Introduction: Two Hearings, One Week, and the End of an Era

Consider a single week in July 2026, because it captures the entire argument of this paper in miniature. On Tuesday, July 14, Jeffrey Kessler, the U.S. Under Secretary of Commerce for Industry and Security, sat before the House Foreign Affairs Committee and confirmed that Nvidia had finally begun shipping its H200 artificial-intelligence processors to approved Chinese buyers — but only, in his words, “very few” of them, under a licensing regime so conditional, so contested, and so politically radioactive that four hundred thousand approved units had at one point sat frozen between a Commerce Department that had greenlit them and other arms of the American state that had not.[12][10] Exactly one week later, on July 21, the Financial Times reported that China’s Ministry of Commerce had been quietly consulting Alibaba, ByteDance, and Zhipu — the champions of China’s world-leading open-weight AI ecosystem — on a proposal that would have been unthinkable three years earlier: export controls on Chinese AI model weights, on the training data behind them, and on Chinese chip designs, including rules that could bar TSMC and Qualcomm from fabricating semiconductors designed by Huawei, Alibaba, or ByteDance.[20][21]

Read those two events together and the symmetry is almost uncanny. In the same month, the United States was rationing its hardware to China chip by chip, license by license, while China was preparing to ration its software, its data, and its design intellectual property to the world, model by model, weight file by weight file. Neither side was merely defending; both were withholding. Neither side commanded the whole stack; each commanded enough of it to hurt the other. This is not the world of 2018, when American export controls were a one-way instrument applied to a dependent rival. It is not even the world of October 2022, when the Biden administration’s sweeping semiconductor rules were described, accurately at the time, as a unilateral siege. It is something structurally new, and it deserves a name of its own.

This paper calls that condition Chokepoint Reciprocity. The term is chosen deliberately, and in deliberate contrast to the most influential framework in the existing literature. In their landmark 2019 article in International Security, Henry Farrell of Johns Hopkins SAIS (then at George Washington University) and Abraham Newman of Georgetown formalized the concept of “weaponized interdependence,” showing how states that sit astride the central nodes of global networks can convert that position into coercive power — states, they wrote, that

“can weaponize networks to gather information or choke off economic and information flows.” — Henry Farrell and Abraham L. Newman, International Security (2019) [1]

Weaponized interdependence, as originally theorized, described an asymmetric world: one dominant power — in practice, the United States, with its jurisdiction over the dollar, SWIFT, and the semiconductor toolchain — exploiting a centralized network against weaker parties.[2] Chokepoint Reciprocity describes what happens when the theory collides with a rival that read the same playbook, spent seven years and several hundred billion dollars building counter-chokepoints of its own, and then began pulling the levers. It is weaponized interdependence gone bidirectional: two states, two distinct sets of network hubs, two extraterritorial legal regimes, and a strategic logic that increasingly resembles Cold War deterrence more than trade policy.

The stakes of getting this right are not academic. The International Monetary Fund has estimated that severe geoeconomic fragmentation could cost the world economy up to 7 percent of global GDP, with losses reaching 8 to 12 percent of output in some countries once technological decoupling is layered on top of trade restrictions.[31] Speaking at the Stanford Institute for Economic Policy Research, the IMF’s First Deputy Managing Director put the warning in a single sentence:

“A very serious decoupling scenario could cost up to 7 percent of [global] GDP.” — Gita Gopinath, First Deputy Managing Director, International Monetary Fund, at Stanford SIEPR [30]

And yet, as this paper will argue, the same reciprocal structure that threatens those losses also contains the seeds of a strange stability. When each side can paralyze the other’s strategic industries, escalation becomes self-limiting; the October 2025 rare-earth confrontation and the one-year truce that followed it in Busan demonstrated exactly this dynamic in real time.[33] Harvard’s Graham Allison, whose Destined for War framed the U.S.–China rivalry as a Thucydidean struggle, has described the two powers as

“the fiercest competitors history has ever seen.” — Graham Allison, Douglas Dillon Professor of Government, Harvard University [32]

But even the fiercest competitors, Allison notes, are condemned by their entanglement to coexist. The pages that follow trace how that entanglement was weaponized from both ends: how unilateral American control collapsed (Section 1); how Washington rebuilt and revised its hardware blockade through 2026 (Section 2); how Beijing constructed its counter-arsenal across materials, markets, models, data, and design (Section 3); how the resulting system can be mapped as five reciprocal gates (Section 4); how deterrence now operates between the two capitals (Section 5); how the world’s most important technology firms are being pulled apart between two sovereigns (Section 6); how third countries and multinationals are adapting through multi-alignment (Section 7); how exposure can be measured through a proposed Reciprocal Chokepoint Index (Section 8); and, finally, what the whole record teaches, distilled into seven pillars (Section 9).

Section 1: The Collapse of Unilateral Technology Control

Every era of international order rests on an architecture, and the architecture of the digital era was astonishingly narrow. The modern semiconductor and AI industry was built on a foundation of American intellectual property and design software, European lithography, East Asian fabrication, and a global market that treated this division of labor as permanent. Three American firms dominated electronic design automation. A single Dutch company, ASML, monopolized the extreme-ultraviolet lithography machines without which no chip below roughly seven nanometers can be economically produced. One Taiwanese foundry, TSMC, fabricated the overwhelming majority of the world’s leading-edge logic. American firms designed the accelerators — above all Nvidia’s GPU lines — that became the engines of the deep-learning revolution. For Washington, this concentration was not a vulnerability but an armory. Because the critical nodes sat inside allied jurisdictions, the United States could project regulatory power through them: the Entity List, the Foreign Direct Product Rule, and the export-control system generally allowed one government to decide, in practice, which nations would be permitted to compute at the frontier.

The instrument was first tested at scale against Huawei between 2019 and 2020, when foreign direct product restrictions severed the company from TSMC and nearly destroyed its smartphone business. It was then generalized on October 7, 2022, when the Bureau of Industry and Security imposed country-wide controls on advanced computing chips, supercomputer end-uses, and semiconductor manufacturing equipment destined for China — a rule that Washington’s own officials described as a shift from maintaining a relative advantage to imposing an absolute ceiling. For a moment, the strategy appeared to vindicate the logic of the chokepoint: China imported chips worth more than its entire annual expenditure on crude oil, its national champions depended on American design software and Taiwanese fabrication, and its AI laboratories trained their models almost exclusively on Nvidia hardware. The siege, it seemed, could only tighten.

But unilateral control is a wasting asset, and it degrades along three predictable channels. The first is substitution: every restriction is simultaneously a subsidy to the restricted party’s domestic industry, because it guarantees local champions a captive market and a national mission. Beijing responded to the 2022 controls exactly as the theory of import substitution predicts, pouring state financing into SMIC, Huawei’s HiSilicon, CXMT, Cambricon, and the equipment and materials firms beneath them, and by 2025–2026 this investment was producing measurable output: Huawei planned to roughly double production of its flagship Ascend 910C accelerator to around 600,000 units in 2026, with as many as 1.6 million Ascend-line dies distributed across China’s AI sector.[18] The Center for a New American Security estimated China’s total 2026 domestic production at the equivalent of roughly 390,000 H200-class processors — far short of American hyperscale deployments, but no longer a rounding error.[11]

The second channel of degradation is enforcement leakage. Chips are small, fungible, and valuable; controls on them summon gray markets into existence. TechInsights teardowns revealed that essentially every early Ascend 910B and 910C sample contained dies fabricated at TSMC’s 7-nanometer node, acquired through a Cayman-registered intermediary in what U.S. authorities concluded was a sanctions-evasion scheme — a violation for which TSMC was fined one billion dollars, and which supplied Huawei with a die bank of roughly 2.9 million units that carried its shipments through 2024 and 2025.[35] Cloud access, smuggled accelerators, and overseas affiliates provided further seams, each of which Washington would spend 2023 through 2025 racing to stitch shut.

