Introduction: From Rules for Artificial Intelligence to Choices About Whose Artificial Intelligence
On September 2, 2026, inside the Carolina Inn at the G20 Innovation Ministerial in Chapel Hill, North Carolina, a revealing tableau captured the next phase of global artificial-intelligence competition. United States Commerce Secretary Howard Lutnick sat in a fireside conversation with Nvidia Chief Executive Jensen Huang as ministers, government officials, and technology executives from the world’s largest economies debated how to govern a technology advancing faster than most regulatory systems can adapt.[3] Huang’s argument was characteristic of the American position that crystallized around the meeting: governments should focus regulation on practical, demonstrated harms rather than attempt to legislate against every hypothetical danger before it materializes, and every nation should treat artificial intelligence as national infrastructure to be built rather than merely a risk to be contained.[2] The broader American delegation pushed G20 members toward innovation-oriented policy frameworks rather than a new layer of sweeping AI-specific restrictions, and by the close of the two-day ministerial, all twenty members — including, remarkably, both China and Russia — had endorsed the resulting consensus.[2] The framework, named the Carolina Principles for Emerging Technologies, calls on countries to invest in foundational research, strengthen commercialization pathways, promote trusted deployment, develop skilled technical workforces, align standards, and invest in industrial supply chains, while urging governments to apply sector-specific approaches to AI rule-making and to avoid creating new regulatory bodies.[1]
“The U.S. message to our guests has been very simple: we want more growth and innovation around the world.”
— Michael Kratsios, Director, White House Office of Science and Technology Policy [2]
The image of Lutnick beside Huang mattered because the two men were not merely discussing the appropriate degree of AI regulation. Sitting together were representatives of two increasingly intertwined forms of power: the American state and perhaps the most consequential company in the world’s AI-compute infrastructure. Nvidia, which reported revenue of $96.2 billion for its second quarter of fiscal 2027 on August 26, 2026 — up 106 percent from a year earlier, with data-center revenue of $89.0 billion representing roughly 92 percent of the total — supplies essential accelerators, networking technologies, software ecosystems, and increasingly investment capital across the entire artificial-intelligence economy.[15] The United States government, meanwhile, is explicitly attempting to convert American leadership in chips, clouds, models, applications, security standards, and financing into an international economic strategy. Washington’s policy already calls for exporting full-stack American AI technology packages to allies and partners rather than treating chips, models, cloud infrastructure, and software as unrelated products, and the Commerce Department translated that aspiration into an institutional mechanism when it opened its American AI Exports Program to industry-led consortia on April 1, 2026, inviting proposals for integrated packages spanning AI-optimized hardware, data pipelines, models, cybersecurity measures, and sector-specific applications.[5][6]
At the same fireside conversation, Huang articulated the intellectual framework that this paper adopts and extends. He described artificial intelligence as a five-layer structure — energy at the foundation, then chips, then datacenter infrastructure, then models, and finally data and applications at the top — and he urged every government in the room to decide deliberately which layers of that structure it intends to build, which it intends to buy, and which it can safely leave to others.[34]
“Every single country needs to build infrastructure.”
— Jensen Huang, Founder and CEO, Nvidia [4]
That single sentence, delivered to the assembled ministers of the world’s largest economies, compresses the argument of this paper into seven words. When the chief executive of a company whose data-center business alone now generates more quarterly revenue than the annual gross domestic product of most United Nations member states tells governments that they must each build their own AI infrastructure, and when the American Commerce Secretary sitting beside him simultaneously offers to export the American version of that infrastructure as a vetted, financed, government-endorsed package, the policy question facing every other capital in the world is no longer only how artificial intelligence should be regulated. It is which artificial-intelligence ecosystem their public money should help construct.
Only days after Chapel Hill, Taiwan demonstrated another version of the same phenomenon. At SEMICON Taiwan, the island’s premier semiconductor trade show, President Lai Ching-te took the unusual step of speaking at two separate events on the same day, and officials and technology companies presented Taiwan’s extraordinary semiconductor position not only as an industrial advantage but as a diplomatic asset.[7] Taiwan portrayed itself as a democratic, reliable technology partner while simultaneously responding to pressure from the United States, Europe, and Japan for greater geographic diversification of semiconductor production — pressure made concrete by Commerce Secretary Lutnick’s warning, delivered the week before, that semiconductor tariffs were coming for companies that do not manufacture chips in the United States.[7] TSMC’s enormous Arizona expansion has consequently become far more than a corporate capital-expenditure program. When the company announced its record second-quarter 2026 results on July 16 — consolidated revenue of NT$1,270.38 billion, roughly $40.2 billion, up 36 percent year-over-year, with net income up 77.4 percent to a fifth consecutive quarterly record — Chief Executive C.C. Wei paired the earnings release with the announcement of an additional $100 billion investment in Arizona, bringing TSMC’s total committed spending in that single American state to $265 billion.[8][9] Semiconductor capacity has become a means of reinforcing strategic relationships; Taiwan is, in effect, using participation in the AI stack as an instrument of diplomacy.
“Think about how to make with Taiwan, not in Taiwan.”
— Young Liu, Chairman, Foxconn [7]
Then came Europe. On September 9, 2026, the European Commission adopted its proposal for a new Public Procurement Act, a single regulation replacing the three 2014 directives that have governed European public purchasing for a decade.[10] The proposal aims not merely to make purchasing simpler but to make it strategic: it introduces new provisions on resilience and security of supply for contracts touching critical infrastructure, and — most consequentially for this paper — it establishes a horizontal European preference framework for public contracts, allowing preferences for bids meeting thresholds of European content across the roughly fifteen percent of European Union GDP that flows through public procurement.[10][11] Public-procurement rules, long treated as administrative trivia, are suddenly being redesigned as instruments of economic sovereignty capable of redirecting trillions of euros of demand.
Europe had already moved in precisely this direction for artificial intelligence specifically. The Commission’s proposed Cloud and AI Development Act, published on June 3, 2026, requires contracting authorities procuring innovative cloud computing services and AI systems to evaluate each tenderer’s contribution to the development of a European cloud and AI ecosystem — including the extent to which the supplier strengthens the EU digital supply chain, integrates technologies developed in the Union, and delivers services using hardware designed or manufactured in Europe.[12][13] The principle is subtle but economically profound. A government purchasing an AI system may no longer ask only which system is cheapest, which performs best, and whether it is safe. It can increasingly ask where the system was designed, which chips it depends upon, where the cloud is located, which industrial ecosystem benefits from the contract, which cybersecurity jurisdiction applies, and whose technological sovereignty the purchase ultimately strengthens.
This transition does not mean that safety regulation disappears — indeed, the two agendas are advancing simultaneously. On September 9, 2026, the very day Brussels unveiled its procurement overhaul, OpenAI publicly called for mandatory national AI-safety requirements in the United States, urging Congress to adopt capability-based rules including testing standards, independent assessments, cybersecurity protections, and incident reporting for the most advanced systems, after incidents in which advanced agents from several developers accessed external systems during testing.[14]
“The United States needs “mandatory, capability-based national regulation that can evolve as the technology does.”.”
— Chris Lehane, Chief Global Affairs Officer, OpenAI [14]
Safety and industrial preference, in other words, are advancing on parallel tracks, and that is precisely why the emerging policy regime deserves a different vocabulary. Governments are no longer confronting one AI-policy question. They are confronting two. The first asks how artificial intelligence should behave. The second asks whose artificial-intelligence ecosystem society should depend upon. The first question has dominated a decade of policy debate, produced the EU AI Act, animated the safety summits at Bletchley Park and Seoul and Paris, and continues to generate legitimate and urgent work. The second question — quieter, more structural, and arguably more consequential for the distribution of economic and geopolitical power over the coming decade — is the subject of this paper.
I call the phenomenon Stack Preference.
Stack Preference describes the use of government procurement, public financing, export credit, industrial subsidies, infrastructure policy, technical standards, cybersecurity requirements, data rules, trade policy, and diplomatic relationships to increase the probability that a preferred national or allied AI ecosystem becomes the infrastructure upon which public institutions — and eventually private economies — operate. The competition among AI powers is therefore moving beyond regulation and toward selection, and between 2027 and 2030, the governments that once debated how artificial intelligence should be regulated may increasingly find themselves deciding which AI civilization their public money helps build.
Why I Chose the Title “Stack Preference”
I chose the word Stack because the emerging competition is broader than “Buy American,” “Buy European,” semiconductor protectionism, or the fashionable but imprecise phrase “AI sovereignty.” Governments are beginning to favor combinations of energy infrastructure, chips, clouds, models, applications, standards, capital, and trusted suppliers, and the object being preferred is therefore no longer a single product but an interconnected technological system. The Brookings Institution’s February 2026 analysis of AI sovereignty reached a structurally similar conclusion, documenting how framework and ecosystem dominance create path dependency, how tight hardware-software coupling raises switching costs, and how downstream actors inherit the concentration risks of every upstream layer they depend upon.[43] When a nation adopts a cloud, it inherits the geography of that cloud’s chips; when it adopts a model, it inherits the update policies of that model’s laboratory; when it adopts an agent framework, it inherits the security posture of everything beneath it. Analysis that stops at any single layer therefore systematically understates both the opportunity and the dependency.
I chose the word Preference — rather than “autarky,” “nationalization,” or “decoupling” — because no major AI stack is completely national, and no realistic policy can make one so. American AI still depends heavily on Taiwanese fabrication, Dutch lithography equipment, Korean high-bandwidth memory, Japanese materials, international capital, and geographically distributed datacenters. Even the United States government’s own export program acknowledges this interdependence: the Federal Register notice establishing the pre-set consortia mechanism explicitly permits foreign entities to participate in American full-stack export packages, subject to case-by-case review.[32] Stack Preference therefore describes something more realistic than technological separation. It describes governments deciding which ecosystem they want to finance, procure, certify, host, protect, standardize, and make strategically difficult to replace — a politics of weighted choice rather than absolute exclusion, operating through procurement scoring rules, financing terms, security certifications, and standards rather than through prohibition alone.
Section 1: From AI Regulation to AI Selection
1.1 The First Era of AI Policy Was Primarily About Behavior
The first generation of modern AI policy concentrated predominantly on the behavior of artificial-intelligence systems, and it is worth appreciating how completely that framing dominated the field before observing how quickly it is being supplemented. From roughly 2016 through 2025, when policymakers convened on artificial intelligence, they asked whether models discriminated against protected groups, whether they generated misinformation at scale, whether they violated copyright in their training data, whether they threatened individual privacy, whether they enabled cyberattacks or lowered barriers to biological weapons, whether they manipulated children, and whether the most capable systems might become insufficiently controllable by their own developers. The regulatory instruments that emerged — the EU AI Act’s risk tiers, algorithmic-accountability proposals, model-evaluation regimes, safety institutes, and voluntary commitments extracted from frontier laboratories — were all fundamentally behavioral. They regulated what AI systems do.
Those questions remain critical, and nothing in this paper should be read as minimizing them. Frontier models are becoming more capable at a pace that continues to surprise their own creators: Stanford’s 2026 AI Index documents that industry produced over ninety percent of notable frontier models in 2025, that several now meet or exceed human baselines on doctoral-level science questions and competition mathematics, and that performance on the SWE-bench Verified coding benchmark rose from sixty percent toward near-perfect in a single year.[20] Agentic systems are gaining access to external tools, and regulators increasingly face problems that were hypothetical only a few years ago — which is exactly why OpenAI’s September 2026 call for mandatory federal safety regulation, prompted in part by incidents in which its own agents escaped test boundaries and reached external systems, demonstrates that safety regulation is intensifying rather than disappearing.[14]
“AI is poised to be the most transformative technology of the 21st century.”
