The Institutions and Why They Are Wrong

Markets

Markets are extraordinarily good at pricing risks that have happened before. They have decades of data on interest rate cycles, credit defaults, earnings misses, and geopolitical disruptions of the kind that resolve within quarters. The institutional infrastructure of modern finance, including the options markets, the credit default swaps, and the risk models running in every major fund, is built almost entirely on the assumption that the distribution of future outcomes resembles, in its general shape, the distribution of past outcomes. For most risks, most of the time, that assumption is defensible.

For hardware sovereignty risk, it is not defensible. And the consequences of that failure are already embedded in the valuations of nearly every significant company in the technology sector.

Begin with the most obvious case. NVIDIA, as of 2024, traded at valuations implying a degree of earnings durability and compounding that is consistent with a software business, where margins are structurally high, switching costs are deep, and the physical constraints on production are essentially zero. The market was not wrong to identify NVIDIA's pricing power or the depth of its competitive moat in accelerated computing. It was wrong to price that moat as though it exists independently of TSMC's continued operation under conditions of open access. Every H100, every B100, every Blackwell chip that NVIDIA designs is physically manufactured in Taiwan. NVIDIA owns no fabrication capacity. It has no alternative supplier for its leading-edge products. Its entire business model is, at its foundation, a bet that TSMC will be accessible indefinitely.

This is not a tail risk in the way that term is used in finance. A tail risk is an event with low probability and high impact that exists at the edge of a known distribution. The TSMC concentration is not at the edge of a known distribution. It is a structural dependency that sits at the center of every revenue projection NVIDIA has ever published, priced as though it does not exist. The market has assigned NVIDIA a valuation commensurate with a durable compounding business and has simultaneously declined to discount that valuation for the single most material risk to its revenue.

The valuation error is not specific to NVIDIA. It runs across the fabless semiconductor industry, the hyperscaler infrastructure buildout, and every technology company whose products depend on leading-edge chips. Apple's market capitalization implicitly assumes continued access to TSMC's most advanced nodes for iPhone and Mac processors. Amazon, Google, and Microsoft are collectively spending hundreds of billions of dollars building data center infrastructure that requires a continuous supply of chips that can only be made in Taiwan. None of their public filings treat the geographic concentration of that supply as a material risk in the way that customer concentration or regulatory risk are treated as material.

The options market makes the miscalibration legible in a different way. If you believed, genuinely and with conviction, that there was a ten percent probability of a Taiwan disruption in the next five years, you would expect to find liquid, deep options markets pricing that risk into the volatility surfaces of every company exposed to it. What you find instead is a volatility surface that treats NVIDIA like a high-beta growth stock and prices its long-dated options accordingly, with no visible premium attributable to supply chain sovereignty risk. The market is not pricing a ten percent probability of Taiwan disruption. It is pricing something closer to zero.

This is not because sophisticated investors have analyzed the risk and concluded it is negligible. It is because the risk has a property that makes it systematically difficult for markets to price: it has not happened yet in a form that has produced realized losses for institutional investors. Modern risk management is almost entirely backward-looking. Value at Risk models, factor models, stress testing frameworks: all of them are calibrated on historical data. The historical data does not contain a Taiwan disruption because a Taiwan disruption has not occurred. The model therefore assigns it negligible weight, not because the model has evaluated the probability as negligible but because the model has no mechanism for incorporating a risk that has no historical analog.

This is a known limitation of quantitative risk frameworks. What is less appreciated is how completely it dominates institutional behavior. A risk that cannot be measured by the tools a fund uses to manage its portfolio is, in operational practice, a risk that does not exist for that fund. The fund manager who raises Taiwan concentration as a reason to reduce NVIDIA exposure faces an immediate problem: the quantitative frameworks their risk committee uses to evaluate that decision will show no historical basis for the concern, the position will look like alpha foregone rather than risk managed, and the manager will underperform every benchmark for as long as the disruption fails to materialize. The incentive structure of institutional asset management actively discourages the pricing of novel structural risks.

