Introduction: The Death of the Borderless Algorithm
On May 28, 2025, buried in the routine language of a quarterly earnings release, NVIDIA — then, as now, the most valuable company ever built on the sale of arithmetic — disclosed a single accounting item that captured the end of an era. The company recorded a $4.5 billion charge for excess inventory and purchase obligations on its H20 processor, a chip that had been deliberately engineered downward, hobbled by design, to squeeze beneath the performance thresholds of American export controls so that it could still legally be sold to China. In April 2025, the United States government had informed NVIDIA that even this diminished product would now require an export license, and demand evaporated overnight. The company reported that it had managed to sell $4.6 billion of H20 systems into China before the door closed.[1] Consider what that line item actually represents. A private corporation designed a product not for a customer, but for a regulation; a government redrew the regulation, and billions of dollars in silicon became, in a stroke, geopolitically stranded assets. No purely economic theory of technology can explain that accounting entry. Only a political one can.
It was not supposed to end this way. For roughly two decades, the governing mythology of the digital age held that information technology was inherently borderless — that code, data, and eventually machine intelligence would flow across jurisdictions as freely as radio waves, dissolving the parochial geographies of the nation-state. The early rhetoric of artificial intelligence inherited this utopianism wholesale. AI was to be a global public good, an open-source gift economy in which a graduate student in Lagos and a research lab in Palo Alto drew from the same wells of pre-trained models and shared benchmarks. Corporate mission statements promised to develop intelligence ‘for the benefit of all humanity.’ Academic papers circulated without borders; model weights were posted for anyone to download; the cloud, that great deterritorializing metaphor, floated serenely above the map.
That world is dead. In its place stands a landscape of chip blockades and licensing regimes, of closed-source hoarding and state-backed national champions, of sovereign compute enclaves ringed by regulatory fences, of model weights treated with the reverence and paranoia once reserved for nuclear centrifuge designs. The United States has constructed the most ambitious technology-denial regime since the Cold War’s CoCom system, and then — in a twist that only confirms the underlying logic — begun selling exceptions to it, monetizing the choke point itself through a 25 percent levy on advanced chips exported to China.[24] Europe has legislated a continental regulatory shield and simultaneously poured billions into ‘sovereign AI’ infrastructure. The Gulf monarchies have converted hydrocarbon rents into GPU clusters. China, walled off from the most advanced silicon, has responded with a state-orchestrated campaign of semiconductor self-sufficiency and a flood of open-weight models designed to set global standards from below. The chief executive of NVIDIA, the company whose products constitute the strategic commodity of the age, has traveled from capital to capital preaching a doctrine indistinguishable from mercantilism:
“Every country needs its own sovereign AI – to produce intelligence rather than import it.”
— Jensen Huang, CEO of NVIDIA, World Government Summit [5]
When the world’s leading vendor of a technology markets it explicitly in the vocabulary of national sovereignty — intelligence as a domestic product, imports as dependency — the borderless illusion is not merely fading. It has been formally repudiated by the very actors who once profited from proclaiming it.
Defining Algorithmic Realism in Geopolitics
This paper proposes a name for the intellectual framework adequate to this new landscape: algorithmic realism. The term requires careful definition, because it is being deliberately adapted from another discipline. In computer science, ‘algorithmic realism’ was coined by Ben Green and Salomé Viljoen in a landmark 2020 paper to describe a methodological evolution away from ‘algorithmic formalism’ — the internalist habit of evaluating algorithms purely on their mathematical properties, as if optimization objectives and benchmark scores exhausted their meaning. Green and Viljoen, drawing an explicit analogy to the twentieth-century movement from legal formalism to legal realism, argued that algorithms must instead be judged by their real-world social consequences: by what they do to actual people embedded in actual institutions, not by what they prove on paper.[2]
This paper performs the same operation one level up, at the scale of the international system. If algorithmic realism in computer science means looking past the mathematics of a model to see its social harms, algorithmic realism in international relations means looking past the mathematics to see algorithms as they actually function among states: as raw instruments of national power, economic coercion, and security competition. The formalist illusion being punctured is different but structurally identical. Where the computer scientist’s formalism imagined the algorithm as socially neutral, the geopolitical formalism — call it techno-formalism, or the Silicon Valley view — imagined the algorithm as territorially neutral: a placeless mathematical object indifferent to flags, borders, and armies. Both illusions flatter their holders, and both collapse on contact with reality. A frontier model is not an equation. It is a congealed stack of physical and political facts: lithography machines that exist in one Dutch town, advanced fabrication concentrated overwhelmingly in one Taiwanese company, data centers drawing gigawatts from national grids, training data scraped from particular languages and cultures, and export-licensing regimes that determine, chip by chip, who is permitted to run the mathematics at all.[6]
The central argument of this paper can be stated plainly. The rise of AI nationalism and model protectionism is not a policy accident, a temporary trade spat, or a betrayal of technology’s true nature. It is the inevitable materialization of algorithmic realism at global scale. Because artificial intelligence demonstrably shifts the balance of economic and military power, and because the international system remains anarchic — possessing no authority above states capable of enforcing restraint — states are structurally compelled to treat AI model weights, data pipelines, and compute infrastructure as sovereign national assets. The predictable result, already visible in the policy record from 2022 through mid-2026, is the permanent fracturing of the once-global digital order into heavily guarded technological spheres of influence.
Why This Paper, and This Framework, Are Named “Algorithmic Realism”
The choice of name is deliberate, and it carries four justifications that the reader should hold in mind throughout. First, the name performs an act of intellectual inheritance: it borrows the realist move from Green and Viljoen — strip away the formal abstraction, attend to consequences and power — and honors that lineage while extending it from the sociology of algorithms to the geopolitics of algorithms.[2] Second, the name anchors the analysis in the oldest and most durable tradition of international relations theory. ‘Realism’ in IR — from Thucydides through Morgenthau to Waltz and Mearsheimer — denotes the view that states pursue power and security under anarchy, that relative gains matter more than absolute ones, and that interdependence is always shadowed by vulnerability. Naming the framework ‘algorithmic realism‘ declares that AI has now entered the domain where those iron laws apply. Third, the name is diagnostic rather than celebratory: to call the present order ‘realist’ is to insist that we describe the world as it is — export controls, model hoarding, data fences — rather than as the borderless utopia its architects promised. Realism, in both of its parent traditions, is first of all an ethic of disillusionment. Fourth, and finally, the name is predictive. A realist framework generates falsifiable expectations: that global AI treaties will fail, that export controls will be evaded and then tightened and then traded away and then reimposed, that every state with the fiscal capacity will pursue sovereign compute, and that the weights of frontier models will be guarded like state secrets. As this paper documents, every one of those predictions has already been vindicated by events between 2020 and 2026. A framework that predicts is a framework worth naming.
The paper proceeds in nine sections. Section 1 constructs the theoretical foundation, contrasting algorithmic realism with techno-formalism and bridging technology policy with structural realism. Section 2 anatomizes the domestic pillars of AI nationalism: state-backed champions, compute sovereignty, and regulatory shields. Section 3 dissects the international mechanics of model protectionism: the open-weights calculus, the Silicon Curtain of export controls, and the sovereign fencing of data. Section 4 presents the defining case study of the era, the US–China chip war, tracked through the primary record of regulations and corporate earnings up to the second quarter of 2026. Section 5 surveys the sovereign AI movements of Europe and the Middle East. Section 6 assesses the systemic consequences: algorithmic bipolarity, epistemicide, and cultural imperialism. Section 7 draws the implications for global governance. Section 8 formalizes the paradigm shift in tabular form, and Section 9 distills the argument into five pillars of algorithmic realism before the conclusion issues its final warning.
Section 1: Theoretical Foundations — Algorithmic Realism versus Techno-Formalism
Every geopolitical era rests on an implicit theory of its most important technology, and every such theory eventually meets the world. This section builds the conceptual scaffolding for the rest of the paper. It begins by deconstructing the reigning corporate-academic narrative of AI as neutral mathematics — techno-formalism — then bridges technology policy with classical International Relations realism, and closes with the materialist turn: the insistence that analysis must descend from the ‘cloud’ metaphor to the physical, finite, and geographically bound inputs of machine intelligence. The section’s burden is to show that algorithmic realism is not a mood or a metaphor but a coherent theoretical position with identifiable intellectual parents and testable implications.
1.1 The Critique of Techno-Formalism
Techno-formalism is the name this paper gives to the constellation of beliefs that dominated elite discourse about AI from roughly 2012 to 2022, and that still structures much of Silicon Valley’s public rhetoric. Its core tenets are familiar. Artificial intelligence models are ‘just math’ — matrices of floating-point numbers whose meaning is exhausted by their mathematical description. Scientific progress in AI is objective and universal, belonging to no nation because it belongs to the discipline. Openness is both an engine of progress and a moral default; restrictions on the flow of models, papers, and talent are frictions to be minimized. And the appropriate governance posture is light-touch, technocratic, and global, since a borderless technology logically demands borderless rules. On this view, an export control on a GPU is as conceptually confused as an export control on the Pythagorean theorem.
The formalist narrative was never entirely cynical; it reflected the genuine sociology of a research field built on open publication, shared benchmarks, and internationally mobile talent. But it systematically obscured three facts. First, it concealed the interests of its proponents: ‘borderless AI’ was also a business model, in which a handful of American platforms aggregated the world’s data, talent, and customers while framing any national resistance as backwardness. Critical scholars at the AI Now Institute have documented how the vocabulary of open, global innovation coexisted comfortably with an unprecedented concentration of compute, capital, and decision-making power in a few corporate hands.[45] Second, formalism obscured the dual-use character of the technology. The same transformer architecture that autocompletes an email can guide a loitering munition, triage signals intelligence, or design a persuasion campaign; a technology that touches every sector of the economy and every domain of warfare cannot remain politically weightless, whatever its mathematics look like. Third, and most fundamentally, formalism mistook a historical interlude for a natural law. The borderless internet of 1995–2015 was not the spontaneous order of information; it was an artifact of unipolar American power, a period in which one state was so dominant that it could afford to experience openness as costless. As that dominance eroded, the openness went with it.