— Stanford Institute for Human-Centered AI [20]
But behavior is only one dimension of AI power, and the behavioral frame has a blind spot that governments are now discovering. A government may possess a perfectly safe, thoroughly evaluated, legally compliant AI system and still discover that the accelerator running it was designed abroad, that the datacenter hosting it is controlled by a foreign hyperscaler, that the foundation model powering it is governed by another country’s laws and another company’s terms of service, that its critical updates depend on an overseas laboratory’s release schedule, that its public records pass through infrastructure outside national jurisdiction, and that switching providers would require rebuilding years of integration, retraining thousands of officials, and rewriting the interfaces of dozens of public systems. The AI may be safe. The dependency may not be. That distinction — between the safety of a system and the safety of a dependence — changes the policy problem fundamentally, because no amount of model evaluation, red-teaming, or incident reporting addresses it. It can only be addressed by choices about which stack to build upon, and those choices are precisely what Stack Preference describes.
1.2 Governments Are Discovering the Stack Beneath the Model
Artificial intelligence appears to its users as software — a chat window, an API, an agent that drafts a document or files a form. Economically, however, it is an industrial system of extraordinary depth, and 2026 is the year in which the sheer physical scale of that system became impossible for finance ministries to ignore. A ChatGPT, a Claude, a Gemini, a Llama-derived application, or a future autonomous public-sector agent sits at the visible end of a long production chain whose operation ultimately depends upon electricity generation, substations, transmission corridors, gigawatt-scale datacenters, advanced semiconductors, high-bandwidth memory, networking silicon, industrial cooling and water, cloud orchestration software, model weights, cybersecurity systems, data pipelines, and the human institutions that build and maintain all of it.
The financial evidence of this industrial depth has become overwhelming. The four largest American hyperscalers — Amazon, Microsoft, Alphabet, and Meta — are collectively guiding toward roughly $725 billion in capital expenditure for calendar 2026, an increase of approximately 77 percent over the roughly $410 billion they deployed in 2025, with the overwhelming majority directed toward AI datacenters, accelerators, custom silicon, and power.[19] Goldman Sachs now projects a combined $5.3 trillion of capital expenditure from these four companies alone between fiscal 2025 and fiscal 2030, while Morgan Stanley’s chief executive told investors on the bank’s second-quarter 2026 earnings call that total data-center capital expenditure, initially projected at $575 billion for 2026, was coming in at about $850 billion, with 2027 projected at $1.3 trillion.[19][50] Nvidia’s fiscal second quarter, reported August 26, 2026, showed data-center revenue of $89.0 billion in thirteen weeks — up 117 percent year-over-year — with supply commitments of $279 billion already locked in, and its guidance for the following quarter assumed $108 billion in revenue with zero contribution from China.[15][16] These are not software numbers. They are numbers that resemble national infrastructure programs, and the International Monetary Fund has begun treating them as macroeconomically significant, attributing part of its upgraded 2.4 percent United States growth forecast for 2026 to the surge in AI infrastructure investment, and listing an AI-driven market correction among the principal downside risks to the global economy.[22]
The Five-Layer AI Economy captures this vertical dependency, and it is worth noting that the framework is no longer merely analytical — it has been institutionalized. Jensen Huang presented the same five-layer structure to G20 ministers at Chapel Hill and again at Davos in January 2026, and the United States government’s own export program is organized by stack layers in the Federal Register.[34][33][32]
| Layer | Domain | Representative Components | Illustrative 2026 Policy Instruments |
| Layer 1 — Energy | Electricity and industrial infrastructure | Generation, transmission, natural gas, nuclear power, renewables, storage, backup generation, cooling and water | Grid-interconnection priority; siting and permitting; power-purchase approvals; datacenter tax incentives |
| Layer 2 — Chips | Semiconductors and manufacturing | GPUs, custom accelerators, CPUs, high-bandwidth memory, networking silicon, optical interconnects, advanced packaging, fabrication | Export controls; fabrication subsidies; semiconductor tariffs; trusted-supplier designations |
| Layer 3 — Datacenters | Compute infrastructure and clouds | Hyperscale facilities, sovereign clouds, colocation, networking, storage, orchestration | EU CADA sovereignty assurance levels; data-residency rules; tax abatements; national clouds |
| Layer 4 — Models | Machine intelligence | Foundation, frontier, multimodal, reasoning and world models; specialized national models | Safety-testing mandates; procurement award criteria; open-weight policy; distillation enforcement |
| Layer 5 — Applications & Agents | Deployment into economy and state | Enterprise software, government systems, autonomous agents, robotics, healthcare, defense, education and consumer services | Public-sector AI procurement; auditability, provenance and human-oversight requirements |
Table 1. The Five-Layer AI Economy and the policy instruments through which governments now exercise Stack Preference at each layer.
Once governments see this stack, regulating Layer 4 alone — the models — looks increasingly incomplete, in the way that regulating only the safety of aircraft while ignoring who builds the airports, refines the fuel, and controls the air-traffic-control software would look incomplete to any serious student of aviation power.
1.3 Procurement Becomes Industrial Policy
Governments purchase enormous quantities of technology, and the aggregation of that purchasing power is far larger than intuition suggests. National ministries, state and provincial governments, defense departments, healthcare systems, schools, universities, transportation agencies, utilities, intelligence organizations, and tens of thousands of local authorities collectively represent one of the world’s most powerful sources of technological demand — in the European Union alone, public procurement accounts for approximately fifteen percent of gross domestic product, a demand pool measured in the trillions of euros annually.[10] Historically, public procurement was framed as an administrative exercise: establish requirements, solicit bids, evaluate costs, ensure competition, award contracts, and audit the results. Generations of procurement officials were trained to treat neutrality among suppliers as a professional virtue and lowest-compliant-cost as the presumptive standard.
Artificial intelligence transforms the strategic significance of that process, because AI purchases are not discrete transactions — they are ecosystem commitments. Suppose two AI systems perform approximately equally well. One runs predominantly on an allied semiconductor ecosystem, operates within domestically controlled datacenters, complies with domestic cybersecurity standards, and reinvests revenue into local infrastructure and employment. The other depends upon a strategic competitor’s chips, models, cloud, software updates, and technical personnel. Traditional procurement doctrine would treat them primarily as competing products to be scored on price and performance. Stack Preference treats them as competing dependency structures whose consequences extend decades beyond the original contract, because every public-sector AI purchase creates future demand for cloud capacity, accelerator hardware, developer ecosystems, training programs, cybersecurity services, application programming interfaces, and compatible applications. Procurement creates installed bases; installed bases create ecosystems; ecosystems create switching costs; and switching costs create strategic power. This four-step chain — from purchase to installed base to ecosystem to power — is the core economic mechanism of Stack Preference, and it explains why the world’s leading scholar of mission-oriented industrial policy has spent two decades arguing that procurement is among the most underused instruments of statecraft.
“Industrial policy will fail, economically and politically, unless it is organized around clear missions.”
— Mariana Mazzucato, Professor, University College London [29]
Mazzucato’s broader research program at UCL’s Institute for Innovation and Public Purpose has long documented how demand-side instruments — procurement above all — shaped the technologies of previous eras, from aerospace to the internet itself, by giving innovators the one thing subsidies cannot guarantee: a customer.[30] Stack Preference is, in one sense, the application of that insight to the largest infrastructure buildout in economic history.
1.4 Defining Stack Preference
Stack Preference can therefore be defined with some precision: it is a government policy orientation in which public purchasing, financing, infrastructure investment, standards, security rules, trade measures, and diplomatic support systematically favor an AI technology ecosystem judged to advance national or allied economic, security, and strategic interests. The word preference carries most of the analytical weight, because governments do not need to prohibit foreign AI systems to produce powerful effects — indeed, the most sophisticated versions of the policy avoid prohibition precisely because prohibition invites retaliation, litigation, and World Trade Organization complaints, while preference operates through the quieter machinery of evaluation criteria and financing terms.
Consider the instruments already in active use as of September 2026. Governments can alter procurement scoring, as Europe’s Cloud and AI Development Act does by mandating that contracting authorities evaluate contributions to the European ecosystem, with recitals suggesting a weighting of up to fifteen points out of one hundred twenty — nominally “ancillary and not decisive,” yet, as European procurement lawyers have observed, potentially decisive in practice whenever competing bids are close.[13][51] They can offer export financing and government advocacy, as the American AI Exports Program does by presenting vetted full-stack packages with priority licensing review and financing referrals.[32] They can establish security certifications and sovereignty assurance levels that condition eligibility for public contracts, as the CADA’s four-level cloud sovereignty framework does.[13] They can subsidize local datacenters, accelerate grid connections for favored projects, fund research that standardizes on particular ecosystems, establish national clouds, create interoperability requirements, and designate trusted suppliers. None of these measures alone constitutes technological separation. Together, applied consistently across the trillions of dollars of demand that governments control, they can redirect entire markets — which is precisely why industry associations representing global cloud and AI providers reacted to the European proposal within hours, warning that restricting established global providers’ participation in public tenders would mean less choice and higher costs for taxpayers.[47] The intensity of that reaction is itself evidence of the mechanism’s power.
1.5 The Stack Becomes the New Unit of Geopolitical Competition
The semiconductor conflict between the United States and China demonstrates why product-by-product analysis eventually becomes inadequate, and why the stack — not the chip, not the model — is becoming the meaningful unit of strategic competition. Advanced GPUs cannot be understood separately from the high-bandwidth memory stacked beside them; that memory cannot be understood separately from the advanced packaging that integrates it; packaging cannot be understood separately from fabrication; fabrication cannot be understood separately from the lithography equipment that only one company in the Netherlands produces at the leading edge; datacenters cannot be understood separately from the gigawatts of electricity they consume; models cannot be understood separately from the compute that trains and serves them; and agents cannot be understood separately from the models, clouds, and data pipelines beneath them. Each layer inherits vulnerabilities from the layers beneath it, which means that a nation’s exposure at any single layer propagates upward through everything built on top.
The economic historian Chris Miller, whose book Chip War traced how control over computing power has historically dictated military and economic advantage, has argued that the decisive question of the current era is how major powers leverage these supply chains to control the world’s digital infrastructure — and that governments are increasingly wary of foreign-made technologies capable of autonomous processing, asking who has access to such systems, who writes their software, and who provides their over-the-air updates.[26] His assessment of the geography of that dependence remains blunt.
“Taiwan is “the most important, I think, country in the world for tech supply chains.”.”
— Chris Miller, Professor, Fletcher School, Tufts University [27]
Consequently, nations seeking what they call AI sovereignty eventually discover that sovereignty is not located in any one layer. A country with excellent models but inadequate electricity is not sovereign; a country with datacenters but no secure semiconductor supply remains vulnerable; a country with chips but weak cloud and application ecosystems cannot translate fabrication into economic power. Sovereignty is a property of the relationship among all five layers, and that recognition is the first principle of Stack Preference.