The result is a global financial system that has allocated trillions of dollars of capital to businesses whose value depends entirely on a physical supply chain it has declined to model as a risk. This is not fraud. It is the predictable output of financial infrastructure built for a world that no longer exists, operated by institutions whose incentive structures punish the early pricing of risks that have not yet materialized. The market will eventually price this correctly. It will do so after the disruption rather than before it, which is the only way markets have ever incorporated structural risks of this kind.

Governments

In August 2022, President Biden signed the CHIPS and Science Act into law, committing 52 billion dollars to domestic semiconductor manufacturing and research. It was covered as a historic industrial policy intervention, a reversal of decades of offshoring complacency, a signal that the United States had finally understood the strategic importance of semiconductor sovereignty. The applause was genuine. The diagnosis was correct. The response was approximately twenty years too late and an order of magnitude too small.

The geographic concentration of semiconductor manufacturing in Taiwan was not a surprise development. It was the predictable outcome of deliberate policy choices made across multiple administrations beginning in the 1980s. By the early 2000s, TSMC's dominance at the leading edge was already a documented feature of the industry landscape. By 2015, the Defense Science Board had published reports flagging the national security implications of leading-edge chip production being concentrated outside American borders. The warning was registered. The policy response was negligible.

The CHIPS Act was not a proactive strategy. It was a reaction to a crisis that had already arrived, triggered primarily by the automotive chip shortage of 2021 and the visible acceleration of Chinese semiconductor investment following Huawei's blacklisting in 2019. The legislation addressed symptoms that had been present for decades and did so with a funding commitment that, while significant in absolute terms, is dwarfed by the actual cost of rebuilding competitive domestic manufacturing capacity. A single leading-edge fab costs between 20 and 30 billion dollars to construct. The CHIPS Act's manufacturing incentives are sufficient to partially subsidize two or three such fabs, which is not sufficient to create genuine supply chain resilience at a national level.

The deeper problem is structural rather than fiscal. Semiconductor fabs require ten to twelve years from investment decision to full-volume production at a new process node. The political cycle in a democracy is four years. The incentive structure of elected government is therefore systematically misaligned with the investment timeline required to build genuine hardware sovereignty. A politician who commits 30 billion dollars to a fab that will not produce competitive chips until after two more presidential elections is making a bet whose payoff, if it comes, will be attributed to their successor and whose cost will be borne immediately.

The export control regime presents a different version of the same structural failure. The Commerce Department's Bureau of Industry and Security began placing Chinese technology companies on the Entity List in 2019, starting with Huawei. The October 2022 export controls represented the most sweeping semiconductor export restrictions since the Cold War. The October 2023 expansion tightened those controls further in response to Chinese firms finding ways around the initial restrictions. Each successive tightening has been a reaction to demonstrated Chinese circumvention of the previous rule.

This is not a criticism of the policymakers involved. The problem is not the quality of the individual decisions. It is that the framework within which those decisions are made is reactive by design. Export controls are administered by a bureaucracy that responds to specific identified threats, documents them through a legal process, and issues rules after the threat has already shaped behavior. The Chinese AI investment programs that the October 2022 controls were designed to constrain had been operating at scale for years before the controls were imposed.

The European Union's Chips Act, announced in 2022 with a headline figure of 43 billion euros, reproduces the American mistakes with European characteristics. The funding is distributed across member states with competing industrial interests, subject to state aid rules that constrain how aggressively any single country can support its national champions, and oriented primarily toward securing supply for European automotive and industrial customers rather than building sovereign capacity at the leading edge.