The intellectual resources for this critique were assembled inside computer science itself before they were needed by geopolitics. Green and Viljoen’s diagnosis of ‘algorithmic formalism’ identified precisely the epistemic habits — internalist reasoning, faith in neutrality, indifference to downstream consequence — that this paper finds reproduced at the level of national strategy. Their prescription, a ‘porous’ mode of analysis attentive to the social world beyond the model’s formal boundary, is the methodological seed of the present framework.[2] What they did for the algorithm embedded in a welfare office or a courtroom, this paper does for the algorithm embedded in the international system: it refuses to let the mathematics launder the politics.
1.2 The Intersection of Software and State Realism
If techno-formalism supplies the thesis to be negated, classical International Relations realism supplies the engine of the negation. Structural realism, in the tradition of Kenneth Waltz, begins from a single austere premise: the international system is anarchic, meaning that no authority above the state can guarantee its survival, and therefore states must ultimately rely on self-help. From this premise flows the security dilemma — one state’s defensive accumulation of capability reads as threat to its neighbors — and the primacy of relative gains: what matters is not whether cooperation makes both parties richer, but whether it makes my rival relatively stronger. John Mearsheimer’s offensive realism sharpens the point: great powers do not merely seek security, they seek regional hegemony and act to prevent the rise of peer competitors, because in anarchy the only reliable insurance is preponderance.
Insert artificial intelligence into this machinery and the outputs are overdetermined. AI is what economists call a general-purpose technology and what strategists call an omni-use one: it simultaneously raises total factor productivity, transforms intelligence collection and analysis, accelerates weapons design, enables autonomous systems, and reshapes the information environment in which political legitimacy itself is contested. A technology with that profile alters the balance of power along every axis at once. Under anarchy, no great power can responsibly allow a rival to command a decisive lead in such a technology, whatever the absolute gains from cooperation might be — and every great power knows that every other great power reasons the same way. The result is a structurally forced competition for algorithmic dominance, in which restraint is punished and hedging is rational. The remarkable fact is how early and precisely this dynamic was predicted. In June 2018, the British investor Ian Hogarth — who would later chair the United Kingdom’s frontier AI safety taskforce — published an essay titled ‘AI Nationalism’ forecasting that machine learning’s transformation of economies and militaries would destabilize the international order and force governments to act, with closed economies, blocked acquisitions, and restricted talent flows to follow.[3]
“AI policy will become the single most important area of government policy.”
— Ian Hogarth, “AI Nationalism” (2018) [3]
Eight years later, Hogarth’s forecast reads less like prediction than like minutes taken in advance. Susan Ariel Aaronson’s 2024 study for the Centre for International Governance Innovation documented the intervening consolidation: a growing roster of states adopting what she characterizes as a neo-mercantilist approach to nurturing AI, willing to impede foreign competitors and subsidize domestic ones in the name of sovereignty, security, and competitiveness.[4] The theoretical point deserves emphasis, because it is what distinguishes algorithmic realism from mere commentary on a trade war. Realism explains why the pattern is general rather than idiosyncratic. The United States did not restrict chip exports because of one administration’s temperament, and China did not pursue semiconductor self-sufficiency because of one leader’s ideology; both behaviors are what structural theory predicts of great powers confronting a power-shifting technology under anarchy. The same logic explains why middle powers — France, the United Arab Emirates, Saudi Arabia, India, South Korea — have converged on ‘sovereign AI’ strategies despite wildly different political systems: in a fragmenting order, dependency on any single foreign provider of intelligence infrastructure is a strategic vulnerability, and states purchase insurance against it when they can afford to.
1.3 The Materialist Turn in AI Studies
The third foundation of algorithmic realism is methodological: a materialist turn that refuses the ‘cloud’ metaphor and insists on analyzing the physical, finite, and geographically bound inputs of artificial intelligence. The cloud is perhaps the most successful ideological metaphor of the digital age, conjuring an ethereal, placeless realm of computation. The reality is brutally terrestrial. Machine intelligence is manufactured from three material feedstocks — advanced silicon, energy, and data — and each is unevenly distributed across the map in ways that create precisely the choke points, dependencies, and rents that realist theory expects states to fight over.
Consider silicon first. The 2026 Stanford AI Index — the most comprehensive annual measurement of the field — notes that a single company, TSMC, fabricates almost every leading AI chip, rendering the entire global AI hardware supply chain dependent on one foundry complex on the island of Taiwan, the most militarily contested piece of geography on Earth.[6] Behind TSMC stands an even narrower choke point: the extreme-ultraviolet lithography machines required to pattern leading-edge chips are produced by exactly one firm, ASML of the Netherlands, whose export licenses have accordingly become instruments of alliance diplomacy. Chris Miller’s influential history ‘Chip War’ traced how this concentration emerged and why it converts semiconductor supply chains into strategic terrain; the Tufts University historian has spent the years since the book’s publication as perhaps the most-cited public interpreter of the export-control struggle.[9] Energy is no less material. The AI Index measured global AI data-center power capacity at 29.6 gigawatts — roughly the peak demand of New York State — and found the United States hosting 5,427 data centers, more than ten times any other country; a single frontier training run now carries the carbon signature of a small city’s annual driving.[6] Electric grids, unlike ideas, do not cross borders freely; they are licensed, subsidized, and defended by states. And data, the third feedstock, is the congealed cultural and behavioral record of particular populations — which is why, as Section 3 will show, states increasingly treat it as a sovereign natural resource, subject to localization mandates and extraction controls, rather than as a free-flowing global commodity.
Once the analysis descends to this material level, the formalist picture becomes untenable. An ‘algorithm’ that requires a Taiwanese fab, a Dutch lithography machine, an American design house, a Gulf sovereign wealth fund’s capital, a national electric grid’s gigawatts, and a population’s linguistic corpus is not a placeless mathematical object. It is a geography. And geography, as every realist from Mackinder onward has understood, is what states contest. Algorithmic realism, then, rests on three legs: the epistemic critique of formalism inherited from critical computer science, the structural logic of competition inherited from IR realism, and the materialist attention to physical inputs inherited from political economy. With this scaffolding in place, the paper turns from theory to the observable architecture of the new order.
Section 2: The Pillars of AI Nationalism
If Section 1 explained why states must compete for algorithmic power, this section examines how they do it at home. AI nationalism is not a single policy but an architecture — a mutually reinforcing set of domestic interventions through which states attempt to build self-reliant AI ecosystems and weaponize their internal markets, budgets, and legal systems in the service of strategic autonomy. Three pillars carry the structure: the direct cultivation of state-backed foundational-model champions; the enclosure of compute and energy as sovereign infrastructure; and the deployment of regulation as a geopolitical shield. Each pillar would be legible on its own as ordinary industrial or consumer-protection policy. Read together, and read against the theoretical backdrop of Section 1, they reveal a coherent strategic project: the conversion of the national territory into a defensible algorithmic base.
2.1 State-Backed Foundational Models: The Return of the National Champion
The national champion — the flagship firm nurtured by the state to carry the flag in a strategic industry — was supposed to have died with twentieth-century dirigisme. Artificial intelligence has resurrected it in less than five years. The pattern is now global and remarkably uniform. In France, Mistral AI has been elevated to the status of continental standard-bearer: President Emmanuel Macron shared a stage with Mistral’s chief executive Arthur Mensch and NVIDIA’s Jensen Huang at the VivaTech conference in June 2025 to bless the launch of Mistral Compute, a sovereign European AI cloud built in its first phase on 18,000 NVIDIA Grace Blackwell systems, an initiative the French president hailed as historic.[14] Mensch, for his part, had already told the Davos audience in January 2025 that the company was
“not for sale”
— Arthur Mensch, CEO of Mistral AI, World Economic Forum, Davos [14]
— a three-word declaration that would be commercially unremarkable were it not so obviously addressed to geopolitics rather than to shareholders: a promise to the French state that its champion would not be absorbed, as DeepMind had been, into an American platform. In the United Arab Emirates, the state-linked Technology Innovation Institute has built the Falcon family of open-weight models into the default foundation for Arabic-language AI across the region, while the national champion G42 anchors a compute empire spanning the Microsoft-backed Stargate UAE data-center program and a Cerebras-based supercomputing fleet.[13] Saudi Arabia answered in May 2025 by founding HUMAIN as a wholly owned subsidiary of the Public Investment Fund — a national AI champion conjured directly from the sovereign wealth of the state.[13] China’s champions — Baidu, Alibaba, Tencent, and the newer model labs DeepSeek, Zhipu, Moonshot, and MiniMax — operate within an explicit national strategy that has targeted global AI leadership by 2030 since 2017, financed by government guidance funds that one Stanford-cited estimate places at $912 billion deployed across strategic industries over two decades.[6]
The strategic logic of the champion model is straightforward once viewed through a realist lens. A domestic frontier lab is simultaneously an economic asset (capturing rents from the general-purpose technology of the era), a security asset (guaranteeing that intelligence infrastructure critical to government, defense, and finance does not run on a rival’s servers under a rival’s kill switch), and a diplomatic asset (something to offer partners in the emerging economy of compute treaties and model-sharing agreements). It is also, crucially, an insurance policy against coercion: the whole point of strategic autonomy is that a foreign government’s export ban, sanctions package, or terms-of-service change cannot decapitate one’s own AI capacity. The events of 2025–2026 repeatedly validated the insurance logic. When American export policy could strand $4.5 billion of NVIDIA inventory in a quarter, and when access to frontier American models could be modulated by presidential directive, every capital in the world drew the same conclusion: dependency on another state’s algorithmic stack is a standing vulnerability.[1]
2.2 Compute Sovereignty and the Enclosure of the Grid
The second pillar treats access to computation itself — advanced GPUs, the data centers that house them, and the electricity that feeds them — as a national-security priority on par with energy security or food security. The vocabulary is telling: governments now speak routinely of ‘sovereign compute,’ ‘AI factories,’ and ‘national AI infrastructure,’ phrases that would have been unintelligible to policymakers a decade ago. The substance matches the vocabulary. The United Kingdom’s Isambard-AI cluster in Bristol, fully operational since mid-2025 with 5,448 NVIDIA GH200 superchips, is explicitly framed as a dedicated national AI factory within a broader £18 billion infrastructure program designed to let British startups and government agencies build models under British jurisdiction.[15] France’s Jean Zay supercomputer was extended in 2025 specifically for ‘sovereign training.’ NVIDIA announced in 2025 a wave of European partnerships — more than twenty AI factories across the continent — with its chief executive framing the buildout in unmistakably infrastructural terms:
“AI is the essential infrastructure of our time.”