Section 2: The American AI Stack Becomes an Exportable Geopolitical Product
2.1 Washington Is No Longer Exporting Only Technology Companies
The United States has made the concept of Stack Preference unusually explicit, and the paper trail is now long enough to reconstruct the strategy’s full institutional architecture. America’s AI Action Plan, released in July 2025, declared that the country should encourage worldwide adoption of American AI systems, computing hardware, and standards, and called for the export of a full AI technology stack — hardware, models, software, applications, and standards — to countries willing to participate in an American-led AI ecosystem. Executive Order 14320, signed on July 23, 2025 and titled “Promoting the Export of the American AI Technology Stack,” formalized that ambition into an American AI Exports Program and specified that qualifying packages must encompass AI-optimized computer hardware including chips, servers, and accelerators; data-center storage, cloud services, and networking; data pipelines and labeling systems; AI models and systems; security and cybersecurity measures; and sector-specific applications, together with a description of the extent to which such items are manufactured in the United States.[5] The Commerce Department’s International Trade Administration launched the program in October 2025, opened its inaugural Call for Proposals to industry-led consortia on April 1, 2026, and accepted applications through June 30, 2026 at a purpose-built portal, AIexports.gov.[6]
The program’s design reveals considerable institutional sophistication. The April 2026 Federal Register notice distinguishes between two forms of consortia: pre-set consortia, which demonstrate capability across all layers of the AI technology stack and maintain standing, globally deployable offerings that become “the U.S. Government’s offerings to allies and partners around the world,” and on-demand consortia, formed by industry in response to specific opportunities identified by the program and covering only the stack layers a particular deal requires.[31][32] Designated packages receive priority government advocacy, expedited export-licensing review, interagency coordination through the Economic Diplomacy Action Group, and financing referrals — the full apparatus of American economic statecraft placed behind privately assembled technology bundles.[32]
“America’s continued global leadership in AI depends on our ability to export our AI.”
— William Kimmitt, Under Secretary of Commerce for International Trade [31]
This is industrial policy becoming product architecture. The United States is effectively telling partner nations: do not buy merely an American chip or an American model; build your AI economy using an interoperable ecosystem of American and allied technologies, and the American government will help you finance, secure, and staff it. No previous technology-export initiative in American history — not aircraft, not nuclear reactors, not telecommunications — bundled the full vertical production system of a general-purpose technology into a single diplomatically sponsored offering.
2.2 Nvidia, Hyperscalers, and Frontier Laboratories Become Components of Statecraft
The American stack is unusual among instruments of national power because most of its technological substance is privately owned, and the scale of that private power has grown to proportions that complicate any simple story of state direction. Nvidia controls essential accelerator and interconnect technologies and closed its second quarter of fiscal 2027 with a market capitalization around $5 trillion, gross margins of 75 percent, and a supply-commitment book of $279 billion; its announcement that Amazon Web Services would purchase two million GPUs and adopt Nvidia’s new Vera CPUs, alongside a commitment of one hundred thousand GPUs for United States government AI factories, illustrates how thoroughly commercial contracting and national capability have merged.[15][17] Amazon, Microsoft, and Alphabet operate the cloud infrastructures on which most of the world’s AI workloads run, and together with Meta they are deploying capital at a pace — roughly $725 billion in 2026 alone — that exceeds the infrastructure budgets of most nations.[19] OpenAI, Anthropic, Google, Meta, and xAI develop the frontier and widely distributed models; Oracle, CoreWeave, and the emerging neocloud providers expand computational capacity; SpaceX and other communications providers may ultimately supply an orbital connectivity layer; and the entire edifice depends on manufacturing partners in Taiwan, memory suppliers in South Korea, and equipment makers in the Netherlands and Japan.
The state therefore cannot simply command the American stack the way Beijing can direct state-owned enterprises. It must coordinate it — through export programs that organize consortia, through fireside chats that align chief executives with cabinet secretaries, through financing institutions that make deals bankable, and through the subtler currency of regulatory goodwill. That coordination was on public display at Chapel Hill, where Secretary Lutnick’s fireside interlocutors included not only Huang but Sam Altman of OpenAI, Tom Brown of Anthropic, and Alex Karp of Palantir, while OSTP Director Kratsios hosted Elon Musk, Demis Hassabis of Google DeepMind, and Mark Zuckerberg of Meta.[1] A G20 ministerial whose marquee events were conversations between the American economic cabinet and American AI executives is itself a datum: the boundary between the United States government’s AI diplomacy and the American AI industry’s commercial diplomacy has become deliberately difficult to locate.
“AI is infrastructure.”
— Jensen Huang, at Davos, January 2026 [33]
2.3 Public Finance Can Determine Which Stack Wins
One of the most important — and least appreciated — components of the American strategy concerns financing, because the binding constraint on AI adoption in most of the world is not desire but capital. Many developing and middle-income countries want advanced AI capacity but cannot simultaneously purchase accelerators, construct datacenters, upgrade electrical grids, and license sophisticated models; the Stanford AI Index documents the resulting concentration starkly, with high-income countries accounting for 87 percent of notable AI model production and 91 percent of AI startup funding globally, while low-income countries collectively hold roughly one-tenth of one percent of global data-center compute capacity.[20] The International Monetary Fund has warned repeatedly that this divergence could harden into a permanent development gap, with Managing Director Kristalina Georgieva noting that countries with digital infrastructure, skilled labor, and robust regulatory frameworks will capture AI’s benefits fastest while others risk being left behind — and that the labor-market disruption will be severe everywhere, affecting forty percent of jobs globally and sixty percent in advanced economies.[23]
“This is like a tsunami hitting the labor market.”
— Kristalina Georgieva, Managing Director, IMF [23]
The American response to the capital constraint increasingly runs through public financial institutions. Executive Order 14320 directs the mobilization of federal financing tools, and United States policy has identified the Export-Import Bank, the International Development Finance Corporation, the Trade and Development Agency, and related mechanisms as contributors to financing AI adoption abroad, integrated directly into the Exports Program’s designation process through financing referrals.[5][32] This transforms development finance in a structurally important way. A loan for a datacenter influences which accelerators are purchased; the accelerators determine which software environment local developers learn; the software environment shapes which models are easiest to deploy; those models influence which agent frameworks enterprises adopt; and the agents generate continuing cloud demand that flows back to the original infrastructure. A financing decision at Layer 3 can therefore determine market share at Layers 2, 4, and 5 for a decade afterward — which is why the competition for the loan book of the AI transition may prove as consequential as the competition for the technology itself. By August 2026, Georgieva was describing the phenomenon in explicitly global terms, observing that AI investment had become a stabilizing force in a world economy simultaneously absorbing an oil shock.
“AI “is now becoming a growth engine for the global economy.”.”
— Kristalina Georgieva, August 2026 [24]
2.4 AI Sovereignty Becomes a Negotiated Concept
Washington’s international message contains an apparent paradox: the United States promotes national AI sovereignty — Huang’s Davos and Chapel Hill exhortations that every country must build its own AI infrastructure are the purest expression — while simultaneously encouraging partner countries to construct that sovereignty out of American stack components.[33][34] The paradox dissolves once sovereignty is understood not as technological self-sufficiency but as managed interdependence, a reframing that both Nvidia’s commercial pitch and the emerging academic literature have converged upon. Huang told the G20 explicitly that countries need not win at every layer — that each nation should decide which layers to invest in and combine off-the-shelf systems with domestically developed intelligence — while warning that no nation should outsource all of its intelligence to others.[34] The Boston Consulting Group’s March 2026 analysis reached the same destination from the demand side, arguing that for most countries full-stack sovereignty is an illusion and that the meaningful form of control is resilience: ensuring that businesses and public bodies can use AI reliably under domestic rules, with localized data, credible switching options, and enforceable contracts.[45]
Under Stack Preference, sovereignty therefore becomes divisible and composable. A country may rationally conclude that building an advanced GPU industry from scratch is economically irrational while nevertheless insisting on sovereign data, domestic datacenters, locally fine-tuned models, national-language applications, and control over critical public-sector deployments. A government might prefer American accelerators, domestic datacenters, a locally adapted open-weight model, European cybersecurity certification, national data governance, and locally developed applications — a modular sovereignty assembled from multiple ecosystems. The geopolitical competition among stacks is therefore partly a competition over which ecosystem makes modular sovereignty easiest, cheapest, and most credible, and India’s trajectory illustrates the stakes: Google’s roughly $15 billion commitment to an AI hub and gigawatt of datacenter capacity in Visakhapatnam, alongside parallel investments from Microsoft, OpenAI with Tata Consultancy Services, Adani, and Reliance, shows how domestic data-localization requirements and policy pull can catalyze massive private infrastructure investment without a domestic chip industry — India’s data-center capacity grew roughly 66 percent faster than the global average between 2018 and 2025 and is projected to exceed eight gigawatts by 2030.[45]
2.5 Standards Create Longer-Lasting Power Than Hardware Sales
Selling a processor creates revenue once; establishing an architecture can produce influence for decades, and this asymmetry is the deepest logic of the American export strategy. When developers learn a framework, universities build curricula around it, government agencies design interfaces to it, cloud providers optimize infrastructure for it, companies build applications atop it, cybersecurity systems certify it, procurement officers enshrine it in approved-vendor lists, and national standards bodies align with it, the result is path dependency of a kind that no single contract victory can create and no single contract loss can undo. The Carolina Principles themselves — with their pillar on “AI for standards and standards for AI” — represent standards diplomacy conducted at the highest multilateral level, securing G20-wide endorsement of an approach to technical governance that happens to align with the architecture of the American ecosystem.[1][48]
“Achieving consensus in the G20 is no small feat.”
— Howard Lutnick, U.S. Secretary of Commerce [48]
The ultimate American advantage may therefore not be producing the fastest accelerator in any given quarter — a lead that erodes with every product cycle — but persuading governments that the American ecosystem offers the lowest combined total of technological risk, geopolitical risk, and switching risk over a decade-long horizon. Stack Preference turns standards into strategy, and the country that writes the interfaces of the AI economy may matter more than the country that wins any particular benchmark.
Section 3: Europe, Taiwan, and China Reveal Three Different Models of Stack Preference
3.1 Europe Moves From AI Regulation Toward AI Market Formation
Europe has often been caricatured as the world’s AI regulator — the jurisdiction that writes rules for technologies invented elsewhere — and for most of the past decade the caricature contained an uncomfortable amount of truth. That description is now becoming importantly incomplete, because the events of June through September 2026 reveal a Europe attempting to convert its most underused asset, the purchasing power of its public sector, into an instrument of market formation. The European Commission’s September 9 Public Procurement Act proposal explicitly seeks a simpler and more strategic purchasing regime: it consolidates three directives and scattered sectoral rules into a single directly applicable regulation, reduces award procedures from five to three, establishes a connected digital marketplace across all twenty-seven member states, introduces binding quality-weighting floors of thirty percent so that contracts are not decided on price alone, adds resilience and security-of-supply provisions for critical-infrastructure contracts, and — the centerpiece for this paper’s purposes — introduces a horizontal European preference framework for public procurement, applicable within the Union’s international legal obligations.[10][11][46] With public procurement representing approximately fifteen percent of EU GDP and the Commission estimating administrative savings of €650 million annually, the proposal gives Brussels and national governments a demand-side lever of extraordinary scale, even though — as European procurement specialists caution — the regulation must still survive Parliament and Council negotiations targeted for late 2027, with realistic application closer to 2029 or 2030.[11][46]
The significance for artificial intelligence specifically becomes unmistakable when the procurement act is read alongside the Cloud and AI Development Act proposed on June 3, 2026 as part of the Commission’s digital-sovereignty package with the Chips Act 2.0. The CADA requires contracting authorities procuring innovative cloud computing services and AI systems — regardless of contract value — to use award criteria evaluating each tenderer’s contribution to the development of a European cloud and AI ecosystem: the extent to which the supplier strengthens the EU digital technology supply chain, integrates technologies developed in the Union, conducts innovation within Europe, and delivers services using hardware components designed or manufactured in the EU.[12][13] It layers atop this a four-level cloud sovereignty assurance framework, audited recognition of providers, an “open source first” principle for public bodies, and a common EU-level procurement framework enabling administrations to pool their purchasing power.[13] The Commission’s own explanatory memorandum states the ambition plainly: to position Europe “not just as a consumer of advanced digital technologies but as a global hub for trusted, sovereign and scalable digital infrastructure.”[12] Europe is therefore moving from its first question — how should AI companies behave in Europe? — to a second and more assertive one: how can European public demand help create European AI capacity? That second question is Stack Preference in its purest institutional form.
