The AI Stack and the State

The AI Stack and the State

I. Compute as Territory

AI training compute has become the central terrain of geopolitical power. Nations rise or fall based on their access to advanced chips and the electricity required to use them at scale. This is not a soft shift in economic priorities. It is a material transformation in the structure of influence.

The United States has tightened its grip on this emerging order through a series of export controls that restrict the flow of high-performance GPUs and semiconductor tools to strategic competitors. These policies shape the pace of global AI development and force rivals into slower or more costly pathways. China has felt this pressure most sharply, as access to advanced chips remains essential for training large-scale models and sustaining competitive research. The impact is compounded by the world’s dependence on Taiwan for advanced fabrication. TSMC remains the producer of nearly all cutting-edge chips, and its location anchors an entire global hierarchy in a narrow band of territory vulnerable to disruption.

The energy demands of frontier AI training deepen this divide. Large models require staggering amounts of electricity, and the projected growth of data center power consumption in the coming decade will favor states with abundant, stable, and inexpensive energy sources. Nations that lack these conditions will remain subordinated in the coming order, regardless of talent or ambition.

Compute is no longer a technical resource. It is a form of strategic territory, and the nations that control it will command the direction of technological power for decades.

II. The New Lords of Silicon

The commanding heights of the AI age are controlled by a narrow circle of firms whose capabilities exceed those of many states. Their fabs, datacenters, and proprietary architectures form the spine of global compute, concentrating power in a way reminiscent of earlier epochs when a few actors controlled entire resource chains. Nvidia, TSMC, ASML, and the major American cloud providers occupy this position today. Their influence is not symbolic. It is operational.

Nvidia shapes the tempo of AI development through its near monopoly on training chips. Delays in its supply chain ripple across research labs, startups, and national strategies alike. This dominance persists because TSMC remains the only foundry capable of fabricating the most advanced chips at commercial scale. ASML reinforces this hierarchy through its exclusive hold on EUV lithography machines, equipment without which cutting-edge fabrication is impossible. Together, these three actors form a strategic triangle that defines who can enter the frontier model race and who must remain downstream.

Cloud giants extend this hierarchy. Their hyperscale datacenters, distributed energy agreements, and global fiber networks give them the ability to host and train models at a scale few governments can match. These firms determine which countries receive early access to advanced compute clusters and which are offered limited or delayed capacity. The decisions are commercial on the surface but geopolitical in consequence.

Control of compute has consolidated into a small elite. Their choices, investments, and alignments shape the global distribution of intelligence itself.

III. The Vassal States

Beneath the small circle of compute powers lies a wide tier of states that possess technical talent, functional research ecosystems, and ambitious digital agendas yet lack the hardware and energy needed to train frontier-scale models. Their position resembles technological vassalage. They can participate in the AI economy, but only through channels controlled by stronger actors.

Countries such as Singapore, South Korea, Israel, and Canada occupy the upper end of this structure. They can host significant datacenters, attract cloud investment, and fine-tune or deploy advanced models. Yet they cannot produce or train frontier systems without relying on American hardware, Taiwanese fabrication, or access granted by U.S. cloud providers. This reliance curtails autonomy. It shapes national AI strategies around what is available rather than what is desired.

The next tier is larger and more constrained. Dozens of states in Latin America, the Middle East, Africa, and Eastern Europe depend entirely on access to foreign cloud infrastructure. Their AI industries operate through API calls to firms headquartered elsewhere. Their domestic models rarely exceed moderate scale. Their digital ecosystems are vulnerable to price shifts, policy changes, or export updates enacted by distant powers. They possess talent. They lack sovereignty.

This structure does not reflect intellectual weakness. It reflects material scarcity. States can educate engineers and build research labs, but without fabrication capacity, energy abundance, or privileged cloud access, they cannot escape dependency.

A hierarchy built on compute becomes self-reinforcing. Those without early access struggle to gain it later.

IV. Energy as Destiny

AI training requires a material foundation far more demanding than the digital rhetoric surrounding it suggests. Chips alone do not determine capability. Without vast quantities of reliable electricity, frontier models cannot be trained, deployed, or scaled. Energy capacity is therefore beginning to function as a geopolitical filter, separating states that can sustain AI development from those that cannot.

The projected rise in global data center electricity demand highlights this shift. The International Energy Agency forecasts that by 2030, the energy footprint of data centers could approach the consumption of medium-sized industrialized nations. AI workloads account for much of this trajectory. The concentration of hyperscale facilities in the United States and China reflects their unique ability to allocate land, water, and steady power to these projects. Both countries have already begun restructuring regional grids to support long-duration, high-density compute clusters.

Regions without abundant or inexpensive energy struggle to compete. Europe faces mounting challenges due to high energy costs, grid constraints, and environmental regulations that complicate datacenter expansion. States with intermittent or fragile power supplies find themselves locked out entirely. Their limitation is structural. Even with skilled engineers or strong policy ambitions, the absence of stable energy prevents the formation of competitive AI infrastructure.

This emerging divide is not a temporary imbalance. It is a durable pattern rooted in geography, resources, and grid resilience. Energy determines whether a state can host compute at the scale required for sovereign AI development. Those who lack it remain dependent on actors who do.