The third and most consequential channel is retaliation — the discovery by the targeted state of chokepoints running in the opposite direction. This was the channel Washington’s planners most consistently underweighted. China refines roughly 90 percent of the world’s rare earth elements and produces the overwhelming majority of its permanent magnets; it dominates gallium, germanium, and graphite processing; it is the world’s largest single market for semiconductors and the manufacturing hub through which Western electronics, automotive, and defense supply chains pass.[9] Beginning with gallium and germanium controls in 2023, accelerating through the April 2025 rare-earth restrictions, and culminating in the sweeping October 2025 regime analyzed in Section 3, Beijing converted these positions into legal instruments consciously modeled on America’s own — including, remarkably, Chinese versions of the de minimis rule and the Foreign Direct Product Rule, asserting jurisdiction over goods made outside China with Chinese inputs or technology.[6]

By the end of 2025, then, the unipolar architecture had collapsed into something the original theorists of weaponized interdependence had not fully anticipated: a two-hub network. The United States still commanded the physical high ground of the computing stack — leading-edge logic, advanced accelerators, design software, lithography access through its allies. But China now commanded the material floor beneath that stack, the largest demand market within it, and — as Sections 3 and 4 will show — an expanding share of the software and model layer above it. The conflict, to borrow a military metaphor, had evolved from a siege into a war of position: two entrenched powers, each holding ground the other cannot take, each capable of inflicting severe harm, and each increasingly aware that the other’s capacity for harm is the only durable limit on its own ambitions.

Section 2: The U.S. Toolset — Revising the Hardware Blockade

The United States continues to anchor its strategy in the physical layers of the computing stack, and to understand why, one must appreciate what those layers are worth. Advanced AI training is, at bottom, an exercise in concentrated energy and silicon: whoever controls the supply of leading-edge accelerators, the high-bandwidth memory stacked beside them, the networking that binds them into clusters, and the tools that fabricate all of the above controls the pace at which any nation can climb the capability curve. Washington’s export-control regime is best understood not as a single wall but as a living structure — repeatedly rebuilt, patched, extended, and occasionally partially dismantled — whose evolution between 2022 and 2026 reveals both the reach and the limits of American power. Four phases stand out.

Phase one: the ceiling (2022–2023). The October 2022 rules established compute-density and interconnect thresholds above which chips could not be sold to China, and cut Chinese fabs off from the equipment needed to produce logic below roughly 14–16 nanometers, DRAM below 18 nanometers, and NAND above 128 layers. Nvidia’s response was commercially rational and strategically corrosive: it engineered China-specific processors — first the A800 and H800, then the H20 — calibrated to sit just beneath each successive threshold. Washington answered in October 2023 by tightening the thresholds to capture the A800 and H800, forcing another redesign. The pattern that would define the entire era was already visible: every bright-line rule invites engineering to the line, and every instance of engineering to the line invites a new rule. Perpetual revision was built into the system’s DNA from the start.

Phase two: closing the seams (2024–2025). In December 2024, the Bureau of Industry and Security extended controls to high-bandwidth memory — the true physical bottleneck of the AI era — and added well over a hundred Chinese entities, including equipment makers, to the Entity List, while further restricting the software and hardware needed for advanced-node production. The significance of the HBM controls can hardly be overstated: SK Hynix alone controls roughly 60 percent of the global HBM market and is sold out through 2026, while Micron can satisfy only 50 to 66 percent of demand from its core customers, meaning that memory, not logic, is the binding constraint on both Chinese and Western accelerator production.[11] In parallel, Washington moved up the technology curve preemptively, signaling controls around next-generation transistor architectures — the gate-all-around structures that define the 2-nanometer era now entering volume production at TSMC — and around the electronic design automation tools required to design them, so that the frontier would be fenced before China reached it.[34]

The jurisdictional expansion of this phase deserves particular attention, because it is where the “extraterritorial trap” described later in this paper was fully sprung. In January 2025, the outgoing administration published the AI Diffusion Rule, an attempt to ration advanced compute for the entire planet through a three-tier country system; the incoming Trump administration rescinded it in May 2025 as unworkably broad, but replaced it with instruments no less extraterritorial. On September 29, 2025, BIS adopted the so-called Affiliates Rule, automatically extending Entity List restrictions to overseas subsidiaries majority-owned by listed Chinese companies — a rule aimed squarely at the cloud-renting and affiliate-procurement workarounds through which restricted parties had continued to access American silicon from third countries.[6] Chinese officials would later cite this September 29 expansion, explicitly and by date, as the proximate justification for their own extraterritorial rare-earth controls ten days later — a sequencing that captures the reciprocal logic of the era better than any theoretical model could.[6]

Phase three: the April 2025 shock and the price of maximalism. On April 9, 2025, the administration informed Nvidia that even the H20 — a chip deliberately designed to comply with every existing threshold — would henceforth require a license for export to China, citing supercomputer end-use risk. The financial consequences were immediate and, for the first time, large enough to be legible on the income statement of the world’s most valuable company: Nvidia recorded a $4.5 billion charge in its first fiscal quarter of 2026 for H20 inventory and purchase obligations, disclosed that it had sold $4.6 billion of H20 product in the quarter before the restriction landed, and guided that roughly $8.0 billion of expected H20 revenue would vanish from the following quarter.[5][4] Chief Executive Jensen Huang, who spent 2025 publicly campaigning against the policy, framed the outcome in market-share terms that no policymaker could comfortably ignore:

“We went from 95 percent market share to 0 percent.” — Jensen Huang, CEO, Nvidia, on the company’s position in China [17]

Huang’s protest contained a strategic argument as well as a commercial one: an American absence from the Chinese market, he warned, does not de-compute China; it merely transfers the market, the developer ecosystem, and the standards-setting power to Huawei. The events of late 2025, described in Section 3, would lend that warning considerable force — because by the time Washington reconsidered, Beijing had begun locking American chips out on its own initiative.

Phase four: the December 2025 reversal and the era of conditional flow (2025–2026). In December 2025, President Trump announced that the United States would permit Nvidia to sell the H200 — a genuinely powerful Hopper-generation processor, not a deliberately hobbled export variant — to approved Chinese customers. The formal BIS rule took effect on January 15–16, 2026, shifting license review for the H200 and AMD’s MI325X from a presumption of denial to case-by-case scrutiny, an adjustment that CNAS calculated as roughly a thirteen-fold increase in the computing power previously permitted for export, bounded by an aggregate cap on the order of 850,000 H200-equivalents and paired with conditions designed to keep domestic American supply prioritized.[10][11] Attached to the flow was an unprecedented fiscal claim: a 25 percent charge on China-bound H200 revenue, described in various accounts as a tariff and in others as a federal revenue share, layered atop an earlier 15 percent arrangement covering H20-class sales — the American state, in effect, taking equity in its own chokepoint.[36][39]

What followed demonstrated that in a reciprocal system, opening a gate is no easier than closing one. Congress rebelled: within days, the House Foreign Affairs Committee voted 42–2 to advance the AI OVERWATCH Act, which would ban Blackwell-class exports to China for two years, impose a 30-day congressional review on each H200 license, and grant Congress a standing veto over future AI-chip export decisions; by February 2026 a bipartisan group of lawmakers was demanding a full country-wide ban on semiconductor-equipment exports to China, and the era of Validated End-User exemptions that had allowed TSMC, Samsung, and SK Hynix to supply their own China fabs without individual licenses was brought to an end.[10][36] The Commerce Department approved roughly ten Chinese buyers — Alibaba, Tencent, ByteDance, and JD.com among them — for up to 75,000 H200 units each, later adding ZTE’s Kangxun unit and the server assembler Maginfra; and yet, as of mid-May 2026, not a single unit had been delivered, and as of Kessler’s July testimony the cumulative volume remained, in his phrase, “very few.”[38][13][12] Nvidia, for its part, excluded China compute revenue from its guidance entirely, quarter after quarter, treating the world’s second-largest AI market as a contingent asset of American foreign policy.[27]

Step back from the four phases and the shape of the American toolset becomes clear. It is, first, physical: anchored in accelerators, memory, equipment, and the foundry access of allies. It is, second, perpetually revisionist: the H200 rule of January 2026 was the fifth major recalibration of chip thresholds in forty months, and the tug-of-war between an executive inclined to monetize the chokepoint and a legislature inclined to weld it shut guarantees further revision. It is, third, extraterritorial by construction, reaching overseas affiliates, third-country fabs, and any product touched by American tools or software. And it is, fourth — this is the crucial point for the argument of this paper — no longer unanswered. Every phase after 2023 unfolded against Chinese counter-moves that shaped, constrained, and in the April–October 2025 sequence arguably provoked American policy. The blockade still stands, but it now stands opposite a blockade.

