontier development. The mechanism is different; the structural effect — creating a protected environment in which a small number of favoured organisations can compound capability advantage — is similar.
Frontier AI training capability is concentrating in a pattern that mirrors the TSMC/Taiwan pattern at an earlier stage of development. Anthropic (Delaware public benefit corporation, with a significant London research and policy presence), OpenAI (Delaware, San Francisco), and Google DeepMind (US headquarters, London research base) account for the overwhelming majority of frontier model training runs at scale. The geographic and organisational concentration is not coincidental — it reflects the same compounding dynamics of talent density, capital access, compute access, and institutional knowledge that made Tainan the centre of global advanced semiconductor manufacturing.
The UK's position in AI requires more careful framing than a simple dependency narrative. The concentration of frontier AI capability in San Francisco is a venture capital and data centre story; it is not the intellectual origin story. The foundational knowledge that underpins every modern neural network was developed in Britain.
Alan Turing, working at Cambridge and then at Manchester, produced the theoretical architecture of computation and the conceptual framework of machine intelligence that defines the field at its roots. Geoffrey Hinton — UK-born, Cambridge undergraduate — developed the backpropagation techniques and deep learning methods that every frontier model in production today is built upon; he spent decades in academia before his work was industrialised by US capital. Demis Hassabis, who read computer science at Cambridge and completed his PhD at UCL, founded DeepMind in London in 2010, and the lab's foundational research on reinforcement learning and protein structure prediction remains among the most scientifically significant work in the field. Oxford and Cambridge continue to supply a disproportionate share of the research talent at every major frontier lab. AI is, at its foundations, applied mathematics; and the British mathematical tradition — from Turing and the wartime cryptanalysis infrastructure at Bletchley through to the current generation of researchers — is the origin of that mathematics.
The correct structural description is therefore not that the UK is a junior participant in a US-controlled compute stack. It is that the UK exported the human capital that built the field, and the US provided the capital stack to industrialise it. The concentration in San Francisco reflects where the venture capital, the hyperscaler infrastructure, and the regulatory forbearance were located — not where the underlying knowledge originated. DeepMind operating within Google's compute infrastructure is a consequence of that capital asymmetry, not evidence that the UK's contribution was peripheral.
This makes the UK's structural position in AI closer to Japan's position in the early semiconductor industry than to a simple dependency relationship. Japan's VLSI programme in the late 1970s generated the process chemistry knowledge and manufacturing techniques that underpinned the entire subsequent semiconductor supply chain; Japan originated the knowledge commons, then lost manufacturing dominance to Taiwan and Korea as capital concentration and production scale shifted. The UK originated the intellectual commons of modern AI, then lost the industrialisation race to US capital. In both cases, the knowledge origin is not in question; what transferred was the capital-intensive production layer. The UK government's AI ambitions, expressed through the AI Safety Institute and various compute investment commitments, are real policy intentions facing a structural capital and infrastructure deficit — one that is genuinely difficult to overcome at speed, but whose origins lie in capital allocation rather than intellectual capacity.
If Anthropic or OpenAI become as structurally critical to US national security as TSMC is to global semiconductor supply — and the trajectory of AI capability deployment into defence, intelligence, and critical infrastructure suggests this is not a distant scenario — the US faces its own version of the TSMC hostage problem. Policy becomes constrained by the need to protect the champion, not just support it. Regulatory action that might otherwise be warranted becomes politically and strategically difficult when the regulated entity is simultaneously a critical national security asset. The concentration that the protected environment created becomes the constraint on the protection that the concentration requires.
This inversion is not a future hypothetical. The US government's reluctance to impose binding frontier AI regulation, despite substantial internal debate about AI risk, is already partially explicable by the strategic logic of not wanting to constrain the organisations whose capability advantage is considered central to competition with China. The pattern by which strategic asset status insulates a champion from regulatory accountability is already operating.
The One Company Country pattern in AI differs from its semiconductor predecessor in ways that make the concentration both potentially less durable and potentially more dangerous. The critical difference is that the compartmentalisation mechanism — tacit knowledge in one geography — is structurally weaker for AI than for semiconductor manufacturing. A significant and growing portion of AI capability is codified in published research. Transformer architectures, scaling laws, RLHF methodology, and the key techniques behind current frontier models have been published in peer-reviewed papers accessible to any research group with the compute to implement them. The process recipe for N3 does not exist in the open literature. The process recipe for GPT-4-class capability substantially does.