— Jensen Huang, CEO of NVIDIA, announcing Europe’s AI infrastructure buildout [12]
Infrastructure, of course, is precisely the category of asset that states have never allowed to remain stateless: roads, rails, grids, ports, and telecommunications networks are universally licensed, regulated, subsidized, and secured by governments. To declare AI ‘infrastructure’ is to invite — indeed to demand — its nationalization in the regulatory sense. The declaration also has a hard physical corollary. Because training and inference at frontier scale are energy-gluttonous, compute sovereignty entails grid sovereignty. States are now securing dedicated generation for AI campuses, fast-tracking permits for data centers as a matter of national urgency (the American executive order of July 2025 on accelerating federal permitting of data-center infrastructure is the paradigm case), and treating electricity for AI as a strategic reserve.[11] The Stanford AI Index’s finding that state-backed investment in AI supercomputing is rising in parallel with national AI strategies — particularly among developing economies — confirms that the enclosure of compute is not a great-power idiosyncrasy but a general movement: a worldwide fencing of the computational commons into national paddocks.[6]
2.3 Regulatory Nationalism as a Geopolitical Shield
The third pillar is legal. Regulation is conventionally analyzed in the register of consumer protection, safety, and rights — and the great AI statutes of the 2020s genuinely serve those purposes. But an algorithmic-realist reading attends to their second function: regulation as a protectionist rampart, a means of insulating domestic markets and values from foreign — primarily American — algorithmic dominance, and of projecting regulatory power outward. The European Union’s AI Act, in force since August 2024 as the world’s first comprehensive horizontal AI regulation, is the exemplary case. Its risk-based architecture imposes conformity assessments, documentation, and oversight obligations that fall most heavily on the foreign hyperscalers whose general-purpose models dominate the European market, while its extraterritorial reach exports European standards to any provider who wishes to touch the single market — the celebrated ‘Brussels effect’ repurposed for the algorithmic age. European officials have been candid about the sovereignty dimension: the Council’s own May 2026 communiqué on amending the Act framed the reform explicitly as strengthening the Union’s digital sovereignty and competitiveness.[17]
Yet the trajectory of the AI Act between 2024 and 2026 also exposes the tensions inside regulatory nationalism, and the paper would be incomplete without recording them. By late 2025 the implementation of the Act’s high-risk regime was visibly off track, and on 19 November 2025 the European Commission tabled a ‘Digital Omnibus on AI’ proposing to defer the core high-risk obligations. After failed trilogues in April 2026, the co-legislators reached provisional agreement on 7 May 2026: obligations for stand-alone high-risk systems were postponed to 2 December 2027, and for AI embedded in regulated products to 2 August 2028, while transparency duties largely held to their original August 2026 date and a new prohibition on AI-generated non-consensual intimate imagery was added.[16] Civil-society critics read the retreat as capitulation to industry and American pressure — one analysis noted that 69 percent of the Commission’s 2025 lobbying meetings on the file were with business groups — while the Commission presented it as pragmatic sequencing.[18] For the purposes of this paper, the episode cuts both ways and both cuts are instructive. On one hand, it demonstrates that regulatory shields are costly to hold when a jurisdiction lacks the underlying industrial base: Europe discovered that it could not simultaneously constrain the technology and covet it. On the other hand, the direction of the adjustment — simplification in the name of ‘digital sovereignty,’ explicit favoritism toward European small mid-caps, acceleration of the sovereign-compute agenda — confirms rather than refutes the nationalist reading. Europe did not abandon the shield; it reforged the shield into something closer to a sword, aligning its regulatory calendar with the needs of its own champions. Regulatory nationalism, like every other pillar in this section, bends toward the same end: a domestic ecosystem robust enough to survive in a fragmented world. Beyond Europe, the same instrument appears in cruder forms — data-localization statutes, security reviews of foreign algorithms, national cybersecurity laws conditioning market access on source-code disclosure — which Section 3 takes up as part of the international mechanics of protectionism.
Section 3: The Mechanics of Model Protectionism
Where Section 2 examined the domestic construction of AI power, this section examines its international defense. Model protectionism is the outward-facing complement of AI nationalism: the system of barriers, both offensive and defensive, that states erect around algorithmic assets at the border. Its mechanics operate on the three material layers identified in Section 1. At the layer of models, states and firms calculate over the release or hoarding of weights. At the layer of hardware, export controls draw a Silicon Curtain across the map. At the layer of data, localization laws and scraping restrictions fence the informational commons into sovereign paddocks. Each mechanism deserves extended treatment, because each reveals a different face of the same underlying logic: in an anarchic system, the flow of power-relevant technology is never neutral, and every channel of flow becomes a lever of statecraft.
3.1 The Weaponization of Openness: Open Weights versus Closed Weights
No question in contemporary AI policy is more strategically charged than whether the numerical weights of a trained model should be published. The techno-formalist framing treats this as an engineering or business choice; the realist framing recognizes it as a decision about the international distribution of capability. Two opposing calculi are in play, and the fascinating fact of the mid-2020s is that the two superpowers have largely swapped the positions one might have predicted for them. The closed-weights calculus, dominant among the leading American frontier labs, treats weights as crown jewels: publishing them surrenders a hard-won capability lead to every rival and adversary simultaneously, forecloses the ability to monitor or revoke misuse, and — in the language of nonproliferation that increasingly saturates the discourse — constitutes an irreversible release of a dual-use asset. Model hoarding, on this view, is simply the preservation of a technological monopoly, and the collapse of transparency documented by Stanford’s Foundation Model Transparency Index — whose average score fell from 58 to 40 in a single year, with frontier labs disclosing progressively less about training data, compute, and architecture — is the visible signature of weights becoming strategic secrets. The erosion extends even into the nominally open ecosystem: by 2025, for the first time, downloads of opaque open-weight models — weights published, but training data, licensing, and provenance undisclosed — outnumbered downloads of models meeting basic open-source criteria, a collapse of transparency that coincides exactly with the rising geopolitical stakes of the open commons.[6][22]
The open-weights calculus inverts every term. Publishing capable weights commoditizes the layer where a rival holds pricing power, seeds a global developer ecosystem on one’s own architectural standards, and converts soft-power goodwill into technical lock-in. This is the strategy China has executed with startling success since January 2025, when the little-known lab DeepSeek released its R1 reasoning model as open weights under a permissive MIT license — an event that briefly erased roughly a trillion dollars of American market capitalization and has entered the lexicon as the ‘DeepSeek moment.’[19] The aftershocks reorganized the global open ecosystem. By 2026, Alibaba’s Qwen family had overtaken Meta’s Llama in cumulative downloads on Hugging Face, Chinese open models had surpassed American ones in total downloads according to MIT research, and the majority of trending open models were either Chinese or derivatives of Chinese bases — with the top positions on the leading open-model leaderboards occupied entirely by Chinese systems, whose combination of near-frontier performance, permissive licensing, and prices a fraction of Western equivalents has pulled global developer mindshare decisively eastward.[19][20] Industry observers stopped hedging:
“DeepSeek and its Chinese peers are no longer the underdogs.”
— Ben Lorica, editor of Gradient Flow, on China’s open-weight ascendancy [21]
Washington understood the stakes and answered in kind. The White House’s July 2025 AI Action Plan broke with the instinct to restrict and instead endorsed American open models as instruments of national power, observing that
“Open-source and open-weight models could become global standards in some areas.”
— The White House, “Winning the Race: America’s AI Action Plan” (July 2025) [11]
— and drawing the explicit conclusion that such standards carry geostrategic value, which the United States must contest by ensuring the world builds on American, rather than Chinese, foundations.[46] The symmetry is the point. Both hoarding and releasing are now moves in the same great game; ‘openness’ has been fully weaponized, valued not as a scientific norm but as a distribution channel for influence. A model given away freely is not a gift; it is an export of standards, values, dependencies, and telemetry. Meanwhile the same openness is read by security establishments as a proliferation leak — every capable open model is instantly available to every intelligence service, cartel, and militia on Earth — which is why the open-weights question now sits, unresolved and perhaps unresolvable, at the intersection of industrial strategy and arms control. The techno-formalist sees a licensing choice. The algorithmic realist sees the twenty-first century’s version of the fight over whether to internationalize atomic energy — replayed at the speed of software release cycles.