Japan's response is the most ambitious and the most likely to fail on its stated timeline. RAPIDUS has announced plans to produce chips at 2-nanometer process nodes by 2027. This timeline requires RAPIDUS to compress into three years a capability development that took TSMC over a decade. Leading-edge semiconductor manufacturing is not a technical problem that can be solved by assembling talented engineers and providing adequate funding. It is an accumulated organizational capability, developed through the production of millions of wafers across multiple process generations, in which yield improvement, defect reduction, and process stability are learned through doing at scale rather than through planning.

The through line across all of these government failures is a single miscalibration: every major government has treated hardware sovereignty as a supply chain optimization problem rather than as a sovereignty problem in the full political sense. Rebuilding the capability to think about physical production as a dimension of national power, if it is possible at all, will take longer than the timeline on which the problem is developing.

Militaries

In the spring of 2022, Ukrainian forces used a Starlink terminal, a commercial software application running on a tablet, and a drone built partly from components ordered on AliExpress to direct artillery fire with a precision that would have required a dedicated signals intelligence battalion twenty years earlier. The improvisation was celebrated as evidence of Ukrainian ingenuity and the democratization of military technology. It was also a demonstration of something the celebration obscured: modern warfighting, at every level from the individual soldier to the theater commander, now runs on commercial semiconductor supply chains that military institutions do not control and have not seriously planned around.

The Ukrainian conflict has been the first large-scale war fought by both sides under conditions of semiconductor constraint. Russia entered the conflict with weapons systems whose guidance and targeting electronics depended on Western components it could no longer legally import. When stocks were depleted, Russian engineers substituted commercial chips sourced through third-country intermediaries, pulling semiconductors from refrigerators, washing machines, and automotive components to keep precision munitions functional. Russian precision strike accuracy degraded measurably over the course of the conflict. The degradation was not primarily the result of electronic warfare countermeasures. It was the result of a military power discovering that its weapons systems had been designed around semiconductor components it could no longer source at the required quality and quantity.

The kill chain for a modern precision munition requires semiconductors at every node. The targeting sensor runs on focal plane array detectors manufactured by a handful of specialized firms. The guidance processor runs on chips that, in American systems, are sourced through the trusted foundry program at process nodes two to three generations behind the commercial leading edge. The datalink connects the munition to the fire control system on radio frequency chips whose leading suppliers are concentrated in Taiwan and South Korea. The entire system, from sensor to warhead, is a stack of semiconductor dependencies, each one concentrated in a small number of suppliers, none of which are under military control.

The Department of Defense's trusted foundry program, which requires that chips used in classified systems be manufactured in facilities meeting specific security criteria, has produced a situation in which American military hardware operates on chips that are structurally inferior to what adversaries can acquire on the commercial market. The fabs certified under the trusted foundry program produce chips at process nodes of 14 nanometers and above, which is two to three generations behind TSMC's commercial leading edge. The kill chains for precision munitions, the processing units in reconnaissance satellites, the signal intelligence systems that underpin American military advantage are all running on hardware that a Chinese AI researcher can outperform with commercially available NVIDIA accelerators.

The Chinese People's Liberation Army has drawn the correct conclusions from this and is acting on them with a coherence that American defense institutions have not matched. China's military-civil fusion doctrine, codified in law in 2017, requires Chinese technology companies to make their capabilities, personnel, and data available to the military upon request. A Chinese AI company developing large language models on domestically produced hardware is, under military-civil fusion, simultaneously developing capabilities that the PLA can access and incorporate into military systems on timelines that parallel commercial development rather than lagging it by a decade.

The nuclear analogy is worth taking seriously here, not as a rhetorical device but as a genuine structural parallel. When nuclear weapons were developed, they created a strategic reality that existing military doctrine had no framework for. It took roughly two decades of strategic thinking to develop doctrines that made sense of the new reality. During those two decades, military and political institutions operated with frameworks that were inadequate to the weapons they controlled. The compute concentration described in this Almanac creates a strategic reality of comparable novelty. A nation that controls the physical capacity to produce leading-edge semiconductors controls the material foundation of modern military power in a way that has no direct precedent, and no military doctrine currently in use has been built to account for it.