This codification of AI capability creates a disruption risk that semiconductor manufacturing has never faced. DeepSeek's January 2025 demonstration that a Chinese lab could approach frontier capability at dramatically lower compute cost — using techniques substantially derived from published research — is the clearest illustration of this risk. The open-source development of LLaMA and its successors means that a future open-source model may close the enterprise capability gap more quickly than the commercial frontier labs' moats would suggest. If that happens, the US AI concentration faces a disruption that TSMC has never faced and arguably cannot face: its equivalent would be if someone successfully open-sourced the N3 process recipe in sufficient detail to enable a new entrant to replicate it. That is not possible in semiconductor manufacturing. In AI, the analogous event is a live risk.
The One Company Country pattern in AI may therefore be less durable than its hardware predecessors. But the structural dynamics of concentration — the protected environment, the compounding advantage, the critical dependency, the eventual policy constraint — are following the same arc. Understanding the full trajectory of the semiconductor case is the best available framework for estimating how far the AI case will develop and where it will ultimately invert.
The EU AI Act is the heaviest compliance framework applied to any technology sector to date — mandatory conformity assessments, prohibited use categories, and liability structures that add meaningful operational overhead before a product reaches market. This is the regulatory environment that prevents a European equivalent of TSMC or Samsung from emerging in AI. EU distributed regulatory sovereignty is structurally incompatible with the concentrated industrial bet that produces a monopoly champion: no single member state can act decisively, consensus is driven by the most cautious member, and the result is a framework optimised for risk containment rather than competitive acceleration. The jurisdictions that have produced monopoly champions share one structural feature — a single authority capable of making and sustaining an industrial bet. The EU is, by design, not that.
The UK post-Brexit sits outside this gravity well. The AI Safety Institute frames regulation as pro-innovation with guardrails rather than precautionary by default, and there is no binding AI Act equivalent on the UK statute book. The more accurate framing is not that the UK is a European AI player with a lighter touch; it is that the UK occupies a distinct regulatory and knowledge jurisdiction — sharing language, security architecture, and research culture with the US, while carrying none of the EU compliance overhead. That combination is strategically meaningful. The UK's structural position in AI more closely resembles a satellite of the US-led protected environment than a competitor to it.
France is the exception that proves the rule. Mistral — Paris-headquartered, founded by researchers from DeepMind and Meta — is a genuine frontier open-weights lab operating inside the EU. It exists because France has a long and deeply embedded tradition of dirigisme: state-directed industrial policy so thoroughly internalised that French champions emerge almost in spite of EU-level consensus. Nuclear independence, aerospace (Airbus led from Toulouse), agricultural self-sufficiency, cultural protectionism through the cultural exception — France effectively runs a national industrial policy inside a supranational body that officially prohibits one. Mistral is the AI expression of that tradition. It is the exception to the structural argument, not evidence that the EU framework is compatible with producing monopoly champions at scale.
The One Company Country pattern is not a historical curiosity. It is the dominant structural fact of the global semiconductor supply chain, and its logic is working itself forward into the AI industry in real time. Understanding the pattern — its origins in Cold War security guarantees, its mechanism through state-directed industrial policy, its fruition in global monopoly, and its inversion into structural dependency — is prerequisite to understanding both the geopolitical risks that shape global markets and the investment opportunities that those risks create.
Three conclusions deserve emphasis as the terminal point of this analysis:
First, the materials layer is the overlooked monopoly. Japan's Shin-Etsu, Sumco, JSR, Tokyo Ohka, and Resonac hold more durable positions than TSMC in a logical sense: they are prerequisites for TSMC's manufacturing, they are geographically safer, they are less visible to the political attention that creates risk, and they have already demonstrated their leverage in the 2019 Korea export controls. OVERWEIGHT the Japanese materials layer is, from a risk-adjusted return perspective, potentially the most attractive single expression of the structural thesis in this report.
Second, the HBM concentration is the next TSMC. SK Hynix's position in HBM3E for AI accelerators is already a single-point-of-failure for the global AI industry, and it receives approximately one-tenth of the political and media attention of TSMC. The forced diversification of HBM supply will be the next major industrial policy theme in semiconductors. Micron's strategic value to US policymakers is correspondingly higher than its commodity DRAM cycle exposure would suggest.