3.2 Export Controls and the Silicon Curtain
If model weights are the soft frontier of protectionism, hardware is its hard frontier — and it is here, in the machinery of export controls, that algorithmic realism achieves its purest physical expression. The strategy is one of choke-point control. Because the semiconductor supply chain narrows at several points to a handful of firms in a handful of allied jurisdictions — ASML’s monopoly on extreme-ultraviolet lithography in the Netherlands, TSMC’s dominance of leading-edge fabrication in Taiwan, NVIDIA’s dominance of AI accelerators and the CUDA software ecosystem in the United States, and American and Japanese control of critical tooling and materials — a coalition that governs those choke points can, in principle, deny an adversary the physical capacity to run frontier-scale mathematics. Since October 2022, when the United States imposed sweeping restrictions on the export of advanced chips and chipmaking equipment to China, this logic has been elaborated into the most ambitious technology-denial architecture since the Cold War: performance thresholds ratcheted downward, the Entity List expanded by dozens of Chinese firms, allied controls on lithography synchronized with Washington’s, and enforcement extended in June 2026 even to the overseas subsidiaries of Chinese-headquartered companies.[28] The scholar most identified with the historical analysis of this strategy has framed its dilemma with characteristic directness:
“We have to be very careful when deciding which countries and which companies” receive advanced AI chips.
— Chris Miller, Tufts University, author of “Chip War,” on CNBC [9]
The deeper significance of the Silicon Curtain is conceptual: it is the ultimate manifestation of the materialist turn described in Section 1.3. Export controls do not regulate ideas, papers, or even code — all of which travel too easily to police. They regulate objects: wafers, lithography machines, accelerator boards, the physical substrate without which the mathematics cannot be executed at scale. In doing so they concede, at the level of official policy, everything the formalists denied. If AI were really borderless math, blockading GPUs would be as futile as blockading geometry. The fact that the world’s most powerful state has staked its technological primacy on preventing adversaries from acquiring specific rectangles of silicon is the strongest possible admission that intelligence is a manufactured, material, and therefore controllable good. Section 4 will examine the empirical career of this strategy — its achievements, evasions, reversals, and monetization — in full detail; here it suffices to register its systemic effect. The Silicon Curtain has partitioned the world into compute zones: a trusted inner circle with unrestricted access to frontier hardware, a negotiated middle subject to licensing and caps, and a denied periphery forced toward smuggling, domestic substitution, or algorithmic efficiency. Every state on Earth now knows which zone it inhabits, and plans accordingly.
3.3 Data Colonialism and Sovereign Fences
The third mechanism of protectionism operates on the rawest material of all: data. Here the analytical vocabulary comes not from arms control but from the critique of empire. A substantial scholarly literature — anchored by Nick Couldry and Ulises Mejías’s theory of data colonialism and extended across African, Asian, and Latin American cases — argues that the global data economy reproduces the extractive geometry of historical colonialism: raw material (behavioral, linguistic, cultural data) is harvested from the periphery at negligible cost, processed in the metropole into high-value products (foundation models), and sold back to the periphery as a service, with the value captured almost entirely at the center.[37] A 2026 viewpoint in Development in Practice, surveying evidence from three continents, concludes that digital platforms systematically extract value from the Global South through mechanisms mirroring historical colonial patterns, and that conventional regulatory responses treat symptoms rather than the underlying asymmetry of power.[37]
States have responded by fencing the resource. Data-localization laws — requiring that citizens’ data be stored and processed on national territory — have proliferated from China’s Cybersecurity Law through India’s data-protection regime to dozens of jurisdictions; the OECD counted the trend accelerating years ago, and the AI boom has supercharged it, because training data is now understood as a strategic input rather than a compliance afterthought. Web-scraping, the extractive practice on which the first generation of foundation models was built, is being restricted by copyright litigation, robots-exclusion standards with legal teeth, national scraping rules, and — most significantly for this paper’s argument — an emergent norm that a nation’s linguistic and cultural corpus is a sovereign natural resource. The sovereign-AI programs surveyed in Section 5 uniformly cite data sovereignty as a core rationale: language, cultural context, and regulatory compliance argue for nationally controlled training data and inference, in the standard formulation.[13] The realist translation is exact. Just as twentieth-century resource nationalism converted oil from a concession-company commodity into patrimony — nationalized, quota-managed, and cartelized — twenty-first-century data nationalism is converting the cultural record of populations into state-managed patrimony. The fence around the data completes the protectionist triangle: weights guarded at the top of the stack, silicon interdicted at the bottom, and the human record enclosed at the base. Nothing about the algorithm, from its physical substrate to its training corpus to its published artifact, now moves across borders as of right. Everything moves, when it moves at all, as a matter of state permission.
Section 4: Case Study — The US–China Chip War, 2022–2026
Every theoretical framework earns its keep in the archive, and for algorithmic realism the decisive archive is the four-year confrontation between the United States and China over advanced semiconductors. No other episode so completely displays the framework’s elements — choke-point coercion, relative-gains reasoning, forced indigenization, the collision of corporate and national interest, and the conversion of a technology market into an instrument of grand strategy. This section reconstructs the confrontation chronologically through the primary record: the regulations of the Bureau of Industry and Security, the earnings disclosures of the companies caught in the crossfire, and the running commentary of the scholars and officials who shaped the debate. The story that emerges is not one of a policy succeeding or failing cleanly, but of a system — the anarchic competition described in Section 1 — metabolizing every move into a countermove, exactly as realist theory predicts.
4.1 The Architecture of Denial, 2022–2024
The opening act was the rule of 7 October 2022, in which the United States restricted the export to China of advanced logic chips, the equipment to make them, and even the American persons who might service them — a scope without precedent in the modern export-control system. The stated theory, elaborated in Congressional Research Service analysis, was twofold: to deny the People’s Liberation Army the computing substrate of military AI, and to arrest China’s construction of an indigenous, self-sufficient semiconductor ecosystem before it reached escape velocity.[8] The controls were tightened in October 2023 to close the loophole through which NVIDIA had sold China-specific variants, and coordinated with the Netherlands and Japan to cover lithography and tooling. NVIDIA responded each time as a rational firm: it engineered new chips — the A800, the H800, finally the H20 — each calibrated to sit precisely beneath the latest performance threshold, a product line whose specifications were authored as much in Washington as in Santa Clara. The Biden administration’s parting gesture, the January 2025 ‘AI Diffusion Framework,’ attempted to globalize the architecture into a three-tier licensing system covering nearly every country on Earth — a centrally planned geography of compute that provoked fierce industry resistance and was rescinded by the incoming Trump administration in May 2025 before taking effect.[8]
4.2 Escalation, Reversal, and the Monetized Choke Point, 2025–2026
The years 2025 and 2026 then delivered a sequence of reversals that no formalist account can explain but that realism renders legible as bargaining over a choke point. In April 2025, the government required licenses for even the deliberately degraded H20, stranding NVIDIA’s China inventory and producing the $4.5 billion write-down with which this paper opened; the company sold $4.6 billion of H20 systems in the quarter before the door closed, and subsequently reported zero H20 sales to China-based customers in the following quarter.[1] Then, in December 2025, the direction abruptly inverted: the President announced that NVIDIA’s far more capable H200 would be permitted into China, and on 13 January 2026 the Bureau of Industry and Security codified the shift, moving H200-class and AMD MI325X-class chips from a presumption of denial to case-by-case review under an elaborate scaffolding of end-user certifications and security conditions.[24] The following day the White House imposed a 25 percent tariff on the covered chips — the fiscal expression of the President’s declaration that a quarter of China H200 revenue would be paid to the United States. The choke point, in other words, was no longer merely a valve; it had become a tollbooth.[24]
The reversal split the American strategic community in ways that illuminate the politics of algorithmic realism. The Council on Foreign Relations’ analysis pronounced the new framework strategically incoherent, calculating that shipments of a million H200s would increase China’s installed AI compute in 2026 by 250 percent relative to reliance on domestic chips alone.[25] At the House Foreign Affairs Committee’s pointedly titled hearing, ‘Winning the AI Race Against the Chinese Communist Party,’ the former deputy national security advisor Matt Pottinger demanded legislative correction:
“Congress needs to step in, reverse the policy, and put durable guardrails in place.”
— Matt Pottinger, former U.S. Deputy National Security Advisor, House testimony [26]
The congressional response was strikingly bipartisan. Chris Miller’s own contemporaneous analysis observed that the Senate’s chip-restriction provision was introduced by a legislator close to the President and that, for the first time, parts of Silicon Valley itself — the AI labs and cloud providers whose competitive position is strengthened by weaker Chinese rivals — had come out in favor of restrictions, with the provision’s sponsor framing it in the administration’s own idiom:
“The Gain AI Act is simply an America First amendment.”
— Senator Jim Banks (R-IN), quoted in Chris Miller’s analysis of the chip-export debate [27]
By early 2026 the House Foreign Affairs Committee had advanced the AI OVERWATCH Act to give Congress veto power over chip-export licenses, and the Commerce Department — squeezed between a White House pursuing trade détente ahead of a presidential visit to Beijing and a Congress demanding vigilance — retreated into quiet enforcement of existing rules, suspending its 50 percent Affiliates Rule for a year while issuing the June 2026 clarification that licensing requirements reach Chinese-headquartered firms wherever their subsidiaries sit.[29] The oscillation is the finding. A choke point valuable enough to fight over is valuable enough to sell access to; export controls had become not merely a security instrument but a currency of great-power bargaining, traded against rare-earth access, tariff schedules, and summit atmospherics.[28]
4.3 The Corporate Ledger: Reading Geopolitics in Earnings Reports
The financial record of the period reads like a seismograph of the confrontation, and it is worth dwelling on because corporate earnings are where algorithmic realism becomes auditable. NVIDIA’s fiscal 2026 (ending January 2026) produced revenue of $215.9 billion, up 65 percent year over year, with fourth-quarter data-center revenue alone reaching $62.3 billion — figures without precedent in the history of the semiconductor industry, achieved substantially without the Chinese market that had once supplied a fifth of the company’s data-center demand.[33] The chief executive’s third-quarter formulation entered the vernacular of the boom:
“Blackwell sales are off the charts, and cloud GPUs are sold out.”