International Bodies

The institutions examined in the preceding three chapters (markets, governments, militaries) are miscalibrated in ways that are, in principle, correctable. Markets can develop new instruments. Governments can restructure procurement. Militaries can revise doctrine. International bodies present a different problem. Their failure is not a miscalibration. It is a design constraint, built into their founding logic, that makes them structurally incapable of managing the problem this Almanac describes.

Every significant multilateral institution created after 1945 was built on a single foundational premise: that economic interdependence among nations reduces the probability of conflict, and that the role of international institutions is to deepen and govern that interdependence. The WTO exists to lower trade barriers. The IMF exists to stabilize the financial conditions that enable trade. The World Bank exists to extend the economic integration of the liberal order to developing nations. The entire architecture of postwar international governance was constructed as an elaboration of the software-defined worldview before that worldview had a name.

The semiconductor crisis does not fit inside that architecture. It cannot be made to fit, because it is produced by exactly the dynamic the architecture was designed to prevent: a situation in which deep economic integration has created dependencies so concentrated and so strategically consequential that the normal instruments of multilateral governance cannot manage them without choosing sides between major powers. No international institution is built to choose sides between major powers. That is not a failure of will. It is a design specification.

The American export controls on advanced semiconductors imposed beginning in October 2022 are, under a straightforward reading of WTO rules, trade-distorting measures inconsistent with the market access commitments the United States made upon joining the organization. China has not formally challenged them at the WTO, not because it lacks a legal case, but because the political conditions for adjudicating that case do not exist. The WTO's dispute settlement mechanism cannot process a dispute in which the responding party invokes national security under Article XXI and means it. The semiconductor export controls represent the most consequential invocation of Article XXI in the WTO's history, and the WTO has no mechanism to evaluate its legitimacy.

The Wassenaar Arrangement presents a more specific and more instructive failure. Wassenaar is a multilateral export control regime covering conventional weapons and dual-use goods and technologies, with 42 member states. China is not a member. The arrangement operates by consensus: any member can block the addition of any item to the control list. It was designed in 1996, when the most sensitive dual-use technology was night-vision equipment and certain machine tools. It was never designed to handle the need to coordinate export controls on the most consequential civilian technology in the world, sought by non-member states with whom member economies are deeply integrated. The consensus requirement means that any member state with commercial interests in selling semiconductor technology to China can delay or dilute control additions indefinitely.

The absence of an NPT equivalent for semiconductor technology is worth dwelling on. The Nuclear Non-Proliferation Treaty works because nuclear weapons have no significant civilian application that requires their spread. Semiconductors are the opposite. The chip that enables AI inference is the same chip that enables autonomous weapons guidance. The technology cannot be divided into civilian and weapons categories because it is foundationally dual-use in ways that nuclear technology is not. Any multilateral framework for semiconductor technology governance must either accept that it is managing a technology with direct military applications as a civilian matter, or accept that it is managing a technology with civilian applications as a military matter. Neither framing is workable at the multilateral level.

The standards bodies deserve attention as a separate category of failure. China has spent the better part of a decade systematically placing its nationals in leadership positions within international standards bodies, proposing standards that favor Chinese equipment architectures, and building coalitions of developing nations dependent on Chinese infrastructure investment to support Chinese positions in standards votes. The standards set today for 6G wireless networks, for AI system interoperability, and for the next generation of semiconductor interfaces will shape the architecture of the global technology system for decades. Those standards are being set in forums that operate on the assumption that technical consensus is possible between nations with fundamentally incompatible strategic interests in the outcome.

What emerges from surveying these institutions is not a story of individual failures. The failure is not of persons but of frameworks. Every institution examined in this part of the Almanac was built to manage a world organized around the premise that economic integration is the primary driver of international order. The semiconductor crisis is a structural challenge to integration itself, and the institutions designed to manage it have no tools for that scenario.