Third, the inversion logic applies to AI and the next generation of champion industries. The pattern by which a US security guarantee enables a protected nation to build a global monopoly that subsequently constrains US foreign policy is not a completed historical phenomenon. It is a live process. The AI case is early-stage, but the structural conditions for repeating the pattern — implicit protection, first-mover advantage, compounding network effects, and the gradual accumulation of critical dependency — are already present. Investors who understand the full arc of the semiconductor case are better positioned to identify both the opportunity and the eventual liability in AI champion dynamics.
The fundamental strategic error in the One Company Country model is not the initial security guarantee — that was rational. It is not the industrial policy that enabled the champion — that was also rational. The error is the failure to manage the dependency that inevitably results, and to recognise that a strategic asset with no substitute is, by definition, a strategic liability. The United States has built the most productive and efficient semiconductor ecosystem in history by concentrating it in Taiwan and Japan. It now must live with what it built.
| Year | Event | Significance |
|---|---|---|
| 1952 | US-Japan Security Treaty (revised) | Formalises US military protection; Japan caps defence at ~1% GDP |
| 1953 | Korean Armistice; US-Korea Mutual Defense Treaty | South Korea under US security umbrella; Korean defence partially subsidised |
| 1976–1980 | Japan VLSI Research Programme (MITI) | State-funded technology commons; NEC/Hitachi/Toshiba/Fujitsu/Mitsubishi close gap with US |
| 1979 | Taiwan Relations Act (US) | Replaces US-ROC treaty; establishes de facto security guarantee for Taiwan |
| 1983 | Samsung enters DRAM manufacturing | First chaebols-backed semiconductor push; Korea Development Bank financing |
| 1986 | US-Japan Semiconductor Trade Agreement | Floor-price mechanism constrains Japanese competitiveness; opens window for Korea |
| 1987 | TSMC founded (Morris Chang, ITRI) | Pure-play foundry model; US government implicitly supports Taiwanese semiconductor push |
| 1991 | US-Japan Semiconductor Agreement renewed | Further constrains Japanese pricing; Samsung reaches cost parity |
| Mid-1990s | Korea overtakes Japan in DRAM share | Monopoly transfers rather than diversifies; Samsung/Hynix reach ~60% combined |
| 2001 | Hyundai Semiconductor bailout; renamed Hynix | Korean government sustains the champion; avoids market exit that would have reduced concentration |
| 2010s | TSMC secures Apple, Nvidia, Qualcomm advanced node volumes | US fabless model cements TSMC monopoly; American supply chain dependency becomes structural |
| 2019 | Japan-Korea export controls dispute | First proof of materials layer strategic leverage; photoresists, fluorinated polyimide, HF restricted |
| 2020 | TSMC announces Arizona fab | First acknowledgement at US government level that Taiwan concentration is a liability |
| 2022 | US CHIPS and Science Act ($52B) | Largest peacetime US industrial policy commitment; forced diversification attempt begins |
| 2023 | Japan INCJ acquires JSR; EU Chips Act passed | Japanese government nationalises photoresist monopoly; European diversification attempt |
| 2023 | Sophgo/Huawei A100 chip diversion incident | TSMC concentration creates single enforcement chokepoint; export control evasion via beneficial ownership |
| 2024 | SK Hynix ships first HBM3E to Nvidia | Korean HBM monopoly for AI memory established; next concentration risk crystallises |
| 2025 | TSMC Arizona N4 fab reaches production | First US-manufactured advanced node chips; insufficient volume to change dependency dynamics |
This report is produced by PRZC Research for informational purposes only. It does not constitute financial advice, investment advice, or a solicitation to buy or sell any security. Market share estimates are derived from publicly available industry data, company filings, and PRZC Research analysis; they are subject to revision and should not be treated as precise current figures. The One Company Country framework is an analytical construct developed by PRZC Research and represents the author's structural interpretation of historical events. Past industrial policy outcomes are not indicative of future geopolitical or investment results.
PRZC Research is incorporated in the Republic of Seychelles and is not subject to FCA, SEC, or equivalent regulatory oversight. Reports are addressed to institutional and sophisticated investors capable of independently assessing the risks described.