— Jensen Huang, CEO of NVIDIA, Q3 FY2026 earnings release (November 2025) [34]
The first quarter of fiscal 2027, reported in May 2026, sharpened the picture further: revenue of $81.6 billion, up 85 percent year over year; data-center revenue of $75.2 billion, up 92 percent; and — the geopolitically decisive line — no data-center Hopper shipments to China at all, against $4.6 billion in the year-ago quarter, with the company’s $91.0 billion guidance for the following quarter explicitly assuming zero data-center compute revenue from China.[32] The world’s most important AI company now plans its future on the assumption that the world’s second-largest AI market does not exist for it — even as the January 2026 rule notionally reopened that market. Hyperscaler capital expenditure projections exceeding a trillion dollars for 2027 filled the gap and then some, but the structural lesson stood: the bifurcation of the compute world had been priced in.[35]
4.4 The Verdict of Adaptation: China’s Countermobilization
The final panel of the case study belongs to the target. Export controls were premised on the assumption that denial would durably retard Chinese capability; the record of 2025–2026 shows instead a textbook realist response of forced indigenization and algorithmic adaptation. On the hardware side, SMIC pushed 7-nanometer-class and then 5-nanometer-class production using multi-patterned deep-ultraviolet lithography — costlier and lower-yield than the EUV route it was denied, but functional — expanding advanced-node capacity from roughly 45,000 toward 60,000 wafer starts per month through 2026, while Hua Hong prepared a second 7-nanometer line and Beijing’s 50 percent local-equipment mandate and the semiconductor priorities of the 15th Five-Year Plan (2026–2030) institutionalized the drive.[30] Huawei’s Ascend accelerators, targeted at 1.6 million dies across the product line in 2026 and paired with an open-source software stack positioned against CUDA, emerged as the credible domestic alternative — constrained chiefly by China’s lag in high-bandwidth memory, a gap its champions are working methodically to close.[30] On the model side, the DeepSeek phenomenon demonstrated that constraint itself could be a catalyst: denied maximal compute, Chinese labs pioneered efficiency techniques — mixture-of-experts sparsity, latent attention, aggressive quantization — that compressed frontier-adjacent capability into export-compliant hardware. Palantir’s chief executive drew the American lesson in the starkest competitive terms:
“We have to run harder, run faster, have an all-country effort.”
— Alex Karp, CEO of Palantir, on the DeepSeek moment (CNBC) [36]
By April 2026, DeepSeek’s V4 — open-weight, trained amid persistent American allegations of smuggled Blackwell hardware and illicit distillation of American models, and pointedly optimized for inference on Huawei’s Ascend chips reportedly at Beijing’s direction — signaled the maturation of a parallel stack: Chinese models, on Chinese silicon, under Chinese licenses, distributed globally at prices American labs decline to match.[23] The Stanford AI Index quantified the convergence that same spring: the performance gap between the best American and Chinese models had collapsed to 2.7 percent, from double digits in 2023, even though American private AI investment ($285.9 billion in 2025) exceeded China’s ($12.4 billion) by a factor of twenty-three — a ratio that itself understates Chinese state channels — and even as the migration of AI researchers into the United States fell 89 percent from its 2017 level.[6] Chatham House’s April 2026 assessment drew the conclusion that had become the emerging consensus: chips and hardware controls alone will not prevent China’s advance, because smuggling is pervasive, third-country grey markets flourish, and algorithmic efficiency increasingly substitutes for raw compute.[31] The case study thus ends where realist theory says it must: not with victory or defeat, but with two ecosystems, each hardened by the other’s pressure, dividing the world between them. The chip war did not decide the competition. It constitutionalized it.
Section 5: The Rise of Sovereign AI in Europe and the Middle East
The bipolar confrontation of Section 4 frames the era, but the fuller measure of algorithmic realism is what everyone else did while the giants grappled. This section examines the two regions whose responses have been most institutionally elaborate — Europe and the Gulf — because together they demonstrate the framework’s central claim: sovereign AI is not an ideology confined to superpowers or authoritarian systems but the rational strategy of any state, of any regime type, that can afford it. The phrase ‘sovereign AI’ itself completed a remarkable journey in barely two years, from vendor marketing coinage to budget line: by 2026 it denotes nationally controlled AI capability — the model weights, the compute, the data pipelines, and the talent — and it is a stated strategic priority of most of the G20.[13]
5.1 Europe: Autonomy Between the Blocs
Europe’s sovereign turn is the most theoretically interesting, because Europe is a formal ally of the United States and its sovereign AI program is therefore directed not at an enemy but at a dependency. The continent’s strategic anxiety is easily stated: European enterprises, governments, and militaries run overwhelmingly on American clouds and American models, and the years 2025–2026 furnished repeated demonstrations — export-policy whiplash, presidential directives reshaping model availability, the extraterritorial reach of American law — that this dependency is a live political exposure, not a theoretical one. The response has been a coordinated buildout across the compute, model, and regulatory layers. At the compute layer, the continent’s flagship projects — Mistral Compute’s 18,000-system Grace Blackwell deployment expanding across multiple sites in 2026, Britain’s Isambard-AI factory, France’s extended Jean Zay, and the twenty-plus ‘AI factories’ announced with NVIDIA — constitute the largest peacetime public-private infrastructure mobilization in European technology history.[12] At the model layer, Mistral serves as the designated champion, with the Falcon-style open-weight strategy giving European developers a non-American, non-Chinese foundation. At the regulatory layer, the AI Act and its 2026 Omnibus recalibration, examined in Section 2.3, supply the legal perimeter.
Yet the European case also exposes the hard arithmetic that sovereignty rhetoric tends to obscure. Analysts have noted that Europe’s headline investments, however politically significant, are dwarfed by the capital expenditure of individual American hyperscalers — a single US cloud provider now spends more on AI infrastructure in a year than most European sovereign programs commit across a decade — prompting the uncomfortable question of whether smaller economies might rationally prefer sharing AI power with existing leaders to competing with them.[44] The IMF’s own analysis of Europe’s prospects concluded that AI could lift the continent’s productivity, but only if the single market deepens and regulation is calibrated to enable rather than encumber — precisely the recalibration the 2026 Omnibus attempted.[43] Europe’s sovereign AI, in other words, is best understood not as a bid for parity but as the purchase of strategic optionality: enough indigenous capability to negotiate, to switch providers, to guarantee continuity of government functions in a crisis — insurance against a fragmenting order, priced at tens of billions of euros and paid willingly.
5.2 The Gulf: Hydrocarbon Rents into Algorithmic Rents
The Gulf monarchies present the purest expression of sovereign AI as state strategy, unencumbered by either the coalition politics of Europe or the sanctions perimeter around China. The United Arab Emirates moved first and remains the region’s exemplar. Its state-linked Technology Innovation Institute has iterated the Falcon open-weight family into the default base model for Arabic-language AI — the January 2026 Falcon-H1 releases, built on a hybrid Mamba-Transformer architecture with a 256,000-token context window, were framed explicitly as enabling the Emirates’ sovereign systems to process entire national archives in a single pass — while G42 anchors the compute layer through the Microsoft-backed Stargate UAE program and a Cerebras supercomputing fleet, positioning the Emirates as the AI capital of the Arab world in deliberate alignment with American technology.[15] Saudi Arabia’s answer, HUMAIN, founded in May 2025 as a wholly owned subsidiary of the Public Investment Fund, fuses sovereign wealth, national data, and imported frontier hardware into a single state champion.[13] The strategic design is legible from the realist premises of Section 1: the Gulf states are converting a depleting geological rent (hydrocarbons) into a compounding technological rent (compute and models), using their two structural endowments — capital and cheap energy — to buy a seat at a table from which their population size and research base would otherwise exclude them. In the emerging geography of compute zones, they have chosen the negotiated middle: aligned enough with Washington to receive frontier silicon, sovereign enough to own the infrastructure outright.
The regional surveys make clear how far the pattern generalizes beyond these two theaters — Singapore’s SEA-LION, India’s BharatGen, Japan’s national LLM programs, and dozens of national AI strategies among developing economies tracked by the Stanford Index all rehearse the same three rationales of supply-chain risk, data sovereignty, and strategic autonomy.[13] The universality is the theoretical payoff. When democracies and autocracies, allies and adversaries, oil exporters and city-states all converge on the same policy architecture within the same three-year window, the explanation cannot lie in ideology or leadership idiosyncrasy. It lies in structure — in the anarchic system’s uniform incentive, once AI is understood as power, to bring the algorithm home.
Section 6: Geopolitical Consequences — Fragmentation and Epistemic Hegemony
The preceding sections described a world in the act of dividing itself. This section asks what kind of world results. Two consequences dominate, operating at different depths. The first is structural and visible: the fragmentation of the global technology stack into rival spheres — an algorithmic bipolarity shading into multipolarity — with all the instability that divided systems historically exhibit. The second is deeper and quieter: a struggle over knowledge itself, in which the dominance of a few heavily protected models threatens to export the values, biases, and historical framings of their makers into the cognitive infrastructure of every society that must import its intelligence rather than produce it. The first consequence redraws the map; the second redraws the mind.