Section 3: The Chinese Counter-Strategy — Weaponizing the Material Floor and the Software Ceiling

Beijing’s strategic insight, arrived at gradually between 2018 and 2023 and executed with accelerating confidence thereafter, was that symmetry is a trap. China could not answer the American hardware blockade in kind, because it does not control leading-edge lithography, design software, or frontier accelerators; attempting to match Washington machine-for-machine at bottlenecks it does not hold would have been a war on the adversary’s chosen ground. Instead, China constructed asymmetric chokepoints at the layers of the stack where its position is genuinely dominant: the raw-material floor beneath the semiconductor industry, the demand market inside it, the legal and regulatory space around foreign firms operating in China, and — in the most recent and most conceptually significant development — the software, model, and design layer above the hardware, where Chinese open-weight AI has quietly become world infrastructure. Each instrument deserves careful examination, because together they constitute one half of the reciprocal system this paper describes.

The material floor: rare earths as a mirrored export-control regime. The escalation began in earnest on April 4, 2025, when China imposed licensing controls on heavy rare earth elements and permanent magnets in response to the spring tariff offensive; within weeks, automakers in the United States, Europe, and Japan reported disruptions severe enough to threaten production stoppages, and Washington negotiated a 90-day truce simply to restart shipments.[8] Then, on October 9, 2025 — ten days after the American Affiliates Rule — the Ministry of Commerce issued six coordinated announcements (Notices 55 through 62 of 2025) that transformed a licensing program into a full mirrored export-control regime: controls extended from materials to the technologies for mining, smelting, separation, magnet manufacturing, and recycling; licenses were declared presumptively unavailable for foreign defense users and for certain semiconductor applications, with case-by-case scrutiny for advanced-computing and AI end-uses; fourteen foreign entities were added to China’s Unreliable Entity List; and, for the first time, Beijing formally asserted extraterritorial jurisdiction through its own de minimis and foreign-direct-product mechanisms, reaching goods manufactured outside China with Chinese rare-earth content or technology, effective December 1, 2025.[6][7][9][43]

The architecture was unmistakably learned from Washington — legal scholars at Mayer Brown noted that the regulatory design tracked the American FDPR framework almost clause for clause — but the underlying leverage was distinctively Chinese: roughly 90 percent of global rare-earth processing and magnet production sits inside the People’s Republic, a concentration greater than OPEC ever achieved in oil.[6][9] The empirical record of 2025–2026 shows the weapon working as designed. Chinese customs data analyzed by CSIS show that exports of yttrium — a thermal-coating material without which jet engines cannot be built — to the United States collapsed from 333 tons in the eight months before restrictions to 17 tons in the eight months after, forcing aerospace manufacturers into rationing.[8] Chris Miller of Tufts University’s Fletcher School, the author of Chip War and the most closely read chronicler of this conflict, captured the strategic intent within hours of the October announcement:

“China is signaling ‘we’re willing to threaten your primary growth driver.’” — Chris Miller, Tufts University, author of Chip War, to Axios [3]

That primary growth driver is, of course, artificial intelligence itself — the entire American AI buildout runs on magnets, motors, machined components, and materials that pass through Chinese processing. Miller has separately observed that Beijing’s licensing system does more than restrict; it surveils, giving China supply-chain visibility that enables ever more precise future targeting, while deliberate drip-feeding keeps foreign manufacturers dependent rather than driving them to full substitution:

“‘Just in time’ is central to Beijing’s strategy.” — Chris Miller, in the Financial Times, on China’s rare-earth licensing [34]

The Nexperia affair of late 2025 demonstrated a second face of the same leverage. When the Dutch government seized control of the Chinese-owned chipmaker on security grounds, Beijing paused the export of Nexperia’s China-packaged chips — legacy components used by roughly half of European automakers and virtually the entire European defense industry — and the company’s Chinese unit openly defied instructions from its own Dutch headquarters, an episode that taught European industry how little the distinction between “advanced” and “legacy” semiconductors matters when the legacy parts are the ones your assembly lines cannot run without.[40][34]

The demand market as a weapon: the great expulsion of American silicon. The second instrument is subtler than an export ban because it is an import ban — the deliberate weaponization of China’s own purchasing power. The sequence through 2025 was methodical. In August, regulators required new data centers to source at least half their accelerators domestically, even as Washington was arranging to take a 15 percent cut of resumed H20 sales. In September, the Cyberspace Administration ordered major internet platforms to stop purchasing Nvidia processors, citing security vulnerabilities, while the State Administration for Market Regulation announced that Nvidia had violated the Anti-Monopoly Law in connection with its Mellanox acquisition — a finding held in reserve as legal ammunition.[15][17] In November, Reuters revealed that state-funded data-center projects — a category covering most Chinese data centers, given more than $100 billion in government backing since 2021 — were barred from foreign AI chips entirely, with projects less than 30 percent complete ordered to rip out Nvidia, AMD, and Intel hardware already installed.[14] In December, days before Washington’s H200 announcement, the Ministry of Industry and Information Technology converted encouragement into a written order, adding Huawei and Cambricon chips to the Xinchuang state-procurement list and directing agencies, state-owned enterprises, schools, and hospitals to buy domestic.[42] And by June 2026, Bloomberg reported that the National Development and Reform Commission was drafting a five-year, two-trillion-yuan (roughly $295 billion) national computing-grid program mandating at least 80 percent domestic content — a procurement of historic scale from which Nvidia and AMD are structurally excluded.[16]

The effect of this expulsion campaign is to convert Jensen Huang’s warning into settled fact. Nvidia’s share of the Chinese AI-accelerator market, above 90 percent as recently as 2023, had fallen to roughly half by early 2026 and continues to slide toward zero in state-adjacent demand; Cambricon swung from losses to record profits on revenue growth exceeding 4,000 percent; and Huawei’s Ascend line, backstopped by SMIC’s enhanced 7-nanometer process, is being deployed in CloudMatrix clusters that substitute scale and power consumption for single-chip performance.[36][17][35] The constraint on this substitution is real — SemiAnalysis calculates that China’s Ascend output is bottlenecked not by logic dies, of which SMIC can now produce more than a million annually, but by high-bandwidth memory, with domestic CXMT capacity sufficient for only a few hundred thousand full packages in 2026 once foreign HBM stockpiles are exhausted — but the direction is unambiguous.[19][35] Beijing has decided that dependence on American accelerators is a strategic vulnerability to be eliminated at whatever near-term cost in compute efficiency, and it is willing to enforce that decision with the full procurement power of the state.

The software ceiling: open weights as standards power. The third instrument was not, at first, an instrument at all; it was an emergent property of Chinese AI strategy that Beijing only later recognized as leverage. Following the DeepSeek moment of January 2025, Chinese laboratories embraced open-weight release as their route around both American compute advantages and Western market incumbency — and the results, by 2026, rewired the global AI ecosystem. Alibaba’s Qwen family alone generated 153.6 million downloads in February 2026, more than its next eight competitors combined; by March it accounted for over half of all global open-source model downloads, with more than 180,000 derivative models on Hugging Face — more than Google and Meta combined; Stanford and Berkeley researchers were fine-tuning state-of-the-art systems on Qwen bases for as little as $30 to $50; Singapore selected Qwen over Meta’s Llama as the foundation of its sovereign regional model; and MIT Technology Review’s analysis found that Chinese open-weight models had overtaken American ones as a share of global downloads for the first time.[23] DeepSeek became the most-followed organization on Hugging Face, with Qwen fourth, and the most-read technical papers on the platform now come predominantly from ByteDance, DeepSeek, Tencent, and Qwen.[24] Moonshot AI’s Kimi K3 release in July 2026 — promptly followed by reports that Washington was reviving efforts to ban Chinese models from U.S. government use — confirmed that the frontier gap had narrowed to months.[21][22]

It is against this backdrop that the July 21, 2026 Financial Times report must be read. According to the report, the Ministry of Commerce has consulted Alibaba, ByteDance, and Zhipu on restricting the transfer of key training data abroad and on limiting foreign downloads of model weights for the most advanced systems — while preserving overseas access through hosted APIs, which keep usage visible, revocable, and monetizable; it has sought views on barring overseas foundries, explicitly including TSMC and Qualcomm’s manufacturing chain, from fabricating advanced chips based on Chinese designs from Huawei, Alibaba, and ByteDance; and it is weighing tighter screening of foreign acquisitions of Chinese technology firms, a concern crystallized by Meta’s roughly $2 billion acquisition of the agentic-AI startup Manus, which Chinese authorities ordered unwound.[20][21][22] The measures would enter China’s catalogue of technologies prohibited or restricted from export — the same instrument that already covers rare-earth processing and battery technology — and industry participants have reportedly warned that overreach could sacrifice the global adoption that makes Chinese models strategically valuable in the first place.[20][22]

Whether or not every proposal is enacted, the doctrinal shift is the point. A chip design withheld from TSMC is the mirror image of an EUV machine withheld from SMIC. A weight file withheld from foreign download is the mirror image of an H100 withheld from a Chinese cloud. Training-data controls mirror the compute controls of 2022. Acquisition screening mirrors CFIUS. Layer by layer, instrument by instrument, Beijing has now either built or announced the intention to build a reciprocal of every major mechanism in the American arsenal — anchored not in the hardware bottlenecks it lacks, but in the materials, markets, models, data, and designs it commands. The two-way structure is complete. What remains is to map it.