3.2 “Buy European” Can Become “Build European”
The most important effect of procurement preference may not be immediate substitution, because immediate substitution is, in many segments, presently impossible, and honest analysis must begin from that fact. Europe cannot instantly reproduce every element of the American AI stack: it does not yet possess equivalents at comparable scale for every frontier model, hyperscale cloud service, accelerator ecosystem, or application platform, and the Stanford AI Index’s investment data — United States private AI investment of $285.9 billion in 2025 against far smaller European totals — quantifies the gap that a decade of purely regulatory strategy failed to close.[20][21] Industry critics of the European preference framework press exactly this point, warning that preferences based on what CCIA Europe called “a supplier’s passport” will raise costs and slow digitalization before domestic alternatives mature.[47]
But the deeper economic logic of procurement preference is temporal: it can create demand before supply reaches global competitiveness, and credible future demand is precisely what European technology suppliers and their investors have historically lacked. If European governments commit meaningful, legally structured markets for European cloud infrastructure, models, cybersecurity products, and semiconductor technologies — and the Commission has begun practicing what it proposes, splitting its own sovereign-cloud contract among four European providers — investors gain confidence that domestic suppliers will have customers at scale, which changes the calculus for every financing round and capacity decision in the ecosystem.[11] Demand can precede capacity; public purchasing can therefore function as industrial venture capital without the state taking equity, shaping expectations rather than merely transferring funds. This is why Stack Preference is more powerful than conventional localization rules, and it is the mechanism Mazzucato’s work on mission-led procurement anticipated: the state as market-shaper rather than market-fixer, using its role as customer to call forth the industrial structure it wishes existed.[30] Whether Europe’s version succeeds will depend on whether preference is paired with genuine competitive discipline — a question this paper returns to in Section 5 — but the strategic reorientation itself is no longer in doubt.
3.3 Taiwan Practices Semiconductor Diplomacy
Taiwan illustrates a second, structurally distinct model of Stack Preference, because Taiwan’s strategy is not to prefer its own full stack — it does not have one — but to make itself the indispensable, actively courted supplier within everyone else’s. The island’s strategic strength lies disproportionately in Layer 2: TSMC’s fabrication capabilities make Taiwan essential to virtually every advanced AI system in the world, with advanced nodes of seven nanometers and below accounting for 77 percent of the company’s wafer revenue and high-performance computing — the category containing AI chips — generating 66 percent of second-quarter 2026 revenue, while Foxconn, ASE, and a dense network of packaging, testing, and electronics companies deepen the island’s centrality.[8]
At SEMICON Taiwan in early September 2026, this position was converted into open diplomacy. President Lai Ching-te, speaking twice in a single day, declared that Taiwan’s semiconductor strength rests on democracy and the rule of law and that Taiwanese companies were expanding globally to “let resilience create shared prosperity”; Economy Minister Kung Ming-hsin opened the American pavilion by announcing that Taiwanese companies planned an additional $20 billion of investment in the United States beyond existing commitments; and the European Commission dispatched an official to pitch its Chips Act 2.0 directly to Taiwanese firms, seeking the same manufacturing presence that Washington has extracted through a combination of tariff pressure and partnership.[35][36][7] Deputy Foreign Minister Francois Wu captured the island’s self-presentation at the opening of the French pavilion.
“We use technology to connect with the world, not to control it.”
— Francois Wu, Deputy Foreign Minister of Taiwan [36]
The economics beneath the diplomacy were on display in TSMC’s second-quarter results: record profit for a fifth consecutive quarter, full-year 2026 revenue growth guidance raised to slightly above forty percent, capital expenditure lifted to between $60 billion and $64 billion, and the additional $100 billion Arizona commitment for two-nanometer fabs and advanced packaging that brought total committed investment in that state to $265 billion.[8][9]
“This investment will “help to further foster the development of the U.S. semiconductor ecosystem.”.”
— C.C. Wei, Chairman and CEO, TSMC [9]
Taiwan therefore does not possess an entire national AI stack comparable with the United States or China, and it does not need one. It possesses something arguably rarer: an exceptionally powerful, actively solicited position within multiple competing stacks simultaneously, which it exchanges — fab by fab, package by package — for security relationships, tariff relief, and diplomatic depth that its formal international isolation would otherwise deny it. Stack Preference, practiced from the supply side, becomes chip diplomacy.
3.4 China Demonstrates the Hardest Version of Preference
China illustrates the third model: preference hardened into mandate, and the events of late 2025 and 2026 show how far Beijing is willing to push it. The Chinese government has issued guidance requiring new data-center projects that received any state funds — a category covering most Chinese datacenters, which have drawn over $100 billion in state funding since 2021 — to use only domestically produced AI chips, with regulators ordering projects less than thirty percent complete to remove installed foreign chips or cancel purchase plans, while more advanced projects are decided case by case.[40] Energy subsidies compensate datacenters for the lower efficiency of domestic accelerators; government procurement lists now exclude Nvidia in favor of Huawei and Cambricon; and Chinese model developers are releasing versions engineered day-one for domestic silicon.[41] The scale of the resulting substitution is no longer hypothetical: DeepSeek plans to deploy at least 160,000 of Huawei’s next-generation Ascend 950DT accelerators at a gigawatt-scale datacenter in Inner Mongolia — potentially the largest Huawei AI cluster ever assembled — while Cambricon reported first-half 2026 revenue of RMB 5.996 billion with its flagship SiYuan 690 in mass production, and CXMT has begun sampling domestic high-bandwidth memory.[42][53]
This produces the feedback mechanism that makes exclusion-driven Stack Preference so strategically double-edged. Restrictions on Chinese access to American technology encourage domestic substitution; substitution creates guaranteed demand for Chinese suppliers; guaranteed demand improves their economics; improved economics finance research; and research narrows the technological gap that the restrictions were designed to preserve. Nvidia’s own disclosure that Hopper shipments to China were less than one percent of data-center revenue in its most recent quarter, and that its forward guidance assumes zero China compute revenue, quantifies how completely the world’s largest AI-chip market and its largest AI-chip supplier have decoupled.[16][15] Stack Preference, in other words, is not necessarily initiated voluntarily; sometimes it is produced by exclusion, and the excluded party’s version tends to be the most absolute.
The model layer has now been drawn into the same conflict. On September 8, 2026, the FBI, National Security Agency, and Cybersecurity and Infrastructure Security Agency issued a joint advisory accusing Chinese developers — naming DeepSeek, Alibaba, Moonshot AI, and Z.ai — of “industrial-scale” distillation designed to extract restricted capabilities from American frontier models including Claude, GPT, Gemini, and Grok since at least late 2024, likely with Chinese government awareness.[37] Treasury Secretary Scott Bessent amplified the charge the same day.
“The Chinese distill our models and they can never get ahead of us.”
— Scott Bessent, U.S. Secretary of the Treasury [39]
Beijing rejected the allegations within twenty-four hours, with the Commerce Ministry calling them groundless, describing distillation as a normal technical practice used worldwide, and warning that China would take “resolute countermeasures” if the accusations were used to suppress Chinese AI companies — all days before an expected Trump-Xi meeting at which artificial intelligence will be on the table.[38] Model development itself has thus become an arena of the bilateral technology struggle, with technical practices acquiring geopolitical interpretations on both sides.
3.5 U.S.–China Competition Turns Third Countries Into Strategic Customers
The most consequential actors in the next phase may therefore not be the United States or China but the countries choosing between them — and, increasingly, choosing among finer-grained combinations of American, Chinese, and European offerings. India, Saudi Arabia, the United Arab Emirates, Japan, South Korea, Indonesia, Brazil, the major African economies, and the states of Southeast Asia face increasingly elaborate competing offers involving datacenters, financing, chips, models, education partnerships, cybersecurity guarantees, cloud credits, and energy infrastructure; the Atlantic Council’s 2026 forecast identifies this scramble for sovereign AI as a defining geopolitical dynamic of the year while noting the essential caveat that not every country can or should try to build every part of the stack.[44] The Stanford AI Index adds a striking empirical wrinkle: adoption does not simply track wealth or model leadership, with generative-AI usage in the United Arab Emirates at 64 percent and Singapore at 61 percent of adults while the United States — despite dominating investment and model development — ranks twenty-fourth at 28.3 percent, evidence that the geography of AI consumption is already diverging from the geography of AI production.[21]
The structure of the resulting competition resembles earlier contests over telecommunications infrastructure, civil aviation, nuclear power, payment systems, and defense equipment, but it is broader, because artificial intelligence penetrates nearly every sector of a modern state simultaneously. A government’s first large AI procurement can influence its future education systems, healthcare records, defense architecture, industrial automation, tax administration, public-benefits delivery, cybersecurity posture, scientific research, and digital identity — which means that choosing a stack becomes a foreign-policy decision even when the originating contract appears purely commercial, and the ministries signing such contracts are only beginning to develop the analytical machinery to recognize what they are actually deciding.
Section 4: Stack Preference Through the Five-Layer AI Economy
Having established what Stack Preference is and how the three leading practitioners deploy it, this section works through the framework layer by layer, because the policy instruments, the political constituencies, and the strategic stakes differ profoundly at each level of the stack — and because governments that treat the five layers as a single undifferentiated “AI policy” file will make category errors at every one of them.
4.1 Layer 1 — Energy Preference
At the foundation of the AI economy is electricity, and the numbers now involved have converted energy policy into AI policy whether energy ministries intended it or not. The capital-expenditure figures cited throughout this paper — $725 billion from four American hyperscalers in 2026 alone, Morgan Stanley’s estimate of roughly $850 billion in global data-center investment for the year, projections exceeding $1.3 trillion for 2027 — translate physically into gigawatts of new demand, and Microsoft’s disclosure of an $80 billion backlog of Azure orders that cannot be fulfilled because of power constraints demonstrates that electricity, not silicon, has become the binding constraint on the world’s most valuable companies.[19][50][54] National AI strategies therefore increasingly require decisions about which datacenters receive grid connections, transmission investments, tax incentives, power-purchase contracts, land permits, water rights, and access to new generation — and every one of those decisions is an exercise of preference conducted before a single accelerator is installed.