6.1 Algorithmic Bipolarity, Shading into Multipolarity
The fracture lines of the global stack are now well defined. The Western ecosystem is private-capital-driven and increasingly securitized: frontier development concentrated in a handful of American labs financed by the largest capital-expenditure wave in corporate history — global corporate AI investment reached $581.7 billion in 2025, up 130 percent in a single year, with American private investment of $285.9 billion exceeding every rival by an order of magnitude — and wrapped, since 2022, in the export-control perimeter described in Section 4.[6] The Chinese ecosystem is state-directed and surveillance-integrated, but its decisive 2025–2026 innovation was distributional: the open-weight flood, which converted exclusion from the Western hardware zone into leadership of the global software commons. Between these poles, the rest of the world is not passive terrain but an active scramble: Europe purchasing optionality, the Gulf purchasing centrality, India, Japan, Korea, Singapore, and Brazil each assembling national stacks from some blend of American silicon, Chinese open models, and indigenous data. The Stanford Index’s observation that open-source contributions from the rest of the world now outpace Europe and approach the United States suggests the multipolar layer is thickening beneath the bipolar one.[6]
The strategic character of the divided system deserves precise statement. It is not a clean iron curtain; it is a permeable, contested, and commercially entangled division — chips are smuggled across it, models are distilled across it, tariffed exceptions are sold across it — which in some respects makes it more volatile than the Cold War’s cleaner separation, because every channel of residual interdependence is simultaneously a channel of leverage and a source of dispute. Fragmentation also carries a measurable economic price. The IMF has warned repeatedly that a fragmented world is economically and socially weaker, and its Managing Director has identified deepening fragmentation as a first-order global risk of 2026.[47] Meanwhile the two spheres’ technical trajectories are converging even as their governance diverges — the 2.7 percent model-performance gap documented in Section 4 means neither pole can expect the other’s collapse — so the division must be understood as a durable operating condition of the international system, not a transitional episode. Every institution designed for the unified digital world of 2000–2020, from standards bodies to cloud contracts to academic collaboration norms, is now being retrofitted, painfully, for permanence of division.
6.2 What Is Meant by “Epistemicide”?
The second consequence requires a term of art, and the term must be defined with care because this paper asks it to carry real weight. ‘Epistemicide‘ was coined by the Portuguese sociologist Boaventura de Sousa Santos, in his work on the epistemologies of the Global South, to name the destruction of a people’s knowledge system — the extinction not of the people themselves but of their ways of knowing, their categories, cosmologies, archives, and languages of thought — historically accomplished through colonial schooling, missionary translation, and the delegitimation of local knowledge as superstition.[38] Epistemicide is to knowledge what genocide is to population and ecocide is to environment: the terminal case of a process that begins as marginalization. Allied concepts sharpen the mechanism: Miranda Fricker’s ‘epistemic injustice’ names the wrong done to people specifically in their capacity as knowers, and the decolonial literature on AI applies both ideas to the algorithmic present.[38] Algorithmic epistemicide, then, denotes the version of this destruction executed through artificial intelligence: the process by which the knowledge systems, languages, and interpretive frameworks of less powerful societies are eroded, displaced, or erased because the models that increasingly mediate all knowledge work were trained overwhelmingly on the corpus, and aligned to the values, of a few dominant cultures.
6.3 Cultural Imperialism by Default: The Mechanics of Epistemic Hegemony
The mechanism requires no malice; it is a by-product of scale economics operating through three channels. The first channel is corpus dominance. Frontier models are trained disproportionately on English-language and, increasingly, Chinese-language data; the languages, oral traditions, and knowledge systems of most of humanity are represented thinly or not at all, so the model’s ‘view from nowhere’ is in fact a view from somewhere quite specific — and every student, civil servant, and journalist in a small country who consults it absorbs that somewhere as neutral ground. The scholarly literature documents the result concretely: African, Southeast Asian, and Latin American studies report imported AI tools misaligned with local context, English-language defaults marginalizing local epistemologies, and universities locked into foreign proprietary ecosystems that store their societies’ data — even biometric data — outside national jurisdiction.[38] The second channel is alignment dominance. The behavioral norms of a model — what it will discuss, how it frames contested history, which values it treats as defaults — are set during post-training by its maker, which means that the geopolitical values and historical framings of a handful of firms in two countries are silently exported to billions of users; the American executive order of July 2025 mandating ‘ideological neutrality’ in federally procured models and China’s requirement that public models embody socialist core values are simply the two poles’ explicit acknowledgments that alignment is a site of value projection. The third channel is economic exclusion: nations that cannot afford domestic foundation models — the overwhelming majority — face a forced choice between epistemic dependence on one sphere or the other, or digital abstention. This is the erosion of informational sovereignty in its exact sense: the loss of a society’s capacity to control the epistemic infrastructure through which its own members know the world and themselves.
The international system’s most senior voices have registered the danger in language strikingly close to this paper’s framework. The United Nations Secretary-General has made the divide his signature warning:
“A world of AI haves and have-nots would be a world of perpetual instability.”
— António Guterres, Secretary-General of the United Nations [7]
The IMF’s Managing Director framed the distributive baseline even before the sovereignty wave crested:
“In most scenarios, AI will likely worsen overall inequality.”
— Kristalina Georgieva, Managing Director of the International Monetary Fund [10]
What the epistemicide analysis adds to these warnings is depth: the AI divide is not only a gap in productivity or income, remediable by transfers and capacity-building, but a gap in the means of cultural self-definition. A society that imports all of its intelligence infrastructure has outsourced part of its interpretive sovereignty — its ability to narrate its own history, reason in its own categories, and educate its young in its own voice — to the alignment teams of foreign corporations. The sovereign-AI movements of Section 5 are, on this reading, not vanity projects but acts of epistemic self-defense; and the data-fencing of Section 3.3 is the resource-nationalist instinct applied to the most intimate resource of all. The tragedy, fully visible from the realist vantage, is that the defense is available only to the wealthy. The states most exposed to algorithmic epistemicide are precisely those least able to afford the sovereign infrastructure that would prevent it — which is why the consequence of fragmentation is not many voices but few: a world of two or three epistemic hegemons, a dozen insured middle powers, and a periphery whose ways of knowing survive at the pleasure of someone else’s model card.
Section 7: Policy Implications and the Future of Global AI Governance
A framework that only diagnoses is a framework half-finished. This section draws the governance implications of algorithmic realism, and it begins with a demolition that the diplomatic community has resisted for a decade: under realist conditions, a unified, binding, ‘UN-style’ global regime for artificial intelligence is functionally impossible. It then performs the constructive task, sketching what multilateralism can realistically achieve in a protectionist world — a governance architecture built not on the fiction of harmony but on the management of rivalry.
7.1 The Failure of Global Treaties
The impossibility argument follows directly from the paper’s premises. Binding global governance of a technology requires that the leading powers accept enforceable constraints on capabilities they believe determine their relative position — and under anarchy, no rational great power accepts such constraints while a rival might defect, because national survival and dominance are lexically prior to collective safety. The empirical record of 2024–2026 is a controlled demonstration of the theorem. The United Nations did everything institutional design can do: the Global Digital Compact of September 2024, the General Assembly’s August 2025 resolution establishing an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance, the launch of the Dialogue that September, and its first working session convened in Geneva on 6–7 July 2026 with all 193 member states.[41] The Secretary-General’s framing at the launch captured the aspiration precisely:
“For the first time, every country will have a seat at the table of AI.”
— António Guterres, launching the UN Global Dialogue on AI Governance (September 2025) [39]
And yet the same launch supplied the refutation. The United States — the system’s leading AI power — stood alone in explicit opposition, with the director of the White House Office of Science and Technology Policy rejecting
“centralized control and global governance”
— Michael Kratsios, Director, White House Office of Science and Technology Policy, at the United Nations [40]
of artificial intelligence — a posture consistent with Washington’s earlier refusal to sign the Paris AI summit declaration and with an Action Plan whose third pillar is explicitly the projection, not the pooling, of American AI power.[40] Note carefully what the UN bodies that did achieve consensus have in common: the Scientific Panel assesses, the Dialogue deliberates, the proposed Global Fund builds capacity — none regulates, none verifies, none constrains. The system converged, exactly as realist theory predicts, on the subset of cooperation that is compatible with unconstrained competition. The deeper obstacles are not merely motivational but technical, and they distinguish AI from the arms-control precedents optimists invoke. Nuclear agreements succeeded because warheads are countable, testing is detectable, and the technology’s military character is unambiguous. Model weights are copyable, training runs are increasingly distributable, capabilities are emergent and unverifiable from the outside, and the same system that tutors children can plan attacks. A verification regime adequate to those properties would require intrusion into the crown-jewel facilities of rival states — precisely what rivalry forbids. Global treaty governance of AI thus fails twice over: states will not bind themselves, and could not verify the binding if they did.
7.2 Pragmatic Realist Governance
What remains is not nothing; it is a different architecture, and its outlines are already visible in practice. Realist governance abandons the universal and binding for the plurilateral and interested, and it builds on the one motive that survives anarchy intact: mutual self-interest in avoiding catastrophe and managing the costs of rivalry. Four instruments define the emerging repertoire. The first is the compute treaty — bilateral or small-plurilateral agreements governing access to hardware, of which the era already offers working examples: the security-conditioned frameworks under which frontier silicon flows to the Gulf, the allied coordination of lithography controls among Washington, The Hague, and Tokyo, and even the January 2026 US–China H200 arrangement, which — whatever its strategic wisdom — is functionally a compute treaty with a tariff clause, a negotiated, conditional, revocable opening of a choke point.[24] Because compute is physical, countable, and geolocatable, it is the one layer of the AI stack where verification is tractable; chips, unlike weights, can be audited, and hardware-level attestation and location-verification mechanisms of the kind envisioned in American legislation could become the seismographs of an eventual compute-arms-control regime.[31]
The second instrument is narrow non-proliferation: agreement among rivals not on AI writ large but on discrete, verifiable, mutually feared applications — autonomous cyber-weapons that could destabilize financial systems, AI control over nuclear command and launch decisions, model-enabled biological design. Here the great-power interest is genuinely symmetrical, the taboo logic of chemical-weapons precedent applies, and the UN Scientific Panel’s assessment function finds its realistic use: an early-warning system that gives rivals a shared factual baseline for the small set of things they both fear more than each other.[41] The third instrument is the regional data trust: pooled data governance among culturally or economically aligned states — an African Union corpus commons, an ASEAN health-data trust, Gulf and European variants — which addresses the epistemicide problem of Section 6 at the only scale where the periphery has bargaining power, by aggregating the one asset (culturally unique data) that the model hegemons cannot synthesize. The fourth instrument is capacity transfer as stabilization policy: the IMF and UN programs — the Fund’s AI Preparedness Index, the proposed Global Fund for AI Capacity Development — reframed not as charity but as systemic risk management, on the explicit understanding that an excluded periphery is an unstable periphery. The IMF’s Managing Director has reduced the entry condition to a single sentence:
Without electricity and internet access, “you cannot be part of the AI revolution.”