Section 4: The Five Gates of Reciprocal Control

If Sections 2 and 3 described the two arsenals historically, this section maps them structurally. The central analytical claim of Chokepoint Reciprocity is that the U.S.–China technology conflict is now fought across five distinct gates — five layers of the technology stack at which one state or the other (and increasingly both) can grant or deny passage. The gates are not symmetrical: at some, American control is overwhelming; at others, Chinese control is; at several, control is contested or divided by sub-layer. But every gate now has a gatekeeper on each side of it, and every serious policy move by either capital since 2024 can be located at one or more of them. Understanding the gates individually, and then observing how they interlock, is the fastest route to understanding the whole system.

Gate One — Hardware. This is the gate Washington built first and defends most fiercely: frontier accelerators (Blackwell-class and beyond remain flatly prohibited to China, with even the Hopper-generation H200 flowing only in “very few” units under contested case-by-case licenses), high-bandwidth memory (controlled since December 2024, and the single most binding constraint on accelerator production everywhere), advanced networking, and the semiconductor manufacturing equipment of the American, Japanese, and Dutch toolchain.[10][11][12] Yet even here reciprocity intrudes: the magnets, rare-earth materials, and specialty inputs inside that same equipment — and inside the robots, drives, and cooling systems of the data centers the chips populate — pass through the Chinese material floor, so that the hardware gate’s foundations rest partly on ground the adversary controls.[3][8]

Gate Two — Foundry. The question at this gate is deceptively simple: who may fabricate whose designs, and in which jurisdiction? Washington answers through the Foreign Direct Product Rule, which reaches any fab on earth that uses American tools; through the termination of the VEU exemptions that had let TSMC, Samsung, and SK Hynix operate their China facilities smoothly; and through the $1 billion penalty imposed on TSMC for the Sophgo die-diversion affair, which announced that foundry compliance would be enforced against allies as vigorously as against adversaries.[36][35] Beijing’s answer, per the July 2026 consultations, would be a mirrored rule: Chinese designs from Huawei, Alibaba, and ByteDance could be barred from foreign fabrication without government approval, forcibly consolidating Chinese design demand at SMIC — accepting a two-to-three generation process penalty today in exchange for guaranteeing SMIC the order volume that funds tomorrow’s catch-up.[21][20] Each side, in other words, now treats foundry access itself as a licensable strategic act.

Gate Three — Model. Here the polarity reverses. The United States restricts hardware but has, to date, left model weights largely uncontrolled; its frontier labs simply do not release them. China’s laboratories do release them — and in doing so have captured the open ecosystem: majority share of global open-source downloads, the largest derivative ecosystems, adoption as sovereign infrastructure from Singapore outward.[23][24] The proposed Chinese controls on weight downloads and the parallel American efforts to bar Chinese models from government systems after the Kimi K3 launch mark the model gate’s formal militarization from both directions at once.[20][21] The strategic asset at this gate is not the file itself but the standard: whoever’s models the world builds upon sets the defaults — architectural, linguistic, and ultimately political — for the AI era.

Gate Four — Data. Training corpora, scientific datasets, geospatial information, and industrial records are the least visible and least formalized gate, but the July 2026 proposals would change that, restricting the transfer abroad of key training data as a controlled export.[20][22] China’s advantages here are its population scale, its industrial density — the process data of the world’s factory floor — and a domestic data-governance regime (the Cybersecurity Law, Data Security Law, and Personal Information Protection Law) that already treats data as a sovereign resource requiring security review before export. The United States, for its part, has moved to wall off American genomic, geolocation, and personal data from Chinese access. The data gate is where the two states’ instruments most resemble each other, because both began regulating data as territory before they began regulating it as trade.

Gate Five — Acquisition. The final gate governs the purchase of capability itself: firms, talent, and intellectual property. The American side of the gate — CFIUS review, outbound-investment screening, Entity List designations, visa restrictions — is decades old and well documented. The Chinese side snapped into focus with the Manus affair: Meta’s roughly $2 billion acquisition of the Chinese-founded agentic-AI startup was ordered undone, and the July 2026 consultations explicitly contemplate systematic screening of foreign acquisitions of strategic Chinese technology companies, closing what Beijing regards as a loophole through which frontier capability might be bought rather than built.[21][20] Talent flows — the quiet repatriation of Chinese researchers, the tightening of exchange in both directions — run through this gate as well, and may in the long run matter more than any single transaction.

GatePrimary U.S. InstrumentsPrimary Chinese InstrumentsBalance of Control (2026)
1. HardwareChip/HBM export thresholds; equipment controls; Entity List; AI OVERWATCH Act proposalsRare-earth and material controls with extraterritorial FDPR; Unreliable Entity ListU.S. dominant at frontier logic and memory; China dominant at material inputs
2. FoundryForeign Direct Product Rule; VEU revocation; enforcement fines (TSMC, $1B)Proposed ban on foreign fabrication of Chinese designs; SMIC consolidationU.S. dominant via allied fabs; China building autarkic counter-position
3. ModelClosed frontier weights; proposed federal bans on Chinese modelsProposed weight-download and export-catalogue controls; strategic open-sourcingChina leads open ecosystem; U.S. leads closed frontier
4. DataGenomic/geolocation data walls; cloud KYC proposalsTraining-data export controls (proposed); DSL/PIPL security reviewsDivided; both regimes hardening
5. AcquisitionCFIUS; outbound investment screening; visa policyManus unwind; proposed foreign-acquisition screening; talent repatriationConverging on mutual closure

The table above compresses the argument, but one dynamic property must be added to the static map: the gates interlock. A restriction at one gate migrates pressure to the others. When the hardware gate closed to China, Chinese strategy climbed to the model gate (open weights) and descended to the material floor (rare earths). When Chinese models threatened to become world standards, American policy moved to the model and data gates. When Washington policed the foundry gate against Huawei’s die diversion, Beijing began designing a foundry gate of its own. The system behaves less like a wall than like a hydraulic network: pressure applied anywhere redistributes everywhere, which is precisely why the conflict can no longer be understood — or managed — one export rule at a time.

Section 5: The Mechanics of Reciprocal Deterrence

When two nations each possess the ability to paralyze the other’s strategic industries, the logic of economic conflict changes in kind, not merely in degree, and the closest available analogy — imperfect but illuminating — is the nuclear deterrence theory of the Cold War. This section takes the analogy seriously enough to specify where it holds and where it breaks, because the difference between the two determines whether Chokepoint Reciprocity stabilizes the U.S.–China relationship or merely reschedules its crises.

The analogy holds in three respects. First, both sides now possess assured retaliation. The October 2025 sequence proved it empirically: Washington’s September 29 Affiliates Rule was answered within ten days by the most sweeping export-control action in Chinese history, aimed with precision at the materials on which American AI, automotive, and defense production depend; American officials, by their own account, were negotiating an emergency truce in Busan within weeks, suspending both sides’ measures for one year.[6][33] A White House that had spent three years assuming escalation dominance discovered that the other side could impose acute, immediate, politically intolerable pain — the criterion Chris Miller and former NSC official Chris McGuire identified as the only currency that moves Beijing, and now, demonstrably, the currency Beijing holds against Washington as well.[37]

Second, the deterrence is asymmetric by design, which — as Thomas Schelling taught — is what makes it stable rather than merely dangerous. China does not threaten American chips with Chinese chips; it threatens American growth with material scarcity, targeting yttrium for jet engines, magnets for motors and drives, and legacy microcontrollers for automotive lines, while the United States threatens Chinese ambition with compute scarcity, targeting the accelerators and memory that gate frontier AI.[3][8][11] Because the arsenals do not overlap, neither side can neutralize the other’s weapon by improving its own; each can only build painful countermeasures slowly (rare-earth friendshoring on one side, HBM indigenization on the other) while the deterrent holds in the interim.