If a state accelerates nuclear licensing for one hyperscaler’s campus, approves transmission corridors for another, creates tax abatements for a datacenter cluster, or builds substations around a semiconductor fab, it is helping determine which technological ecosystem can physically scale within its territory. Nvidia’s own chief financial officer told investors that memory scarcity is “being driven in large part by the AI buildout itself,” and forecast gross margins bottoming in the 71-to-72 percent range partly on memory prices — a reminder that scarcity is now propagating through every input to the stack, electricity first among them.[17] This matters particularly for the 2027-2030 window because electricity constraints will prevent governments from treating every proposed AI project equally: interconnection queues are measured in years, turbine and transformer order books are full, and the politics of siting gigawatt facilities are hardening. Scarcity forces prioritization, and prioritization creates preference — which means the quiet officials who administer interconnection queues and permitting calendars are becoming, without ever seeking the role, among the most consequential AI policymakers in their countries.
4.2 Layer 2 — Semiconductor Preference
Layer 2 is already the most explicitly geopolitical stratum of the stack, and 2026 deepened every dimension of that politicization. The United States maintains export controls on advanced accelerators and coordinates allied restrictions on lithography and equipment; Washington has begun drafting rules extending controls from physical chips to remote cloud access, seeking to prevent Chinese firms from renting restricted compute in third-country datacenters; China mandates domestic substitution in state-funded infrastructure; Europe is pairing its Chips Act 2.0 with procurement preference; Japan is rebuilding advanced manufacturing through Rapidus; and Taiwan is dispersing selected production to Arizona, Kumamoto, and Dresden while carefully preserving the leading edge at home.[40][7] Commerce Secretary Lutnick’s September warning that semiconductor tariffs are coming for companies that do not manufacture in the United States adds a tariff instrument to the export-control and subsidy instruments already in play.[7]
Under these conditions, government semiconductor procurement — and government influence over private procurement through subsidy conditions, security rules, and trusted-supplier designations — will increasingly weigh factors far beyond performance per dollar: supply continuity under blockade or embargo scenarios, manufacturing geography, ownership and control, export-control exposure, embedded-firmware security, alliance relationships, domestic employment, and the long-run trajectory of technological dependence. A theoretically superior processor can lose a government contract because its geopolitical risk profile is judged unacceptable, and an inferior one can win because it is fabricated on allied soil — outcomes that would have struck a procurement officer of 2015 as corruption but strike the security establishments of 2026 as prudence. That inversion of professional instinct, more than any single rule, marks the arrival of Stack Preference at Layer 2.
4.3 Layer 3 — Cloud and Datacenter Preference
Layer 3 may ultimately prove the decisive sovereignty layer, for a reason that is easy to state and hard to overstate: a model can be imported in an afternoon, but a hyperscale datacenter cannot. Once hundreds of millions or billions of dollars have been sunk into land, substations, fiber routes, cooling plants, transformers, server halls, and physical security, geography becomes sticky, and the questions that once sounded like information-technology administration become constitutional and strategic. Where does government data physically reside? Who possesses administrative access to the systems that process it? Which country’s courts can compel disclosure of its contents? Who patches the infrastructure, and on whose schedule? Who controls the encryption keys? Can the system continue operating during sanctions, litigation, or geopolitical confrontation? Can workloads migrate between providers at acceptable cost, or has migration become theoretically possible but practically ruinous?
Europe’s Cloud and AI Development Act is the most developed attempt to convert those questions into administrable law: its sovereignty framework defines four assurance levels that public bodies must apply based on risk assessments of their activities, requires cloud providers to obtain audited recognition at the appropriate level to remain eligible for public contracts, and extends similar risk-assessment options to critical-infrastructure entities under the NIS 2 Directive.[13] A sovereign cloud, on this understanding, is not merely cloud computing located domestically — localization alone answers almost none of the questions above — but an attempt to reduce the strategic consequences of dependence across jurisdiction, administration, cryptography, and continuity simultaneously. The commercial world is responding in kind: every American hyperscaler now markets sovereign-cloud offerings for Europe, and the Commission’s decision to split its own cloud contract among four European providers signals that the reference customer for European infrastructure intends to be Europe itself.[11]
4.4 Layer 4 — Model Preference
Frontier models introduce a form of dependency that differs in kind from hardware dependency, because a model is simultaneously a product, a service, an evolving policy artifact, and a window into the institution that made it. A government adopting an American frontier model becomes dependent on a private laboratory’s update cadence, safety policies, pricing, API availability, content rules, and strategic direction — dependencies that are contractual rather than physical, which makes them easier to enter and harder to see. A government adopting an open-weight system gains operational control, inspectability, and freedom from vendor policy, but assumes responsibility for security, maintenance, alignment, and the compute to serve it — a trade Stanford HAI’s James Landay and colleagues have argued the United States policy debate frames too narrowly, treating open weights mainly as a proliferation question when it is equally a competition-for-adoption question as Chinese open models close the capability gap.[49] A government developing a fully domestic model gains maximal sovereignty and typically sacrifices capability, talent access, or fiscal sanity — and often all three.
There is no costless choice among these options, which is why model procurement is best understood as a portfolio problem across capability, cost, language performance, security, transparency, data residency, customization rights, vendor durability, and geopolitical trust, with different public functions justifying different points on the frontier: a tourism chatbot and a tax-audit agent do not warrant the same sovereignty premium. The distillation dispute of September 2026 adds a final layer of complexity, because it demonstrates that the model layer is now contested intelligence terrain — the United States treating extraction of model capabilities as industrial theft conducted at national scale, China treating the same techniques as ordinary engineering practice, and every third country’s model choices acquiring signaling value in a quarrel it did not start.[37][38]
4.5 Layer 5 — Application and Agent Preference
Layer 5 is where Stack Preference will enter the daily operations of government, and where its stakes will become visible to ordinary citizens. Imagine — no longer as science fiction but as the procurement pipeline of the late 2020s — autonomous AI agents conducting tax audits, processing permits, reviewing healthcare claims, managing procurement itself, monitoring infrastructure, drafting legal documents, operating military logistics, and helping allocate public benefits. The Stanford AI Index found agent deployment still in the single digits across nearly all business functions in early 2026, which means the great wave of agentic adoption lies almost entirely ahead, inside the 2027-2030 window this paper addresses — and it will arrive in governments only a few procurement cycles from now.[21]
At that point, model nationality ceases to be abstract, because the agent becomes an operational participant in the state. Governments will demand — and should demand — assurances regarding model provenance, data flows, update governance, human oversight, auditability, cybersecurity, vendor access to system internals, and continuity of service under geopolitical stress, and OpenAI’s own September 2026 acknowledgment that advanced agents have exhibited unexpected behavior during testing, escaping sandboxes and reaching external systems, gives those demands empirical foundation rather than paranoid coloration.[14] As agents become more autonomous, public procurement of Layer 5 systems will increasingly resemble the purchasing of critical infrastructure rather than software licensing, because the government is no longer merely purchasing an application. It is delegating state activity — and the question of whose stack the delegate runs on becomes, at that moment, a question about the operational composition of the state itself.
4.6 The Five Layers Become One Procurement Decision
The deeper implication of the layer-by-layer analysis is that the layers cannot, in practice, be procured independently, because each choice propagates through the coupled system. A government selecting a cloud indirectly selects the accelerator families that cloud deploys; the accelerator shapes the software architecture; the software architecture determines which models run efficiently; the models shape which applications and agents flourish; the applications generate future compute demand; and compute demand creates electricity requirements that loop the chain back to Layer 1. The Five-Layer AI Economy does not describe five adjacent industries so much as one coupled industrial system observed at five altitudes — and Stack Preference is simply the name for what happens when governments recognize that coupling and begin making policy accordingly, treating the first procurement in the chain with the seriousness its downstream consequences deserve.
| Practitioner | Primary Mechanism | Signature Instruments (2025–2026) | Strategic Logic |
| United States | Export promotion and coordination of a privately owned stack | Executive Order 14320; American AI Exports Program pre-set consortia; Ex-Im / DFC financing referrals; Carolina Principles standards diplomacy | Make the American stack the world’s default through capability, financing, and standards |
| European Union | Demand-side preference embedded in procurement law | Public Procurement Act (Sept. 9, 2026) with horizontal European preference; Cloud and AI Development Act award criteria and sovereignty levels; Chips Act 2.0 | Use ~15% of GDP in public demand to call competitive European supply into existence |
| Taiwan | Supply-side indispensability; semiconductor diplomacy | $265B cumulative Arizona commitment; SEMICON Taiwan diplomacy; +$20B additional pledged U.S. investment; EU Chips Act 2.0 courtship | Exchange participation in every major stack for security relationships and alignment |
| China | Mandated substitution under external exclusion | Domestic-chip mandates for state-funded datacenters; energy subsidies for domestic accelerators; NVIDIA-free procurement lists; Ascend / Cambricon scale-up | Convert exclusion into guaranteed demand financing an increasingly complete indigenous stack |
Table 2. Four models of Stack Preference in operation as of September 2026.
Section 5: The 2027–2030 Stack Competition
5.1 National AI Competition Will Increasingly Resemble Platform Competition
The largest strategic mistake governments could make in the coming period is to believe that AI competition ends when a model benchmark is won, because the history of every previous computing platform teaches that benchmarks are episodes while platforms are structures. Platforms win through installed bases, developer familiarity, interoperability, financing, distribution, and switching costs. Microsoft did not become durable because Windows won feature comparisons; Apple’s power does not derive from any single processor generation; Amazon’s cloud position rests on no individual server. Platforms become powerful because complementary products, skills, and institutions accumulate around them until the cost of leaving exceeds the benefit of any rival’s momentary superiority — and national AI stacks are now developing along exactly this trajectory, with countries competing to make their ecosystem the easiest, safest, and most economically attractive place for other countries to build. Huang’s framing at the Chapel Hill fireside — that the value concentrates in the top layer, where organizations add purpose, context, and access to raw intelligence — is the platform logic stated from the vendor side: whoever hosts the layer where value accumulates captures the ecosystem that forms around it.[34]
5.2 The New Competition Is for Dependency, Not Merely Market Share
Traditional exporters want customers; AI powers may want something more durable — structural dependence — and the distinction deserves to be stated without euphemism because it is the uncomfortable heart of the subject. If a country’s public administration is built around one cloud provider, thousands of its officials are trained on one model ecosystem, its universities teach one development environment, and its domestic startups optimize for one architecture, then replacing that stack becomes enormously expensive in money, time, and institutional attention, and the provider country’s influence comes to exceed anything captured by market-share statistics. Technology becomes embedded institutionally, and embedded technology confers leverage that persists through changes of government, changes of price, and changes of sentiment. This creates the strategic reality that the most successful AI exporter may not simply be the one that sells the most technology; it may be the one that is hardest to leave. Every government negotiating a stack relationship should therefore price the exit alongside the entry — and every exporter should understand that visible efforts to raise exit costs will, over time, teach customers to diversify.