V. The End of Open Access

The rhetoric of an open digital world is giving way to a closed architecture defined by scarcity, control, and strategic caution. Frontier AI models are no longer treated as public goods. They are guarded assets. States and firms alike recognize that the capabilities embedded in these systems influence military planning, intelligence collection, industrial policy, and domestic governance. Access is therefore rationed.

The United States has already begun regulating frontier model development as a matter of national security. Reporting requirements tied to training runs, compute thresholds, and model capabilities reflect a strategic shift toward viewing advanced AI as a dual-use technology. These measures parallel the export controls placed on advanced chips. The logic is the same. When a capability shapes national power, it becomes subject to oversight.

Meanwhile, major AI firms are retreating from full transparency. Weight releases have become selective. Training data disclosures are limited or absent. Safety evaluations are published in broad strokes rather than detailed technical form. Even organizations committed to openness have adjusted their posture as model capabilities have grown more powerful. The shift is not ideological. It is structural. Strategic assets are never fully open once their geopolitical value becomes clear.

For states without sovereign compute, this trend closes the path to parity. API access provides utility but not autonomy. Fine-tuning enables adaptation but not independence. As models become more capable, the threshold for meaningful participation rises faster than most countries can respond.

The era of open access has ended. The strategic era has begun.

VI. The Rise of AI Tributaries

As frontier compute concentrates in a few centers of power, a second tier of actors has emerged. These are the AI tributaries. They develop applications, fine-tune models, and adapt systems to local needs, but remain dependent on upstream providers for access to the base architectures. Their role resembles that of historical tributary states that exercised autonomy within limits defined by stronger powers.

Startups across Southeast Asia, Latin America, Eastern Europe, and the Middle East increasingly operate in this position. Their capabilities depend on cloud credits, licensing agreements, and API quotas set by American and Chinese firms. When access is available, they can build competitive products for local markets. When access tightens, innovation stalls. Their strategic horizon is determined elsewhere.

Governments in these regions have begun formalizing relationships with the major cloud and model providers. Some negotiate preferential access through infrastructure partnerships. Others pursue model-sovereignty initiatives, seeking to build domestic alternatives. Yet the gap between ambition and material capacity remains wide. Training frontier models requires silicon supply chains that span continents, datacenters that draw power like industrial complexes, and research teams able to coordinate at global scale. Few tributary states possess these conditions.

This intermediate position brings opportunity but also dependency. Tributaries may grow vibrant AI ecosystems, but they cannot escape the gravitational pull of upstream compute powers. Their policy choices, corporate strategies, and technological futures flow through the decisions of firms and governments located far beyond their borders.

AI has created a layered order. Tributaries inhabit the middle tier, productive yet constrained.

VII. Strategic Implications

The consolidation of compute power produces consequences that reach far beyond the technology sector. Intelligence services, defense ministries, and economic planners increasingly operate within an environment shaped by a small number of actors who control the means of large-scale model training. This arrangement transforms AI from a general-purpose technology into a strategic asset with direct influence on national decision-making. States that command frontier compute are positioned to guide global standards, dominate defense applications, and shape the digital infrastructure of allied and dependent nations.

Military planning reflects this shift. Advanced models improve target identification, intelligence synthesis, and operational logistics. Nations lacking sovereign compute must rely on foreign systems that may be constrained or withheld during crises. This asymmetry alters deterrence calculus. Strategic autonomy shrinks when the tools that interpret the battlespace are imported. Economic competitiveness follows the same logic. Industries adopting AI at scale gain decisive efficiencies, while lagging states face widening productivity gaps that compound over time.

Diplomacy is adapting. Access to compute has become a lever in alliance formation. States negotiate cloud-region placements, hardware allocations, and export exemptions as part of broader geopolitical alignments. Those unable to secure dependable access risk marginalization within emerging digital blocs. The concentration of compute power also increases systemic vulnerability. A disruption at one fabrication node or a targeted attack on a major datacenter can ripple across global markets.

Control of compute now shapes the distribution of power in every sector that depends on intelligence. The consequences are structural and enduring.

VIII. Toward a New Digital Order

The global hierarchy emerging around compute resembles an older world, one structured by asymmetry, strategic dependency, and the material limits of power. The states that hold advanced chips, fabrication capabilities, energy abundance, and hyperscale datacenters define the outer boundaries of the future. Everyone else maneuvers within those boundaries. This order is not enforced by occupation or coercion. It is enforced by scarcity. Power accumulates where compute resides.

The pattern is self-reinforcing. States with early advantages build larger research ecosystems, attract greater investment, and refine national strategies that integrate AI into defense, economic planning, and diplomacy. Their lead compounds with every training cycle. States in subordinate positions confront rising barriers to entry. Even ambitious national AI programs become constrained by hardware availability, cost of power, and reliance on foreign cloud infrastructure. Sovereignty becomes conditional rather than absolute.

Yet this structure is not immutable. New fabrication nodes may emerge, energy grids may expand, and regional alliances may pool resources to build shared compute capacity. Several states are already exploring cooperative datacenter agreements and cross-border AI infrastructure compacts. These efforts reflect a recognition that autonomy in the digital age requires collective investment at scales unreachable for most nations individually.

The central reality remains: compute has become the foundation of geopolitical order. Nations that understand this early will shape the architecture of the century. Those that delay will adapt to systems designed by others. The contest over digital sovereignty has begun, and its outcome will define the strategic landscape for decades.