Third, the system exhibits the crisis-truce-crisis rhythm characteristic of mutual deterrence rather than the cumulative squeeze characteristic of unilateral siege. April 2025: rare-earth restrictions, emergency 90-day truce. October 2025: reciprocal escalation, Busan one-year suspension. December 2025 through July 2026: conditional reopening of the H200 channel, congressional counter-mobilization, Chinese counter-mandates, and a standing awareness in both capitals that the Busan clock expires in late 2026.[8][33][10] Every escalation has been met by an equivalent, asymmetric response, and every response has been followed by negotiation — the tit-for-tat equilibrium that game theory predicts between players who know they must play again.

But the analogy breaks in two places, and the breaks define the residual danger. The first is that economic weapons, unlike nuclear ones, are used continuously at low intensity, which means the system’s participants are perpetually testing thresholds rather than respecting a bright line; Miller’s observation that China remains in the “early stages” of learning what its licenses can do implies a long period of probing, miscalibration, and accident risk.[3] The second is that the arsenals decay. American controls erode as SMIC, CXMT, and Huawei climb the curve; Chinese material leverage erodes as MP Materials, Lynas, and allied processing capacity come online — Miller himself argues Beijing’s rare-earth monopoly is durable in the short run but not the long run.[3][37] Deterrents that decay at different rates invite exactly the preventive logic that made pre-1914 Europe unstable: the temptation to use leverage before it depreciates. The Busan truce, on this reading, is less a settlement than a mutually agreed pause in a race whose participants are simultaneously deterring each other and sprinting to escape the deterrence.

And yet the paradox at the heart of this paper remains: the same interdependence that each side is racing to escape is, for now, the strongest force preventing catastrophe. Complete decoupling is impossible under this structure precisely because each side relies on distinct assets controlled by the other — American AI on Chinese materials and markets, Chinese AI on American-designed compute and allied fabrication — and the IMF’s fragmentation estimates quantify what rupture would cost: up to 7 percent of global output, and 8 to 12 percent in the most exposed economies under full technological decoupling.[31][30] Mutual vulnerability, in the Chokepoint Reciprocity era, is not the problem to be solved. It is the buffer that keeps the problem from becoming a war.

Section 6: Corporate Actors Between Two Sovereigns — The Earnings Record Through Mid-2026

Structural theories are cheap; balance sheets are expensive. If Chokepoint Reciprocity is real, it should be visible in the audited financial statements of the firms standing at the gates — and it is, with a clarity that no policy document can match. This section reads the corporate record through the second quarter of calendar 2026, because the numbers reveal the two deepest truths of the era simultaneously: that the reciprocal conflict imposes enormous, measurable costs on individual firms, and that the AI supercycle has so far been powerful enough to pay those costs and keep growing — a fact that itself shapes how much escalation both governments believe they can afford.

Nvidia: the bellwether of the hardware gate. No company’s financials narrate the conflict more precisely. The April 2025 H20 licensing requirement produced a $4.5 billion inventory charge in the quarter ended April 27, 2025, against $4.6 billion of H20 revenue booked before the door closed, with a further $8 billion of expected sales erased from the following quarter’s guidance.[5][4] Yet by its first quarter of fiscal 2027, reported May 20, 2026, Nvidia posted record revenue, record data-center revenue, and record free cash flow — with China compute contributing exactly nothing and excluded from guidance entirely, since no H200 had shipped despite the December reversal, the roughly ten approved buyers, and the 75,000-unit-per-customer allocations.[27][26][38] Jensen Huang, who had joined President Trump’s Beijing delegation just six days before the earnings call, closed it with a declaration that explains why Washington believes it can afford continued restriction:

“Demand has gone parabolic. The reason is simple: Agentic AI has arrived.” — Jensen Huang, CEO, Nvidia, Q1 FY2027 earnings call, May 20, 2026 [25]

The subtext deserves to be made text: the world’s most valuable company reached a market capitalization above five trillion dollars with zero China revenue, which simultaneously vindicates the hawks (the chokepoint costs America less than feared) and the doves’ warning (the Chinese market, once 95 percent Nvidia’s, is being permanently resettled by Huawei, Cambricon, and the domestic-mandate state).[39][17] By July 2026, with Kessler confirming “very few” actual shipments and Beijing’s procurement bans hardening, the most plausible reading is that both governments — not the market — now jointly determine Nvidia’s China business, and both have chosen, for different reasons, to keep it near zero.[12][14][42]

TSMC: the contested prize of the foundry gate. Taiwan Semiconductor’s second-quarter 2026 results, reported July 16, are the single best evidence that the AI megatrend is absorbing the costs of bifurcation: record revenue of $40.2 billion (NT$1.27 trillion), up 36 percent year over year; record net income of NT$706.56 billion, up 77.4 percent; gross margin of 67.7 percent; the first meaningful revenue from the 2-nanometer N2 node at 3 percent of wafer sales; full-year capital expenditure guidance raised to $60–64 billion; and a fresh $100 billion commitment to Arizona capacity on top of everything previously announced.[28][29][41] Chief Executive C.C. Wei’s framing was serene:

“The AI megatrend continues to drive the need for more and more computation.” — C.C. Wei, Chairman and CEO, TSMC, Q2 2026 earnings call [28]

But serenity at the income statement coexists with maximal exposure at every gate. TSMC sits inside the American hardware and foundry gates (its tools are American-controlled, its China fabs lost VEU status, and it paid a $1 billion penalty over the Huawei die diversion); inside the prospective Chinese foundry gate (the July 2026 proposals would strip it of Chinese design customers by law); inside the material gate (Chinese rare-earth rules explicitly scrutinize semiconductor end-uses); and at the center of the geographic risk that shadows the entire system — the Taiwan contingency that the AI OVERWATCH debate, with its references to 2027 readiness timelines, keeps permanently in view.[36][35][21][9][10] The Arizona expansion is thus best understood not as commercial diversification but as the foundry gate physically relocating itself toward the safer sovereign.

Huawei, SMIC, and the substitution complex. On the Chinese side of the ledger, the story is production physics. Huawei’s planned 600,000 Ascend 910C units for 2026 — up to 1.6 million dies across the Ascend line — represent the substitution ceiling permitted by two constraints: SMIC’s 7-nanometer logic capacity, which is no longer binding, and high-bandwidth memory, which is.[18][19] SemiAnalysis calculates that China procured roughly 13 million HBM stacks before foreign supply closed — sufficient for about 1.6 million 910C packages — after which domestic CXMT output of roughly 2 million stacks per year supports only 250,000–300,000 packages annually until Chinese HBM scales; the exhaustion of the TSMC die bank in early 2026 makes this the decisive constraint of the decade’s second half.[19][35] Cambricon’s swing to record profitability on 4,300 percent revenue growth, and the CloudMatrix architecture’s brute-force answer to Nvidia’s rack-scale systems (roughly double the aggregate throughput at four times the power), complete the picture of an ecosystem that is inferior per chip, inefficient per watt, and nonetheless strategically sufficient — because the mandate economy guarantees it demand at any price.[17][35][16]

Alibaba, ByteDance, Zhipu, and the model-gate champions. The final group of corporate actors holds the newest asset class: globally adopted Chinese AI. Alibaba’s hundred-plus Apache-licensed Qwen releases, ByteDance’s research dominance on Hugging Face, Zhipu’s frontier systems, DeepSeek’s ecosystem leadership, and Moonshot’s Kimi K3 collectively constitute the leverage that the July 2026 export-control consultations propose to formalize — and these same firms are simultaneously the approved H200 buyers on the American side, the mandated Ascend customers on the Chinese side, and the design houses whose access to TSMC the new rules would sever.[23][24][20][38] No corporate cohort in history has stood inside so many gates at once, which is precisely why Beijing consulted them, and why their reported warning — that overreach could forfeit the global adoption on which model-gate leverage depends — may prove the most consequential lobbying of the era.[22]

CompanyLatest Result (through Q2 2026)Chokepoint ExposureGate Position
NvidiaRecord Q1 FY27 revenue (May 2026); $4.5B H20 charge (2025); zero China in guidanceU.S. licenses + Chinese procurement bansHardware (subject of both regimes)
TSMCQ2 2026 revenue $40.2B, +36% y/y; net income +77.4%; capex $60–64BFDPR, VEU loss, $1B fine; proposed Chinese design banFoundry (contested prize)
Huawei / SMIC~600K Ascend 910C planned 2026; up to 1.6M dies; HBM-constrainedU.S. Entity List; domestic-mandate beneficiaryHardware/Foundry (substitution engine)
SK Hynix / MicronHBM sold out through 2026; 50–66% of core demand metHBM export controls; memory scarcityHardware (memory bottleneck)
Alibaba / ByteDance / ZhipuQwen >50% of global open-model downloads; 180K+ derivativesProposed weight/data controls; approved H200 buyersModel/Data (new leverage class)
CambriconRecord profits; ~4,300% revenue growth (H1 2025)Beneficiary of expulsion mandatesHardware (domestic substitute)

Section 7: Global Multi-Alignment and Third-Party Enclaves

A bipolar chokepoint system does not merely constrain its two principals; it restructures the strategic options of everyone else, and the deepest long-term consequences of Chokepoint Reciprocity may lie here, in the choices now being forced on allied and non-aligned states and on the multinational firms headquartered inside them. The old strategy — alignment with the single hegemon that controlled the stack — is no longer available, because there is no longer a single hegemon and no longer a single stack. What has replaced it is a strategy this paper calls multi-alignment: the deliberate cultivation of position within both spheres, redundantly, expensively, and often covertly.