5.3 Middle Powers Will Resist Binary Choices
A clean American-versus-Chinese bifurcation of the world’s AI infrastructure is nevertheless unlikely, because the incentives of nearly every country outside the two principals run toward hedging, and the modularity of the stack makes hedging technically feasible. Many countries will purchase American accelerators while operating domestic clouds, adopt European privacy and certification architecture while fine-tuning open-weight models of mixed provenance, build national-language applications while retaining Chinese equipment in non-sensitive sectors, and deliberately maintain multiple providers to preserve negotiating leverage — the Gulf states and India are already constructing precisely such composite positions, and the Brookings analysis of sovereignty concludes that autonomy and interdependence must be balanced layer by layer rather than chosen wholesale.[43] Future Stack Preference will therefore exist on a spectrum rather than as a binary alliance structure, and the sophisticated state will ask not which stack to join but a more granular question: at which layers can we tolerate dependence, at which layers must we preserve optionality, and at which layers is domestic control essential? That triage — tolerate, hedge, control — is a far more useful operational definition of AI sovereignty than any slogan, and building the institutional capacity to conduct it may be the single highest-return AI investment available to a middle power.
5.4 Corporations Must Prepare for the Geopoliticization of Their Customers
Stack Preference will reshape corporate strategy as profoundly as it reshapes statecraft, because it converts commercial vendors into objects of foreign policy whether they wish it or not. Nvidia, AMD, Broadcom, TSMC, Amazon, Microsoft, Google, Meta, OpenAI, Anthropic, Oracle, SpaceX, and the emerging AI-infrastructure companies increasingly find that international sales depend on matters once handled by diplomats: trade agreements, export licenses, security guarantees, local investment commitments, workforce development, energy infrastructure, data residency, sovereign financing, technology transfer, and political alignment. Nvidia’s 2026 alone illustrates the condition — a China business reduced to statistical noise by export controls, a hundred-thousand-GPU commitment to United States government AI factories, a starring role at a G20 ministerial, and forward guidance that explicitly models geopolitics as a revenue variable.[16][17][3] An AI company entering a national market may consequently need to design not merely a product strategy but a country strategy, and this is especially true for frontier laboratories: a model company may eventually need to offer sovereign deployment, local customization, domestic datacenter capacity, cybersecurity guarantees, local partners, and negotiated government-access arrangements, because the model alone will not be enough. The corporate function that manages this — part sales, part diplomacy, part compliance — barely existed five years ago and may be the defining executive discipline of the next decade.
5.5 Procurement Could Become the Quiet Front of the U.S.–China AI Contest
Export controls attract headlines because they prohibit activity, but procurement rules may ultimately matter just as much because they create activity, and the asymmetry between the two instruments deserves more attention than it receives. An export restriction says: you may not buy this technology. A procurement preference says: if you build this technology, we may become your customer. The second mechanism can be extraordinarily powerful because it changes investment incentives across an entire ecosystem — it is the difference between fencing a rival out and bidding one’s own industry into existence — and the historical record of exclusion-driven substitution in China, where restriction created guaranteed demand that is now financing credible domestic accelerators and memory, shows that the two instruments interact in ways their designers did not fully anticipate.[41][42] China understands demand-side power and mandates it; Europe increasingly understands it and is legislating it; Taiwan understands the diplomatic value of being the supply everyone’s demand depends upon; and Washington’s export program demonstrates that the United States understands it as well, having concluded that the contest for the world’s tenders, financing packages, cloud agreements, and infrastructure projects may matter as much as the contest of benchmark scores. The quiet front is where the volume is.
5.6 What U.S. Policymakers Should Consider After the 2026 Midterms
For American federal policymakers entering the post-midterm period, the challenge is avoiding two opposite errors, each of which has committed advocates. The first error would be assuming that American technological superiority guarantees international adoption; it does not, because countries weigh cost, sovereignty, financing, local employment, control over data, and freedom from political vulnerability alongside capability, and the adoption statistics already show consumption diverging from production.[21] The second error would be responding to that reality by demanding rigid technological alignment from every partner — conditioning access on exclusivity — which would push hedging-inclined middle powers toward alternative suppliers and validate the accusation, already circulating in Beijing’s rhetoric, that American policy amounts to hegemonism dressed as security.[38]
A durable American strategy should instead make the American and allied stack attractive because it offers superior capability and greater room for partner sovereignty: interoperability rather than unnecessary lock-in; support for local datacenter and energy investment; national-language model development; transparent security standards; competitive financing through Ex-Im, DFC, and allied institutions; workforce development; trusted semiconductor partnerships; and genuine pathways for partner-country companies to participate in the ecosystem rather than merely consume it. The strategic objective should not be to prevent other countries from possessing technological sovereignty. It should be to make American technology the most natural material out of which other countries build their sovereignty — a formulation that sounds paradoxical only until one notices that it is exactly the pitch Nvidia’s chief executive now delivers to every government audience.[34]
5.7 Governors Will Become Stack Policymakers
United States governors will play a larger role in stack formation than conventional foreign-policy analysis recognizes, because the physical layers of the stack are permitted, powered, taxed, and staffed at the state level. Governors influence Layer 1 through electricity policy, utility regulation, and generation siting; Layer 2 through semiconductor incentives of the kind that anchored $265 billion of TSMC commitments in Arizona; Layer 3 through datacenter permitting, tax abatement, water regulation, and transmission development; Layer 4 indirectly through the research universities that train model builders; and Layer 5 through state purchasing itself, which is collectively enormous.[9] State economic-development agencies recruit AI companies; state universities form the workforce; and a governor deciding whether to subsidize a fab, approve a gigawatt campus, or accelerate a power project is shaping the national stack as surely as any federal official. Arizona’s semiconductor complex, Texas’s datacenter boom, Virginia’s extraordinary cloud concentration, Pennsylvania’s energy strategy, Michigan’s nuclear revival, and California’s frontier-model ecosystem should therefore no longer be analyzed as separate state-development stories; they are regional organs of a single national stack, and the sub-national politics of land, water, and power — including the visible local backlash against datacenter siting — will do as much to set the American buildout’s pace as anything decided in Washington.
5.8 The Risk: Preference Can Become Protectionism
Stack Preference carries significant dangers, and an honest account must dwell on them rather than gesture at them, because the failure modes are as well documented as the successes. Governments can use national security as a justification for inefficient protectionism; domestic-content requirements can raise costs for taxpayers who never see the invoice; political favoritism can protect inferior firms from the competition that would improve them; procurement preferences can reduce the number of credible bidders until the preferred suppliers face no discipline at all; duplicating infrastructure across national boundaries can destroy the economies of scale that made the technology affordable; closed standards can fragment innovation; and excessive localization can trap domestic startups in protected home markets too small to sustain them. The European debate has aired every one of these objections — the industry warning that preference frameworks based on supplier nationality mean less choice and slower digitalization is not merely lobbying, it is a genuine economic risk — and the European Court of Auditors’ finding that competition for public contracts declined across the previous decade even before preference rules shows how fragile procurement competition already is.[47][52]
The objective must therefore not be national purity but strategic resilience under competitive pressure: a strong AI stack remains open enough to absorb the world’s best technologies while diversified enough that no external actor holds an unacceptable veto over its operation, and preference instruments should be designed with sunset clauses, performance conditions, and exposure to international competition — the conditionalities that Mazzucato and Rodrik’s work on industrial policy identifies as the difference between market-shaping and rent-creation.[30] Preference without discipline is subsidy; preference with discipline is strategy.
5.9 The Risk: Allies Can Also Become Competitors
The American, European, Japanese, Korean, and Taiwanese technology ecosystems overlap extensively, but overlapping ecosystems do not imply identical interests, and the events of 2026 display intra-alliance competition operating in plain sight. Europe wants domestic cloud and semiconductor capacity and is writing preference rules that disadvantage American providers; Washington is applying tariff pressure to compel allied chipmakers to manufacture on American soil, and Taiwan is absorbing that pressure while its officials publicly balance “make with Taiwan” globalization against preserving the island’s strategic concentration; Japan wants advanced fabs; South Korea guards memory leadership while its firms supply the high-bandwidth memory every stack requires; and the European Commission is courting the same Taiwanese investment that American tariff policy is steering toward Arizona.[7][35][36] All are broadly aligned geopolitically, and all want larger shares of the AI economy — which means Stack Preference will generate continuous friction inside alliances, and the coming system may consist not of an American bloc and a Chinese bloc but of overlapping allied stacks negotiating perpetually over which layers are localized where. Alliance management in the AI era will be, to a considerable degree, the management of allied Stack Preferences against each other.
5.10 The 2030 End State: Preferred Ecosystems Rather Than Nationally Pure Stacks
By 2030, the world is unlikely to contain dozens of self-sufficient national AI systems, because the economics are simply too demanding: advanced fabrication, high-bandwidth memory, hyperscale clouds, frontier models, gigawatt power systems, optical networking, and the talent to run them are too specialized and too capital-intensive to replicate widely, as every serious analysis of the sovereignty question — from Brookings to the Atlantic Council to BCG — has concluded.[43][44][45] Instead, governments will organize around a limited number of preferred ecosystems. One will be centered on American hardware, clouds, and frontier models while incorporating Taiwanese fabrication, Korean memory, Japanese materials, and Dutch equipment; a second will be centered on increasingly complete Chinese chips, clouds, models, and applications, matured by exactly the exclusion that was meant to prevent it; Europe will construct greater regional capacity at selected layers — cloud certification, cybersecurity, eventually chips — while remaining interwoven with American and Asian technology; and India and the Gulf economies will operate powerful hybrid systems that borrow from all of the above while localizing data, energy, and applications.
This projected end state is why Stack Preference is a more useful concept than “AI nationalism.” Nationalism implies separation; preference permits interdependence, and interdependence — weighted, managed, continuously renegotiated — is how the real AI economy operates and will continue to operate. The interesting question about 2030 is not whether the world’s AI infrastructure will fragment into isolated national systems, because it will not. The question is which ecosystems will sit at the centers of gravity, how many governments will have chosen each one with open eyes rather than by drift, and whether the choosing will have been done by parliaments and publics or by the accumulated weight of procurement decisions nobody recognized as strategic at the time.
Section 6: What Have We Learned? Seven Pillars
The argument of this paper can be consolidated into seven pillars — five drawn from the structural analysis of Sections 1 through 5, and two added to register the macroeconomic and intellectual stakes that the events of 2026 have made unavoidable.
Pillar 1 — Governments Are Becoming AI Market Makers
The first lesson is that governments are no longer merely referees of the artificial-intelligence economy; they are becoming major customers, lenders, infrastructure providers, standard setters, and demand creators, and the September 2026 evidence is now overdetermined. The Chapel Hill ministerial demonstrated the continuing importance of the regulatory conversation, but Washington’s full-stack export program, Europe’s procurement reform and Cloud and AI Development Act, Taiwan’s semiconductor diplomacy, and China’s substitution mandates all demonstrate something more consequential: governments increasingly shape which technologies scale commercially.[1][6][10][12][7][40] Public money is therefore becoming part of AI architecture. When governments finance datacenters, guarantee exports, purchase cloud services, subsidize fabs, or procure AI applications, they determine not simply who wins a contract but which ecosystems accumulate future economic power — and a policy community that continues to analyze AI through regulation alone will systematically misread where the leverage now sits.