— Kristalina Georgieva, IMF Managing Director, IMF Annual Meetings 2025 [42]
The honest summary of realist governance is modest and should be stated without embarrassment. It will not deliver a harmonized global rulebook, and it will not arrest the fragmentation this paper has documented; nothing will. What it can deliver is what arms control delivered in the last bipolar era: guardrails at the edges of catastrophe, communication channels that survive crises, verified restraint in the narrow zones of symmetrical fear, and enough managed interdependence to keep rivalry expensive rather than existential. In a world of algorithmic realism, that is not a consolation prize. It is the entire achievable agenda — and pursuing anything grander, on the historical evidence now four years deep, means achieving nothing at all.
Section 8: The Geopolitical Shift — Two Paradigms of the Algorithm
The argument of the preceding sections can be compressed into a single structured contrast. Figure 1 sets the two governing paradigms of the algorithmic age side by side: algorithmic formalism, the Silicon Valley view that reigned from roughly 2012 to 2022, and algorithmic realism, the geopolitical view that has replaced it. The table extends the original four dimensions of the contrast with three further pillars — the treatment of data, the treatment of compute, and the ideal of governance — because the paper’s analysis has shown that the paradigm shift operates across the entire material stack, not merely at the level of rhetoric. Each row should be read as a before-and-after photograph of the same object: the algorithm as it was imagined, and the algorithm as it is now governed.
Figure 1. AI Paradigms Compared: Algorithmic Formalism versus Algorithmic Realism [2][3][6][8][11][13][41]
| Dimension | Algorithmic Formalism (Silicon Valley View) | Algorithmic Realism (Geopolitical View) |
| Nature of AI | A borderless, mathematical tool for global productivity; models as neutral scientific artifacts. | A sovereign national asset tied to state power; models as congealed geopolitical facts. |
| Strategic Goal | Interoperability, open access, rapid global deployment, and scale. | AI nationalism: technological self-reliance, strategic autonomy, and containment of rivals. |
| Weapon of Choice | Better code, open-source collaboration, benchmark supremacy, and scaling laws. | Export controls, compute monopolies, model hoarding, tariffs on silicon, and data barriers. |
| View of Data | A free-flowing global commodity; the more crossed borders, the better the models. | A sovereign natural resource; localized, fenced, and guarded like national patrimony. |
| View of Compute | An elastic utility rented from the placeless cloud. | Strategic infrastructure: national AI factories, sovereign grids, and licensed choke points. |
| Governance Ideal | Global, technocratic, light-touch rules for a borderless technology. | Plurilateral compute treaties, narrow non-proliferation pacts, and regional data trusts. |
| Geopolitical Risk | Market monopolies and digital divides. | Algorithmic warfare, compromised sovereignty, epistemicide, and civilizational bias. |
Two features of the table repay attention. First, the right-hand column is not a prediction; every entry in it is documented in this paper from the 2022–2026 policy record — the export controls of Section 4, the sovereign enclosures of Sections 2 and 5, the data fences of Section 3, the governance repertoire of Section 7. The paradigm shift is complete as a matter of practice even where the rhetoric of the left-hand column survives in corporate communications. Second, the left-hand column has not vanished so much as become subordinate: openness, interoperability, and scale are still pursued — but now as instruments within the right-hand column’s logic, as when Washington promotes American open models for their geostrategic value or Beijing floods the commons to set standards from below. Formalism has become a tactic inside realism. That inversion, more than any single policy, is the geopolitical shift this paper set out to name.
Section 9: What Have We Learned? The Five Pillars of Algorithmic Realism
It remains to distill the paper’s findings into the framework’s load-bearing propositions. Five pillars, each established in a preceding section and each carrying a lesson for the policymaker and the scholar, together constitute algorithmic realism as a usable theory.
Pillar One: The algorithm is material. Intelligence is manufactured from silicon, energy, and data — inputs that are physical, finite, and geographically bound — and therefore inherits the geography of its inputs. This is the foundational lesson of Section 1.3 and the entire chip war: whoever imagines AI as placeless mathematics will be perpetually surprised by lithography monopolies, grid constraints, and the strategic weight of one foundry on one island. Analysis must begin from the map, not from the model card.[6]
Pillar Two: Power-shifting technologies are always nationalized in function, whatever their form of ownership. Because AI shifts the balance of economic and military power under anarchy, states are structurally compelled to treat it as a sovereign asset — subsidizing champions, enclosing compute, fencing data — regardless of regime type, alliance membership, or professed ideology. The convergence of France, the Emirates, Saudi Arabia, China, and the United States on functionally identical sovereign-AI architectures within a three-year window is the paper’s decisive evidence that the cause is structural, not idiosyncratic.[13]
Pillar Three: Every channel of technological flow becomes a lever of statecraft. Weights, chips, data, talent, standards: whatever crosses the border can be granted, denied, tariffed, or traded, and therefore will be. The career of the American export-control regime — imposed, tightened, evaded, partially reversed, and finally monetized through a 25 percent levy — teaches that choke points are never merely closed or open; they are priced. Openness itself is weaponized, as the mirrored open-weight strategies of Washington and Beijing demonstrate.[24]
Pillar Four: Denial breeds capability. Export controls and technology blockades impose real costs on their targets, but against a capable adversary they also function as involuntary industrial policy, forcing indigenization and rewarding algorithmic efficiency. China’s 7-nanometer breakout, the Ascend ecosystem, and the DeepSeek efficiency revolution are the era’s proof that a determined state can convert exclusion into parallel capacity — which means denial strategies buy time, not victory, and must be priced accordingly.[30]
Pillar Five: Fragmentation is epistemic before it is economic. The deepest casualty of the divided stack is not trade but knowledge: a world in which a few heavily protected models mediate cognition exports the values and blind spots of their makers into every dependent society, eroding informational sovereignty and, at the limit, committing algorithmic epistemicide against the cultures least able to build alternatives. Governance that addresses only compute and capital, and not corpus and language, will manage the rivalry while losing the civilization.[38]
Together the five pillars answer the question posed by the paper’s title. Strategic autonomy is what states seek (Pillar Two), export controls are how the strong pursue it and provoke it (Pillars Three and Four), and the fragmentation of the digital frontier — material, political, and epistemic — is the result (Pillars One and Five). That is algorithmic realism, stated as doctrine.
Conclusion: The Age of Algorithmic Realism
This paper opened with an accounting entry — $4.5 billion of silicon stranded by a licensing letter — and closed its analysis with a divided planet. Between those two images it has argued a single continuous thesis: the era in which artificial intelligence could be understood as borderless mathematics is over, and it ended not by accident but by necessity. AI nationalism and model protectionism are the inevitable materialization, at global scale, of algorithmic realism — the recognition that a technology which shifts the balance of economic and military power will be seized, subsidized, fenced, and fought over by states whose first obligation, under anarchy, is to themselves. The evidence assembled from the 2020–2026 record is, by now, difficult to argue with. The United States built the most ambitious technology-denial regime in two generations and then sold tolled exceptions through it. China converted exclusion into a parallel stack and captured the world’s open-source commons. Europe legislated a shield, discovered its cost, and reforged it around sovereign infrastructure. The Gulf transmuted oil into compute. The United Nations, doing everything institutional ingenuity allows, achieved a dialogue, a panel, and a fund — deliberation, assessment, and charity — while the one thing it could not achieve, binding restraint, was vetoed by the structure of the system itself. Meanwhile the models converged to within 2.7 percent of each other across the divide, ensuring that neither sphere can expect deliverance through the other’s failure.[6]
Why, in the end, insist on the name ‘algorithmic realism’? The reasons given in the introduction can now be restated as conclusions. The name honors its lineage: like Green and Viljoen’s original, it refuses to let formal abstraction launder consequence — except that the consequence here is measured in fleets and famines rather than in a single courtroom’s docket.[2] The name claims its tradition: it places AI inside the oldest continuous body of thought about power under anarchy, where Thucydides’ strong still do what they can and the weak still suffer what they must — now in tokens per second. The name disciplines its user: realism is an ethic of seeing the world as it is, and the world as it is contains export ledgers, entity lists, sovereign clouds, and alignment teams quietly deciding what a billion students may ask. And the name predicts: every forecast the framework generated — failed treaties, traded choke points, universal sovereign programs, guarded weights — has been redeemed by events, which is the only vindication a framework can earn.
The final word belongs to the warning the analysis compels. Treating artificial intelligence as an ethereal, borderless math problem was always a luxury — the luxury of a unipolar moment that mistook its own dominance for the natural order of information. That luxury is gone, and it is not coming back. The states, firms, and scholars who continue to reason in the old paradigm will be perpetually ambushed by the new one: blindsided by licensing letters, stranded inventories, sovereign mandates, and the quiet epistemic capture of societies that outsourced their thinking. The era of algorithmic realism has arrived. The future of geopolitics will be written not in diplomatic ink, but in the proprietary weights of protected models and the physical borders of compute — and the first task of statecraft, scholarship, and citizenship alike is to learn to read that script.