Consider the dilemmas jurisdiction by jurisdiction. Taiwan hosts the most valuable single node in either network and is therefore the most exposed polity on earth: American policy simultaneously depends on TSMC, fines it, strips its China exemptions, and pays it to replicate itself in Arizona, while Chinese policy simultaneously depends on it, targets it for design-ban decoupling, and claims sovereignty over the ground it stands on.[36][35][21] South Korea lives the memory version of the same dilemma: SK Hynix’s HBM dominance makes it indispensable to the American AI buildout and therefore a primary target of American supply-priority conditions, even as its firms’ China fabs lost VEU protection and its materials chain runs through Chinese processing.[11][36] Japan and the Netherlands contribute the equipment layer of the American gate and absorb Chinese material retaliation in exchange — the Nexperia seizure and its aftermath demonstrated to The Hague that enforcement actions taken inside Europe can be answered within days by export pauses that idle European assembly lines.[40][34] Malaysia and Singapore occupy the most delicate position of all: as the primary Southeast Asian transshipment, packaging, and data-center hubs, they are simultaneously the enforcement frontier of American affiliate rules, the natural home of Chinese overseas capacity, and — in Singapore’s choice of Qwen as sovereign-model foundation — early adopters of the Chinese model gate’s exports.[6][23] The Gulf states are purchasing their way into both networks at once, importing American accelerators under bilateral compute agreements while courting Chinese models and infrastructure, wagering that capital can buy exemption from the choice itself.

For multinational firms, multi-alignment has a harder edge: it means the end of the single global architecture as an organizing principle of the technology corporation. A firm of consequence must now maintain one supply chain compliant with American hardware restrictions and a second insulated from Chinese material and data controls; one product line eligible for the American governmental market and another for the Xinchuang list; one model stack cleared of Chinese weights for Washington’s comfort and another localized for Beijing’s. The bifurcation is not hypothetical: it is visible in TSMC’s $165-billion-plus American buildout, in Nvidia’s parallel maintenance of a China product organization that has had almost nothing to sell for eighteen months, in the ripped-out accelerators of half-built Chinese data centers, and in the compliance departments that now shadow every engineering decision at every major firm.[28][14][27] The IMF’s warning that technological decoupling amplifies fragmentation losses to 8–12 percent of GDP in exposed countries is, at the firm level, simply the sum of millions of such redundancies.[31]

Third-party enclaves, finally, are where the two extraterritorial regimes now collide directly. An American affiliate rule reaching a Chinese subsidiary in Kuala Lumpur; a Chinese FDPR reaching a magnet shipped from Frankfurt; a weight file downloaded in Nairobi from a Hangzhou lab; a Taiwanese fab in Phoenix fabricating for a customer in Shenzhen — each is a point where two sovereignties assert simultaneous jurisdiction over the same transaction, and international law offers no settled rule for the collision. The likeliest medium-term outcome is not resolution but zoning: a gradual, negotiated, permanently unstable division of the world’s technology geography into American-gated space, Chinese-gated space, and a shrinking commons in which multi-aligned states auction their neutrality to both sides. That commons — not Washington and not Beijing — is where the price of Chokepoint Reciprocity will ultimately be paid.

Section 8: Measuring Exposure — A Proposed Reciprocal Chokepoint Index

If the preceding sections are correct that chokepoint exposure is now the master variable of technological statecraft, then it should be measured with the same discipline that trade economists bring to tariff incidence or that central banks bring to financial stress. This section sketches a Reciprocal Chokepoint Index (RCI) — a first-order framework, offered for refinement rather than as a finished instrument — that scores each country’s exposure to the two gated systems simultaneously. The index is built from five components, one per gate, each scored in both directions: exposure to American denial (how much of the country’s technological metabolism passes through U.S.-controlled hardware, foundry, model, data, and acquisition gates) and exposure to Chinese denial (the mirrored question for materials, market access, models, data, and investment). Each component is weighted by substitutability — the realistic time and cost of replacing the gated input — because raw dependence overstates exposure where alternatives exist and understates it where they do not; yttrium at 95 percent Chinese supply with no Western processing is a different order of exposure than solar panels at similar concentration with idle capacity elsewhere.[8][31]

Formally, for each country c, RCI(c) is the weighted sum across the five gates of (dependence × criticality × inverse substitutability), computed separately against each gatekeeper, yielding a two-dimensional score: RCI-US and RCI-CN. The two dimensions matter more than their sum, because strategy differs by quadrant. Countries high on both axes (Taiwan, South Korea, Germany, Japan) are dual hostages whose rational strategy is hedged indispensability — making themselves too valuable for either side to squeeze. Countries high only on the American axis (much of the Gulf) rationally buy Chinese optionality; countries high only on the Chinese axis (resource-processing dependents across the Global South) rationally court American alternatives; countries low on both (a shrinking set) can auction alignment. The United States and China themselves carry nonzero scores on each other’s axes — America’s RCI-CN driven by materials, magnets, and legacy chips; China’s RCI-US driven by HBM, EDA, equipment, and frontier accelerators — and the entire deterrence analysis of Section 5 can be restated as the observation that neither principal’s score against the other is anywhere near zero, nor will be this decade.[19][11][3]

EconomyRCI-US Exposure (illustrative)RCI-CN Exposure (illustrative)Dominant Gate of VulnerabilityStrategic Posture Observed
TaiwanVery HighVery HighFoundry (both directions)Hedged indispensability; U.S. capacity migration
South KoreaVery HighHighHardware (memory) / MaterialsAlliance compliance plus China-fab triage
JapanHighHighEquipment / MaterialsAllied controls plus material friendshoring
Germany/EUModerateVery HighMaterials / Legacy chips (Nexperia)De-risking rhetoric, drip-feed dependence
Singapore/MalaysiaHighHighEnforcement frontier / Model adoptionActive multi-alignment
Gulf StatesHighModerateHardware (compute imports)Capital-funded dual access
United States—HighMaterials / Magnets / Legacy chipsFriendshoring race against truce clock
ChinaHigh—Hardware (HBM, EDA, equipment)Mandated substitution at any cost

Three uses justify the index even in this rough form. First, it disciplines public debate: claims that either power can “win” the technology war dissolve on contact with a two-axis exposure map that shows both principals deeply scored against the other. Second, it prices policy: any proposed restriction can be evaluated by how much it raises the adversary’s exposure relative to how much retaliation it invites against one’s own — the asymmetric-cost criterion that Miller identifies as the only real currency among the three largest economies.[34][37] Third, it forecasts alignment: the quadrant a country occupies predicts its behavior better than its treaty commitments do, which is why formally allied economies keep hedging and formally non-aligned ones keep converging on identical dual-track architectures. The RCI, in short, is the balance-of-power table of the chokepoint era — and like the naval tonnage tables of an earlier age, its chief value is to remind statesmen of what they cannot afford.

Section 9: What Have We Learned? The Seven Pillars of Chokepoint Reciprocity

Every framework must earn its keep by compressing a complicated record into transportable principles, and the 2020–2026 record examined in this paper compresses into seven. The original five pillars of the Chokepoint Reciprocity thesis — asymmetry, revisionism, extraterritoriality, software as kinetic, and interdependence as buffer — survive contact with the evidence; the events of 2025–2026 add two more.

Pillar 1: Asymmetry of Assets.