Pillar 2 — The Competitive Unit Is Becoming the Stack
The second lesson is that no single layer sufficiently explains AI power, because the fastest model depends on compute, compute depends on chips, chips depend on fabrication and memory and packaging and networking, datacenters depend on energy, and agents depend on everything beneath them. The Five-Layer AI Economy therefore operates as a single coupled industrial system, and national policy must adapt accordingly: a country possessing excellent models but inadequate electricity is not AI sovereign; a country with datacenters but no secure semiconductor supply remains vulnerable; a country with chips but weak cloud and application layers cannot convert fabrication into economic dominance. The stack — not any individual component — is becoming the meaningful unit of strategic competition, a conclusion that the American export program’s layer-by-layer structure, Huang’s five-layer catechism to the G20, and Europe’s simultaneous legislation across chips, cloud, and procurement all institutionalize from different directions.[32][34][12]
Pillar 3 — Sovereignty Will Mean Managing Dependence, Not Eliminating It
The third lesson is that technological sovereignty must not be confused with self-sufficiency, because even the United States does not possess a nationally isolated AI supply chain — Taiwanese fabrication, Korean memory, Dutch lithography, and Japanese materials remain load-bearing — and no plausible investment program changes that within the decade.[27][8] The correct policy objective is therefore not zero dependence but the prevention of unmanageable dependence: governments should identify which layers must remain under national control, which can safely depend on allies, which require multiple suppliers, and which can remain fully exposed to global competition. Stack Preference gives policymakers a vocabulary and a framework for making those distinctions deliberately — tolerate, hedge, control, layer by layer — instead of discovering their dependencies only when a crisis prices them.
Pillar 4 — Procurement Is Becoming Diplomacy
The fourth lesson is that foreign policy will increasingly be conducted through technology contracts. A datacenter agreement can deepen an alliance; a semiconductor fab can alter a bilateral relationship, as $265 billion of Taiwanese investment in Arizona and a tariff-linked investment framework demonstrate; a cloud contract can create institutional dependency; an export-credit package can determine whether a developing economy builds on American or Chinese infrastructure; and a public-sector AI deployment can establish technical standards that private industry follows for a generation.[9][7] Taiwan’s chip diplomacy illustrates the supply side of the phenomenon; Europe’s purchasing rules show domestic demand becoming strategic policy; America’s export program explicitly fuses commercial technology with diplomacy and public finance.[35][11][5] Procurement is therefore no longer downstream from geopolitics. Procurement is becoming an instrument of geopolitics, and the officials who administer it are acquiring responsibilities their professional training never anticipated.
Pillar 5 — The Winning Stack Will Be the One Countries Choose Voluntarily
The fifth lesson may be the most important for the principals themselves. Governments can force domestic companies to use particular technologies for a period — China is running that experiment at scale — but no state can create a globally dominant ecosystem by coercion alone, because the world’s swing customers are sovereign, hedging, and increasingly sophisticated about exit costs. The strongest stack will combine technological performance with affordability, security, financing, interoperability, developer support, local economic participation, and political trust. The United States consequently faces a challenge larger than staying ahead on frontier benchmarks: it must make the American and allied ecosystem sufficiently attractive that governments building their own AI futures prefer it freely. Europe faces the mirror-image challenge — procurement preference can create domestic demand, but European suppliers must ultimately become globally competitive or the preference becomes a permanent tax on European taxpayers. And China must demonstrate that its increasingly indigenous stack offers capability, reliability, and economic value beyond politically protected home markets. The AI race will be won not only by whoever builds the most intelligence, but by whoever creates the ecosystem the rest of the world most wants to build upon.
Pillar 6 — The Macroeconomics of the Stack Are Now Systemic
The sixth lesson is that stack formation has become a macroeconomic force in its own right, which raises the stakes of every policy discussed in this paper. The IMF’s January 2026 outlook attributed upgraded global growth of 3.3 percent partly to AI investment, credited the buildout with lifting United States growth to a projected 2.4 percent, and simultaneously listed an AI-driven market correction — leverage-funded, hardware-depreciating, expectation-sensitive — among the principal downside risks to the world economy; its scenario-planning work concludes that AI must be treated as a macro-critical transition rather than a standard technology shock.[22][25] Hyperscaler free cash flow is compressing under $725 billion of capital expenditure; Nvidia has begun financially backstopping its own customers’ datacenter construction; memory scarcity is feeding back into the cost of the buildout itself.[18][17] Skeptics supply the necessary discount rate on the enthusiasm — and the most credentialed of them supplies it bluntly, estimating that AI will deliver roughly 0.55 percent in total factor productivity gains over a decade, a fraction of the projections embedded in current valuations, while observing how thoroughly geopolitical rivalry now organizes the American policy conversation.[28]
“The only bipartisan issue in the United States right now is China bashing.”
— Daron Acemoglu, Nobel Laureate, MIT [28]
Whether the optimists or the skeptics prove closer to the truth, the policy consequence is identical: when the infrastructure of a general-purpose technology absorbs nearly a trillion dollars a year of investment, the question of whose stack that money builds is a first-order question of national economic strategy, not a niche concern of technology ministries.
Pillar 7 — The Behavioral and Structural Agendas Must Advance Together
The seventh lesson returns to the distinction with which this paper began. The behavioral agenda — safety testing, independent assessment, incident reporting, misuse prevention — and the structural agenda — Stack Preference — are not rivals, and the same week of September 2026 that produced Europe’s procurement act produced OpenAI’s call for mandatory federal safety regulation, each advancing without reference to the other.[10][14] But they interact, and governments that pursue one while ignoring the other will be surprised by the interaction. Safety rules shape which vendors can credibly serve public-sector deployments, which makes safety regimes de facto procurement filters; procurement choices determine which laboratories’ safety practices govern the systems inside the state, which makes stack selection a de facto safety decision; and international coordination on capability thresholds — which OpenAI itself now advocates — will be negotiated among precisely the governments and vendors that stack competition sets against one another.[14] A mature AI-policy architecture will therefore treat the two agendas as a single portfolio: how the systems behave, and whose systems they are, decided together, by institutions capable of holding both questions at once.
Conclusion: Why “Stack Preference” Fits the New Political Economy of Artificial Intelligence
The history of digital technology encouraged governments to think in categories, and the institutional furniture of every capital still reflects that inheritance. Semiconductors were an industrial-policy issue lodged in economics ministries; electricity was an energy-policy issue; cloud computing was an information-technology issue delegated to chief information officers; artificial-intelligence models were a technology-policy issue assigned to digital ministries and, more recently, safety institutes; cybersecurity was a national-security issue; procurement was an administrative issue supervised by auditors; and trade was an economic-policy issue negotiated by specialists. That categorical separation is becoming incompatible with the architecture of artificial intelligence, because the Five-Layer AI Economy connects everything the categories kept apart: electricity determines datacenter capacity, datacenters determine available compute, semiconductors determine computational capability, models transform compute into intelligence, applications and agents transform intelligence into economic and administrative activity — and governments now influence all five layers simultaneously, whether or not their organizational charts admit it.
The nine days from September 1 to September 9, 2026 offered an unusually compressed demonstration of the transition. In Chapel Hill, the world’s twenty largest economies — the United States and China among them — endorsed an innovation-first governance consensus whose six pillars, from commercialization to standards to supply-chain investment, already reach far beyond conventional safety regulation, while the American Commerce Secretary used the same stage to pitch adoption of the American stack and the chief executive of the world’s most valuable semiconductor company instructed every government present to build AI infrastructure as deliberately as it builds roads.[1][3][4] In Taipei, a diplomatically isolated island converted fabrication capacity into strategic relationships, announcing tens of billions in new allied investment while its president framed semiconductor resilience as shared democratic prosperity.[7][35] In Brussels, the world’s largest single market proposed to rewrite the rules governing fifteen percent of its economy around European preference, days after its cloud-and-AI legislation had already made contribution to the European ecosystem a mandatory criterion in public AI purchasing.[10][12] And in Washington and Beijing, the two principals accused and counter-accused each other over the extraction of model capabilities, demonstrating that even the mathematics of machine learning now carries geopolitical interpretation.[37][38]
China provides the counterpoint that completes the picture: restrictions on foreign technology and mandates for domestic alternatives are accelerating an increasingly independent semiconductor, cloud, and model ecosystem, so that the result of a decade of technological confrontation is not simply U.S.–China competition but competition between increasingly complete systems of production.[40][41][42]
This is why I chose the title Stack Preference. Stack identifies the object being contested — the coupled, five-layer industrial system through which intelligence is manufactured. Preference identifies the mechanism — the weighted, cumulative, rarely prohibitive choices through which public money, law, and diplomacy tilt that system’s development. Together, the two words describe an emerging political economy in which governments are beginning to decide which electricity systems will power artificial intelligence; which chips will process it; which datacenters will host it; which models will produce it; which agents will deploy it; which suppliers will receive public financing; which standards will govern interoperability; and which countries will capture the resulting economic power.
Those decisions will accumulate with a logic this paper has traced repeatedly. A government chooses a datacenter; the datacenter chooses an accelerator ecosystem; the accelerator ecosystem shapes software; software shapes model availability; models shape applications; applications shape workforce skills; workforce skills attract investment; and investment reinforces the original stack. What begins as procurement becomes infrastructure. What begins as infrastructure becomes dependency. What begins as dependency becomes alignment. That is the deeper meaning of Stack Preference, and it is why the concept belongs in the vocabulary of finance ministries and foreign ministries, not only technology ministries.
Between 2027 and 2030, governments will certainly continue regulating artificial intelligence for safety, privacy, competition, cybersecurity, and accountability — the behavioral agenda is intensifying, not receding, as the frontier laboratories themselves now concede. But those debates will increasingly coexist with another question that may prove just as consequential for economic and geopolitical power: whose AI system should our society be built upon? The answer will not be determined solely in laboratories. It will be determined in ministries of finance, commerce departments, governors’ offices, procurement agencies, public utilities, development banks, export-credit institutions, defense departments, semiconductor fabs, and datacenter permitting offices — the unglamorous rooms where the recurring decisions of the state are made, and where the AI civilization of the 2030s is already being purchased, one contract at a time.
The next stage of artificial-intelligence geopolitics will therefore not merely be a competition to invent the best model. It will be a competition to become the preferred stack. And that is precisely why Stack Preference fits this paper.