Footnotes and Endnotes:
[1] NVIDIA Corporation — NVIDIA Announces Financial Results for First Quarter Fiscal 2026 (SEC Form 8-K, May 28, 2025). https://www.sec.gov/Archives/edgar/data/0001045810/000104581025000115/q1fy26pr.htm
[2] Ben Green & Salomé Viljoen (Harvard University / University of Michigan) — Algorithmic Realism: Expanding the Boundaries of Algorithmic Thought, ACM FAT* 2020. https://www.benzevgreen.com/wp-content/uploads/2020/01/20-fat-realism.pdf
[3] Ian Hogarth — AI Nationalism (June 13, 2018). https://www.ianhogarth.com/blog/2018/6/13/ai-nationalism
[4] Susan Ariel Aaronson (George Washington University) — The Age of AI Nationalism and Its Effects, CIGI Papers No. 306 (2024). https://www.cigionline.org/static/documents/no.306_updated.pdf
[5] Jensen Huang (NVIDIA), via John Sviokla — The AI Cold War and the Race for Sovereign AI, Forbes (November 18, 2025). https://www.forbes.com/sites/johnsviokla/2025/11/18/the-ai-cold-war-and-the-race-for-sovereign-ai/
[6] Stanford University, Institute for Human-Centered AI — The 2026 AI Index Report. https://hai.stanford.edu/ai-index/2026-ai-index-report
[7] António Guterres (United Nations Secretary-General) — Global Digital Compact — AI Panel and Dialogue. https://www.un.org/global-digital-compact/en/ai
[8] Congressional Research Service — U.S. Export Controls and China: Advanced Semiconductors, Report R48642. https://www.congress.gov/crs-product/R48642
[9] Chris Miller (Tufts University, Fletcher School) — ‘Chip War’ Author Chris Miller on the Battle of AI Chip Export Controls, CNBC Squawk Box (December 12, 2025). https://www.cnbc.com/video/2025/12/12/chip-war-author-chris-miller-on-the-battle-of-ai-chip-export-controls.html
[10] Kristalina Georgieva (International Monetary Fund) — AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity, IMF Blog (January 14, 2024). https://www.imf.org/en/blogs/articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity
[11] The White House, Office of Science and Technology Policy — Winning the Race: America’s AI Action Plan (July 23, 2025). https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf
[12] NVIDIA Newsroom — Europe Builds AI Infrastructure With NVIDIA to Fuel Region’s Next Industrial Transformation (June 2025). https://nvidianews.nvidia.com/news/europe-ai-infrastructure
[13] PDP Spectra — Sovereign AI in 2026: Mistral, G42, HUMAIN, BharatGen, and the National-AI Map (May 28, 2026). https://pdpspectra.com/blog/sovereign-ai-initiatives-2026/
[14] TechCrunch — What Is Mistral AI? Everything to Know About the OpenAI Competitor (July 4, 2026). https://techcrunch.com/2026/07/04/what-is-mistral-ai-everything-to-know-about-the-openai-competitor/
[15] TokenRing AI / FinancialContent — The Rise of the Silicon Fortress: How the Sovereign AI Movement Is Redrawing the Global Tech Map (January 9, 2026). https://markets.financialcontent.com/wral/article/tokenring-2026-1-9-the-rise-of-the-silicon-fortress-how-the-sovereign-ai-movement-is-redrawing-the-global-tech-map
[16] Gibson Dunn LLP — EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes (May 27, 2026). https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
[17] Council of the European Union — Artificial Intelligence: Council and Parliament Agree to Simplify and Streamline Rules (May 7, 2026). https://www.consilium.europa.eu/en/press/press-releases/2026/05/07/artificial-intelligence-council-and-parliament-agree-to-simplify-and-streamline-rules/
[18] Joana Soares, Tech Policy Press — EU’s AI Act Delays Let High-Risk Systems Dodge Oversight (April 2, 2026). https://www.techpolicy.press/eus-ai-act-delays-let-highrisk-systems-dodge-oversight/
[19] MIT Technology Review — What’s Next for Chinese Open-Source AI (February 12, 2026). https://www.technologyreview.com/2026/02/12/1132811/whats-next-for-chinese-open-source-ai/
[20] Tim Keary, Forbes — Why China Is Winning the Open Source AI Race (March 25, 2026). https://www.forbes.com/sites/timkeary/2026/03/25/why-china-is-winning-the-open-source-ai-race/
[21] Ben Lorica (Gradient Flow), via Open Source For You — DeepSeek Ignites Global Open Source Race With New High-Efficiency Model (November 2025). https://www.opensourceforu.com/2025/11/deepseek-ignites-global-open-source-race-with-new-high-efficiency-model/
[22] Tech Policy Press — Policymakers Overlook How Open Source AI Is Reshaping Global Power (December 9, 2025). https://www.techpolicy.press/policymakers-overlook-how-open-source-ai-is-reshaping-global-power/
[23] Chris McGuire, Michael C. Horowitz, et al., Council on Foreign Relations — DeepSeek V4 Signals a New Phase in the U.S.–China AI Rivalry (April 29, 2026). https://www.cfr.org/articles/deepseek-v4-signals-a-new-phase-in-the-u-s-china-ai-rivalry
[24] Morgan Lewis LLP — BIS Revises Export Review Policy for Advanced AI Chips Destined for China and Macau (January 16, 2026). https://www.morganlewis.com/pubs/2026/01/bis-revises-export-review-policy-for-advanced-ai-chips-destined-for-china-and-macau
[25] Council on Foreign Relations — The New AI Chip Export Policy to China: Strategically Incoherent and Unenforceable (January 14, 2026). https://www.cfr.org/articles/new-ai-chip-export-policy-china-strategically-incoherent-and-unenforceable
[26] Mayer Brown LLP — Administration Policies on Advanced AI Chips Codified, with Reverberations Across AI Ecosystem (January 22, 2026). https://www.mayerbrown.com/en/insights/publications/2026/01/administration-policies-on-advanced-ai-chips-codified
[27] Chris Miller (Tufts University) — The Shifting Politics of AI Chip Export Controls, Substack (December 11, 2025). https://chrismillersnewsletter.substack.com/p/the-shifting-politics-of-ai-chip
[28] John Power, Al Jazeera — US Says Ban on AI Chip Shipments Applies to Chinese Firms Outside China (June 1, 2026). https://www.aljazeera.com/economy/2026/6/1/us-says-ban-on-ai-chip-shipments-applies-to-chinese-firms-outside-china
[29] East Asia Forum — US Chip Export Controls Have Cooled Down (March 11, 2026). https://eastasiaforum.org/2026/03/11/us-chip-export-controls-have-cooled-down/
[30] Oplexa Research — US China Chip War 2026: Export Impact on Semiconductors (March 24, 2026). https://oplexa.com/us-china-chip-war-2026-semiconductor/
[31] Chatham House, Digital Society Programme — AI Export Controls Are Not the Best Bargaining Chip (April 29, 2026). https://www.chathamhouse.org/2026/04/ai-export-controls-are-not-best-bargaining-chip
[32] NVIDIA Corporation — NVIDIA Announces Financial Results for First Quarter Fiscal 2027 (SEC Form 8-K, May 20, 2026). https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000051/q1fy27pr.htm
[33] NVIDIA Corporation — NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026 (February 25, 2026). https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026
[34] NVIDIA Corporation — NVIDIA Announces Financial Results for Third Quarter Fiscal 2026 (SEC Form 8-K, November 19, 2025). https://www.sec.gov/Archives/edgar/data/0001045810/000104581025000228/q3fy26pr.htm
[35] CNBC — Nvidia Earnings Takeaways: Data Center Revenue Nearly Doubles (May 20, 2026). https://www.cnbc.com/2026/05/20/nvidia-nvda-earnings-report-q1-2027.html
[36] CNBC — In Fighting AI Chip Trade War With China, There’s One Big Mistake U.S. Can’t Afford to Make (February 11, 2025). https://www.cnbc.com/2025/02/11/deepseek-ai-chip-export-ban-trade-war-us-wont-win.html
[37] Development in Practice (Taylor & Francis) — Data Colonialism and Digital Sovereignty in the Global South (January 7, 2026). https://www.tandfonline.com/doi/full/10.1080/09614524.2025.2609155
[38] Frontiers in Education — Algorithmic Dependence and Digital Colonialism: A Conceptual Framework for AI in Education and Knowledge Systems of the Global South (January 2026). https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2026.1720563/full
[39] António Guterres (United Nations Secretary-General) — Remarks at the Launch of the Global Dialogue on Artificial Intelligence Governance (September 25, 2025). https://press.un.org/en/2025/sgsm22839.doc.htm
[40] Center for Strategic and International Studies (CSIS) — What the UN Global Dialogue on AI Governance Reveals About Global Power Shifts (October 14, 2025). https://www.csis.org/analysis/what-un-global-dialogue-ai-governance-reveals-about-global-power-shifts
[41] United Nations — Global Dialogue on AI Governance — First Session, Geneva, July 6–7, 2026. https://www.un.org/global-dialogue-ai-governance/en
[42] George Washington University Regulatory Studies Center — Toward a New Multilateralism for AI: Insights from the IMF Annual Meetings 2025. https://regulatorystudies.columbian.gwu.edu/toward-new-multilateralism-ai-insights-imf-annual-meetings-2025
[43] International Monetary Fund — World Economic Outlook Update (January 2026). https://www.imf.org/-/media/files/publications/weo/2026/january/english/text.pdf
[44] TIME — Why Europe’s Efforts to Gain AI Autonomy Might Be Too Little Too Late (February 2026). https://time.com/6695958/ai-infrastructure-germany-uk/
[45] AI Now Institute — AI Nationalism(s): Global Industrial Policy Approaches to AI — Executive Summary (2025). https://ainowinstitute.org/publications/ai-nationalisms-executive-summary
[46] Sidley Austin LLP, Data Matters — The Trump Administration’s 2025 AI Action Plan and Related Executive Orders (July 30, 2025). https://datamatters.sidley.com/2025/07/30/the-trump-administrations-2025-ai-action-plan-winning-the-race-americas-ai-action-plan-and-related-executive-orders/
[47] Kristalina Georgieva (IMF), World Economic Forum — AI, Skills and the Global Economy in 2026: Meet the Leader, Davos (January 2026). https://www.weforum.org/podcasts/meet-the-leader/episodes/ai-skills-global-economy-imf-kristalina-georgieva/