Reciprocity does not mean similarity. The United States controls the physical bottlenecks of high-end compute — accelerators, high-bandwidth memory, EDA software, lithography access through allies, and the foundry network’s tooling layer. China controls the material floor (roughly 90 percent of rare-earth processing), the largest demand market, the world’s dominant open-model ecosystem, continental-scale training data, and, prospectively, the export of its own designs and weights.[11][9][23] The arsenals are non-overlapping by construction, which is precisely what makes them mutually deterrent rather than mutually cancelling: neither side can build its way out of the other’s leverage quickly, and both know it.

Pillar 2: Perpetual Revisionism.

The boundaries of technological chokepoints are never static, because every bright-line rule summons engineering to the line. Five major recalibrations of American chip thresholds in forty months; Chinese catalogue revisions extending from rare-earth technology to battery lines to, prospectively, model weights; the H20 designed to the rule and then banned, the H200 banned and then conditionally unbanned and then congressionally besieged — the record shows that a chokepoint regime is not a law but a metabolism, and states that cannot revise continuously cannot hold the gate at all.[36][10][20]

Pillar 3: The Extraterritorial Trap.

Chokepoints cannot be contained within domestic borders, because the networks they gate are global; therefore every serious chokepoint regime is driven, step by inexorable step, to claim jurisdiction over other people’s territory. The American FDPR and Affiliates Rule reach fabs in Taiwan and subsidiaries in Malaysia; China’s December 2025 extraterritorial rare-earth rules reach magnets made in Germany from Chinese inputs; and each regime now cites the other as its justification.[6][36] True leverage, the record shows, is defined less by what a nation makes than by how far its regulatory writ can be made to run — and the collision of two such writs over the same third-country transaction is the signature legal event of the era.

Pillar 4: Software as a Geopolitical Kinetic.

Geopolitical leverage has climbed the computing stack. Restricting physical silicon is no longer the sole objective, nor even necessarily the decisive one: controlling model weights, training corpora, algorithmic standards, and the open-source distribution channels through which half the world now acquires AI capability is equally potent, and China’s proposed weight and data controls mark the first time any state has formally treated a downloadable file as a strategic export commodity on par with a lithography machine.[20][22][23] The model gate differs from the hardware gate in one crucial respect — copies, once released, cannot be recalled — which is why it is being fortified prospectively, against future frontier systems rather than the installed base.

Pillar 5: Mutual Interdependence as a Conflict Buffer.

Paradoxically, complete decoupling is impossible under this structure, and its impossibility is the system’s stabilizer. Because both nations rely on distinct assets controlled by the other — and because the measurable cost of rupture runs to several points of global GDP — the mutual power to disrupt operates as a deterrent against total economic warfare, producing the crisis-truce-crisis rhythm of 2025–2026 rather than a spiral to rupture.[31][30][33] The buffer is real but not permanent: it thins at exactly the rate that each side’s substitution programs succeed, which is why both capitals treat the current truce as borrowed time.

Pillar 6: The Open-Weight Paradox.

China’s model-gate leverage exists only because Chinese models are globally adopted, and they are globally adopted only because they are open; restricting them therefore consumes the very asset it protects. This is a genuinely novel strategic dilemma — no equivalent exists in hardware, where a withheld chip loses no value by being withheld — and it explains both the caution of the July 2026 consultations (hosted APIs preserved, controls aimed at future systems) and the industry pushback reported from within them.[20][22][24] The paradox cuts in Washington’s direction too: American pressure to ban Chinese models from Western use, if successful, would relieve Beijing of the dilemma by destroying the adoption anyway. The model gate, uniquely, punishes whichever gatekeeper closes it first.

Pillar 7: The Substitution Clock.

Every chokepoint is a depreciating asset, and both sides now govern by its depreciation schedule. American compute leverage decays as SMIC yields improve and CXMT’s HBM scales past the 300,000-package constraint; Chinese material leverage decays as allied processing capacity comes online; and the rational response to holding a depreciating weapon — using it while it retains value — is the single greatest source of instability in the system.[19][35][3] The statesmanship the era demands is therefore temporal: synchronizing the truces, the substitution programs, and the escalation thresholds so that neither side ever concludes that its window is closing faster than its rival’s. The Busan expiry, the 2027 readiness debates, and the HBM crossover point are all, in this sense, the same clock read from different walls.[33][10][19]

Conclusion:

The emergence of Chokepoint Reciprocity marks the definitive end of post-Cold War technological globalization — not its interruption, and not a detour from which the old single-architecture world will return, but its replacement by a permanent structural condition. The evidence assembled in this paper supports a conclusion that is uncomfortable for maximalists in both capitals. The United States, commanding the summit of the computing stack, discovered between April and October 2025 that the summit rests on a material floor it does not control, and that the world’s largest technology market can be closed against it faster than its licenses can open. China, commanding that floor and that market, has discovered that frontier computation remains gated by memory, tooling, and design ecosystems it cannot yet replicate, and that its newest source of leverage — the open models the world has adopted — is an asset that restriction devalues. Each power holds a knife to something the other cannot quickly replace; neither can use the knife without bleeding; and both are racing, on parallel substitution clocks, to become the first that can.

The U.S.–China technology competition is therefore no longer a race with a clear winner and loser. It is a managed rivalry between two gatekeeping systems, and it will be won — to the extent that word retains meaning — not by the side that decouples first but by the side that manages its dependencies, its allies, its truces, and its revisions with greater discipline over a longer horizon. Absolute technological sovereignty is an illusion for both Washington and Beijing; the pursuit of it to the limit would cost the world economy up to 7 percent of its output and the pursuing powers their own growth engines.[31][30] Security in the era of Chokepoint Reciprocity will be achieved through the sophisticated, calculated management of mutual vulnerabilities: deterrence consciously maintained, escalation deliberately paced, substitution honestly priced, and interdependence — the despised, indispensable interdependence — preserved in exactly the measure that keeps two fierce competitors condemned, in Allison’s phrase, to coexist rather than co-destruct.[32] The gates are built. The gatekeepers face each other. The task of statecraft, for the remainder of this decade, is to ensure that they keep facing each other across open gates rather than closed ones — because in a two-way conflict of chokepoints, the only decisive move available to either side is the one that ruins both.

Footnotes / Endnotes:

[1] Henry Farrell & Abraham L. Newman, “Weaponized Interdependence: How Global Economic Networks Shape State Coercion,” International Security, Vol. 44, No. 1 (Summer 2019), pp. 42–79, MIT Press. https://direct.mit.edu/isec/article/44/1/42/12237/Weaponized-Interdependence-How-Global-Economic

[2] GEOPOL, “Henry Farrell & Abraham Newman: The Theorists of Weaponized Interdependence,” May 2026. https://geopol.uk/thinkers/farrell-and-newman/

[3] Chris Miller (Tufts University), interviewed by Axios, “China threatens access to supplies vital to the most important U.S. industries,” Axios, October 9, 2025. https://www.axios.com/2025/10/09/trump-ai-trade-war-china-minerals

[4] TechCrunch, “Nvidia H20 chip exports hit with license requirement by US government,” April 15, 2025. https://www.techcrunch.com/2025/04/15/nvidia-h20-chip-exports-hit-with-license-requirement-by-us-government

[5] NVIDIA Corporation, “NVIDIA Announces Financial Results for First Quarter Fiscal 2026,” Press Release (SEC Exhibit 99.1), May 28, 2025. https://www.sec.gov/Archives/edgar/data/1045810/000104581025000115/q1fy26pr.htm

[6] Timothy J. Keeler, Tamer A. Soliman & Howard W. Waltzman (Mayer Brown), “PRC Announces New Export Controls on Rare Earth and Battery Materials and Technology,” October 13, 2025. https://www.mayerbrown.com/en/insights/publications/2025/10/prc-announces-new-export-controls-on-rare-earth-and-battery-materials-and-technology

[7] Freshfields, “China Intensifies Export Controls over Rare Earths and Related Technologies,” October 2025. https://www.freshfields.com/en/our-thinking/blogs/risk-and-compliance/china-intensifies-export-controls-over-rare-earths-and-related-technologies-102lpyn

[8] Center for Strategic and International Studies (CSIS), “Rare Earth Export Restrictions One Year Later,” May 2026. https://www.csis.org/analysis/rare-earth-export-restrictions-one-year-later

[9] Reuters, “China tightens rare earth export controls, targets defence, semiconductor users,” October 2025. https://www.aol.com/articles/china-tightens-rare-earth-export-032112473.html

[10] Introl, “BIS H200 Export Policy Shift & the AI OVERWATCH Act,” February 2026. https://introl.com/blog/bis-h200-china-export-policy-ai-overwatch-act-2026