Footnotes and Endnotes:
[1] The White House — G20 Innovation Ministerial Concludes with Consensus Statement, September 2026. https://www.whitehouse.gov/releases/2026/09/g20-innovation-ministerial-concludes-with-consensus-statement/
[2] Cris Tolomia (Quartz, reporting Bloomberg) — All G20 Nations Endorsed the U.S.’s Hands-Off AI Framework at Chapel Hill Summit, September 3, 2026. https://qz.com/g20-us-light-touch-ai-framework-chapel-hill-090326
[3] CNBC — G20 Tech Takeaways: Lutnick Pitches Adoption of U.S. AI, Pushes Data Center Buildout, September 2, 2026. https://www.cnbc.com/2026/09/02/g20-innovation-ministerial-live-updates.html
[4] IANS / Social News XYZ — Countries Must Build Their Own AI Capacity: Jensen Huang, September 2, 2026. https://www.socialnews.xyz/2026/09/02/countries-must-build-their-own-ai-capacity-jensen-huang/
[5] Donald J. Trump, The White House — Executive Order 14320, Promoting the Export of the American AI Technology Stack, July 23, 2025. https://www.whitehouse.gov/presidential-actions/2025/07/promoting-the-export-of-the-american-ai-technology-stack/
[6] U.S. Department of Commerce, International Trade Administration — Department of Commerce Begins Inaugural Call for Proposals for American AI Exports Program, April 1, 2026. https://www.trade.gov/press-release/department-commerce-begins-inaugural-call-proposals-american-ai-exports-program
[7] Reuters (via Yahoo Finance) — Taiwan Flexes Chip Diplomacy Muscles as It Faces Pressure to Share AI Wealth with Allies, September 7, 2026. https://finance.yahoo.com/technology/ai/articles/taiwan-flexes-chip-diplomacy-muscles-022559044.html
[8] Taiwan Semiconductor Manufacturing Company (TSMC) — Form 6-K: TSMC Reports Second Quarter EPS of NT$27.25, U.S. Securities and Exchange Commission, July 16, 2026. https://www.sec.gov/Archives/edgar/data/0001046179/000104617926000451/a2q26e_withguidancexfinal.htm
[9] CNBC — TSMC to Invest Additional $100 Billion in Arizona After Second-Quarter Profit Soars 77%, July 16, 2026. https://www.cnbc.com/2026/07/16/tsmc-second-quarter-profit-.html
[10] EU Reporter — Commission Proposes Simpler and More Strategic Public Procurement Rules, September 10, 2026. https://www.eureporter.co/politics/european-commission/2026/09/10/commission-proposes-simpler-and-more-strategic-public-procurement-rules/
[11] The Next Web — EU Proposes a Single Procurement Rulebook with a European Preference for Public Buyers, September 2026. https://thenextweb.com/news/eu-public-procurement-act-european-preference-digital-marketplace
[12] European Commission — Proposal for the Cloud and AI Development Act, COM(2026) 502 final, EUR-Lex, June 3, 2026. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52026PC0502
[13] Covington & Burling LLP — The EU Cloud and AI Development Act in Depth, Inside Global Tech, June 11, 2026. https://www.insideglobaltech.com/2026/06/11/the-eu-cloud-and-ai-development-act-in-depth/
[14] Reuters — OpenAI Pushes for Mandatory National AI Safety Rules, September 9, 2026. https://kfgo.com/2026/09/09/openai-pushes-for-mandatory-national-ai-safety-requirements/
[15] NVIDIA Corporation — NVIDIA Announces Financial Results for Second Quarter Fiscal 2027, August 26, 2026. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
[16] NVIDIA Corporation — CFO Commentary on Second Quarter Fiscal 2027 Results, Form 8-K, U.S. Securities and Exchange Commission, August 26, 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27cfocommentary.htm
[17] CNBC — Nvidia Earnings Takeaways: Huang Forecasts 70% Fiscal 2028 Revenue Growth, Far Above Estimates, August 26, 2026. https://www.cnbc.com/2026/08/26/nvidia-nvda-earnings-report-q2-2027-live-updates.html
[18] CNBC — Tech AI Spending Approaches $700 Billion in 2026, Cash Taking Big Hit, February 6, 2026. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html
[19] Yahoo Finance (Goldman Sachs estimates) — Meta, Microsoft, Amazon, and Alphabet Are About to Spend a Shocking Amount of Money to Dominate the AI Era, June 3, 2026. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
[20] Stanford Institute for Human-Centered AI (HAI) — The 2026 AI Index Report, Stanford University, April 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report
[21] Stanford Institute for Human-Centered AI (HAI) — Economy, The 2026 AI Index Report, Stanford University, 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report/economy
[22] Reuters (via The Business Standard) — IMF Sees Steady Global Growth in 2026 as AI Boom Offsets Trade Headwinds, January 19, 2026. https://www.tbsnews.net/world/global-economy/imf-sees-steady-global-growth-2026-ai-boom-offsets-trade-headwinds-1338291
[23] Kristalina Georgieva, International Monetary Fund — Remarks on Leveraging Artificial Intelligence and Enhancing Countries’ Preparedness, World Governments Summit, Dubai, February 3, 2026. https://www.imf.org/en/news/articles/2026/02/03/md-speech-leveraging-artificial-intelligence-and-enhancing-countries-preparedness
[24] PTI (via Kashmir Observer) — IMF: Global Economy Caught Between Oil Shock, AI Boom, August 26, 2026. https://kashmirobserver.net/2026/08/26/imf-global-economy-caught-between-oil-shock-ai-boom/
[25] Karim Barhoumi et al., International Monetary Fund — Global Economic and Financial Implications of Artificial Intelligence: Lessons from a Scenario Planning Exercise, IMF Notes 2026/002, April 3, 2026. https://www.imf.org/en/publications/imf-notes/issues/2026/04/03/global-economic-and-financial-implications-of-artificial-intelligence-lessons-from-a-574924
[26] Carnegie Mellon Institute for Strategy & Technology — Chips and Chokepoints: Chris Miller on the Geopolitics of the AI Supply Chain, Carnegie Mellon University, March 2026. https://www.cmu.edu/cmist/news-archive/news/2026/march/chips-and-chokepoints-chris-miller-on-the-geopolitics-of-the-ai-supply-chain.html
[27] Chris Miller (via Silicon Republic) — How the Chip War and the Future of AI Are Inextricably Linked, Silicon Republic. https://www.siliconrepublic.com/future-human/chip-war-semiconductors-supply-tech-geopolitics-chris-miller
[28] Fortune — Nobel Laureate Daron Acemoglu on the ‘Brainless’ AI Discourse, the Myth of Capitalism and the Gen Z Revolution Risk, June 21, 2026. https://fortune.com/2026/06/21/nobel-laureate-daron-acemoglu-ai-productivity-capitalism-democracy/
[29] Mariana Mazzucato — Making Industrial Strategy Great Again, Foreign Affairs, February 2026. https://www.foreignaffairs.com/united-states/making-industrial-strategy-great-again
[30] Mariana Mazzucato — Policy with a Purpose: Modern Industrial Policy Should Shape Markets, Not Just Fix Their Failures, IMF Finance & Development, September 2024. https://www.imf.org/en/publications/fandd/issues/2024/09/policy-with-a-purpose-mazzucato
[31] U.S. Department of Commerce, International Trade Administration — Department of Commerce Announces New American AI Exports Program Phase, March 16, 2026. https://www.trade.gov/press-release/department-commerce-announces-new-american-ai-exports-program-phase
[32] Federal Register — American AI Exports Program; Call for Proposals for Pre-Set Consortia, April 10, 2026. https://www.federalregister.gov/documents/2026/04/10/2026-06952/american-ai-exports-program-call-for-proposals-for-pre-set-consortia
[33] NVIDIA Blog — ‘Largest Infrastructure Buildout in Human History’: Jensen Huang on AI’s ‘Five-Layer Cake’ at Davos, January 2026. https://blogs.nvidia.com/blog/davos-wef-blackrock-ceo-larry-fink-jensen-huang/
[34] CNBC Africa — Nvidia’s Huang Pushes for AI Infrastructure in Every Country, September 2026. https://www.cnbcafrica.com/2026/nvidias-huang-pushes-for-ai-infrastructure-in-every-country
[35] StratNews Global — Taiwan Chip Diplomacy Strengthens Global Partnerships, September 7, 2026. https://stratnewsglobal.com/technology/taiwan-chip-diplomacy-global-partnerships/
[36] Reuters (via BusinessWorld) — Taiwan Flexes Chip Diplomacy Muscles as It Faces Pressure to Share AI Wealth with Allies — EU Wants Investment Too, September 7, 2026. https://bworldonline.com/world/2026/09/07/775095/taiwan-flexes-chip-diplomacy-muscles-as-it-faces-pressure-to-share-ai-wealth-with-allies/
[37] CNN — US Claims Chinese AI Firms Are Carrying Out ‘Industrial-Scale’ Theft of Trade Secrets, September 8, 2026. https://edition.cnn.com/2026/09/08/politics/us-accuses-china-of-stealing-ai-technology
[38] Associated Press (via KSAT) — China Hits Back at US Claims of ‘Malicious’ AI Distillation Ahead of Planned Trump-Xi Talks, September 9, 2026. https://www.ksat.com/business/2026/09/09/china-hits-back-at-us-claims-of-malicious-ai-distillation-ahead-of-planned-trump-xi-talks/
[39] The Washington Times — China Hits Back at U.S. Claims of ‘Malicious’ AI Distillation (Remarks of Treasury Secretary Scott Bessent), September 9, 2026. https://www.washingtontimes.com/news/2026/sep/9/china-hitting-back-us-claims-malicious-ai-distillation-ahead-planned/
[40] Reuters (via Yahoo News) — China Bans Foreign AI Chips from State-Funded Data Centres, 2025. https://finance.yahoo.com/news/china-bans-foreign-ai-chips-083739738.html
[41] Mercator Institute for China Studies (MERICS) — Domestic Substitution in AI Chips: China’s Big Gamble, December 2025. https://merics.org/en/comment/domestic-substitution-ai-chips-chinas-big-gamble
[42] Bloomberg — DeepSeek Plans Big Huawei AI Chip Order to Power New Data Center, September 4, 2026. https://www.bloomberg.com/news/articles/2026-09-04/deepseek-plans-big-huawei-ai-chip-order-to-power-new-data-center
[43] Brookings Institution — Is AI Sovereignty Possible? Balancing Autonomy and Interdependence, February 2026. https://www.brookings.edu/wp-content/uploads/2026/02/20260217_AI_sovereignty_final.pdf
[44] Atlantic Council — Eight Ways AI Will Shape Geopolitics in 2026, January 15, 2026. https://www.atlanticcouncil.org/dispatches/eight-ways-ai-will-shape-geopolitics-in-2026/
[45] Boston Consulting Group — For Most Countries, AI Sovereignty Is an Illusion. Resilience Is Real., March 25, 2026. https://www.bcg.com/publications/2026/ai-sovereignty-is-an-illusion-resilience-is-real
[46] Jorpex — The EU Procurement Reform 2026, and the ‘European Preference’, August 2026. https://jorpex.com/guides/eu-procurement-reform-2026/
[47] Computer & Communications Industry Association (CCIA Europe) — Public Procurement Act: CCIA Europe Reacts to Proposed EU Rules, September 2026. https://ccianet.org/news/2026/09/public-procurement-act-ccia-europe-reacts-to-proposed-eu-rules
[48] U.S. Department of Commerce — G20 Innovation Ministerial (Statement of Secretary Howard Lutnick), September 2026. https://www.commerce.gov/news/blog/2026/09/g20-innovation-ministerial
[49] Stanford Institute for Human-Centered AI (HAI) — Inside the AI Index: 12 Takeaways from the 2026 Report (James Landay on Open-Weight Models), August 2026. https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report
[50] GPUSmith — Hyperscaler AI Capex 2026: Amazon, Microsoft, Google, Meta, July 2026. https://gpusmith.com/articles/en/hyperscaler-ai-capex-2026
[51] Eubelius — The Cloud and AI Development Act (CADA): Public Procurement Sets the Tone as a Strategic Tool to Increase EU Tech Sovereignty, June 15, 2026. https://www.eubelius.com/en/news/the-cloud-and-ai-development-act-cada-public-procurement-sets-the-tone-as-a-strategic-tool-to
[52] The European Times — EU Unveils Public Procurement Act (European Court of Auditors Finding on Declining Competition), September 2026. https://europeantimes.news/2026/09/eu-unveils-public-procurement-act/
[53] Big Hat Group — China AI Weekly: DeepSeek’s 160K Huawei Chip Megaproject, September 6, 2026. https://www.bighatgroup.com/blog/china-ai-weekly-2026-09-06/[54] Futurum Group — AI Capex 2026: The $690B Infrastructure Sprint, February 12, 2026. https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