[11] Center for a New American Security (CNAS), “CNAS Insights: Unpacking the H200 Export Policy,” January 16, 2026. https://www.cnas.org/publications/cnas-insights/cnas-insights-unpacking-the-h200-export-policy

[12] Kristina Roberts (Influencer Magazine UK), “Nvidia’s H200 AI Chip Shipments to China Begin Amidst Ongoing US Export Controls,” July 14, 2026 (reporting Under Secretary Jeffrey Kessler’s House Foreign Affairs Committee testimony). https://influencermagazine.uk/2026/07/nvidias-h200-ai-chip-shipments-to-china-begin-amidst-ongoing-us-export-controls/

[13] Technetbook, “Nvidia H200 AI Chip Shipments to China Expand as US Authorizes New Buyers and Eases Export Controls,” July 2026. https://www.technetbooks.com/2026/07/nvidia-h200-ai-chip-shipments-to-china.html

[14] Reuters, via Tom’s Hardware, “China bans foreign AI chips from state-funded data centers,” November 5, 2025. https://www.tomshardware.com/tech-industry/semiconductors/china-bans-foreign-ai-chips-from-state-funded-data-centers

[15] Foundation for Defense of Democracies (FDD), “Signaling Confidence in Its Domestic Industry, China Bans Foreign AI Chips in State-Funded Data Centers,” November 10, 2025. https://www.fdd.org/analysis/2025/11/10/signaling-confidence-in-its-domestic-industry-china-bans-foreign-ai-chips-in-state-funded-data-centers/

[16] Tech Times (reporting Bloomberg/NDRC plan), “China AI Data Center Grid Locks Out Nvidia With $295 Billion Domestic Chip Mandate,” June 22, 2026. https://www.techtimes.com/articles/318868/20260622/china-ai-data-center-grid-locks-out-nvidia-295-billion-domestic-chip-mandate.htm

[17] International Banker, “Why China Has Banned Domestic Firms from Buying Nvidia’s AI Chips,” November 12, 2025. https://internationalbanker.com/finance/why-china-has-banned-domestic-firms-from-buying-nvidias-ai-chips/

[18] RCR Wireless News (citing Bloomberg), “Huawei to double output of Ascend AI chips,” September 30, 2025. https://www.rcrwireless.com/20250930/ai-infrastructure/huawei-ai-chips-2

[19] SemiAnalysis, “Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM is The Bottleneck,” February 2026. https://newsletter.semianalysis.com/p/huawei-ascend-production-ramp

[20] Marcus Schuler (Implicator.ai), reporting the Financial Times, “China Weighs AI Model and Chip Design Export Controls,” July 21, 2026. https://www.implicator.ai/china-considers-adding-ai-model-weights-and-chip-designs-to-export-list/

[21] Tom’s Hardware, “China is considering export controls on AI technologies, including banning local companies from using TSMC, report claims,” July 21, 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/china-is-considering-export-controls-on-ai-technologies-including-banning-local-companies-from-using-tsmc-report-claims-restrictions-would-also-advanced-ai-models-training-data-and-overseas-acquisitions

[22] TechRepublic, “China Considers Export Controls on AI Models, Training Data and Chip Technology,” July 23, 2026. https://www.techrepublic.com/article/news-apac-china-ai-model-export-controls/

[23] Jon Markman (Forbes), “China’s DeepSeek V4 And Qwen Reshape The Open-Source AI Race,” April 28, 2026 (citing MIT Technology Review download-share analysis). https://www.forbes.com/sites/jonmarkman/2026/04/28/chinas-deepseek-v4-and-qwen-reshape-the-open-source-ai-race/

[24] Hugging Face, “The Future of the Global Open-Source AI Ecosystem: From DeepSeek to AI+,” February 3, 2026. https://huggingface.co/blog/huggingface/one-year-since-the-deepseek-moment-blog-3

[25] Kiplinger, “Nvidia Earnings: Updates and Commentary,” May 2026 (Q1 FY2027 earnings-call coverage). https://www.kiplinger.com/investing/live/nvidia-earnings-live-updates-and-commentary-may-2026

[26] CNBC, “Nvidia (NVDA) Q1 2027 earnings report: Live updates,” May 20, 2026. https://www.cnbc.com/2026/05/20/nvidia-nvda-earnings-report-q1-2027.html

[27] The Motley Fool, “Nvidia (NVDA) Q1 2027 Earnings Transcript,” May 20, 2026. https://www.fool.com/earnings/call-transcripts/2026/05/20/nvidia-nvda-q1-2027-earnings-transcript/

[28] Yahoo Finance, “TSMC posts record Q2 profit as AI demand drives higher revenue and investment outlook,” July 16, 2026. https://finance.yahoo.com/technology/ai/articles/tsmc-posts-record-q2-profit-104626669.html

[29] Investing.com, “TSMC Q2 2026 slides: AI demand drives record margins, HPC surges 20%,” July 16, 2026. https://www.investing.com/news/company-news/tsmc-q2-2026-slides-ai-demand-drives-record-margins-hpc-surges-20-93CH-4794789

[30] Stanford Institute for Economic Policy Research (SIEPR), “IMF’s Gita Gopinath: Geopolitics and its impact on global trade and the dollar” (remarks at Stanford University). https://siepr.stanford.edu/news/imfs-gita-gopinath-geopolitics-and-its-impact-global-trade-and-dollar

[31] Shekhar Aiyar, Andrea Presbitero, Michele Ruta et al. (International Monetary Fund), “Geoeconomic Fragmentation and the Future of Multilateralism,” IMF Staff Discussion Note SDN/2023/001, January 2023. https://www.imf.org/-/media/files/publications/sdn/2023/english/sdnea2023001.pdf

[32] Graham Allison (Harvard Kennedy School), via Harvard Griffin GSAS, “Discussing the Thucydides Trap,” Harvard Graduate School of Arts and Sciences, 2024. https://gsas.harvard.edu/news/discussing-thucydides-trap

[33] War on the Rocks, “The Burn and the Choke: Why Semiconductor Controls Will Outlast China’s Rare Earth Weapon,” January 5, 2026. https://warontherocks.com/the-burn-and-the-choke-why-semiconductor-controls-will-outlast-chinas-rare-earth-weapon/

[34] American Enterprise Institute (AEI), DataPoints (quoting Chris Miller in the Financial Times), “China’s Trade Tactics,” December 2025. https://www.aei.org/featured_data/chinas-trade-tactics/

[35] SemiconductorX, “Huawei / HiSilicon Spotlight — Ascend AI Chips, SMIC, China Bifurcation & Export Controls,” 2026. https://semiconductorx.com/spotlight-huawei-hisilicon.html

[36] Semiconductors Insight, “US China Chip Export Controls H200 2026: The Policy Shift Explained,” April 29, 2026. https://semiconductorsinsight.com/us-china-chip-export-controls-h200-2026/

[37] Jordan Schneider (ChinaTalk), with Chris Miller & Chris McGuire, “Emergency Pod: Rare Earth Export Controls,” October 10, 2025. https://www.chinatalk.media/p/emergency-pod-rare-earth-export-controls

[38] Capital.com, “Nvidia Stock Forecast: H200 China Chip Clearance,” May 18, 2026 (citing Reuters and The Business Times on approved buyers and 75,000-unit allocations). https://capital.com/en-int/market-updates/nvidia-stock-forecast-19-05-2026

[39] Yahoo Finance / Benzinga contributor, “Nvidia’s $5.4 Trillion Moment Hides A Much Bigger Story About China,” May 13, 2026. https://finance.yahoo.com/news/nvidias-5-4-trillion-moment-131120825.html

[40] Sourceability, “China tightens rare-earth export controls amid Nexperia disruption,” October 31, 2025. https://sourceability.com/post/china-tightens-rare-earths-amid-nexperia-disruption

[41] Tech Times, “TSMC Posts Record Quarter as AI Chip Demand Pushes Full-Year Growth Outlook Past 40%,” July 16, 2026. https://www.techtimes.com/articles/320696/20260716/tsmc-posts-record-quarter-ai-chip-demand-pushes-full-year-growth-outlook-past-40.htm

[42] Cryptopolitan (reporting the Financial Times), via MEXC News, “China government mandates domestic AI chips for state use, shunning Nvidia,” December 2025. https://www.mexc.com/news/252618

[43] CMS Law, “China tightens export controls on rare earths: takeaways for global businesses,” March 2026. https://cms.law/en/chn/legal-updates/china-tightens-export-controls-on-rare-earths-takeaways-for-global-businesses