Part III - The Architecture of the Global AI Stack

Note

Protocol Note — v0.9 Release: Part III is presented in compressed form; full-length development of each component will appear in v1.0.

Part II established that law and accounting — the State’s dual governance languages — can no longer classify or measure AI with the precision their institutional authority requires. The Translation Layer has filled that gap by default. But the Translation Layer does not operate in isolation. It operates within a material structure — the AI governance stack — through which legal texts, economic grammars, institutional intermediaries, and computational infrastructure interact to produce operational outcomes.

Part III maps that structure. It traces the shift from horizontal power, organised through territory, to vertical power, organised through the stack — and examines how that vertical architecture concentrates around three blocs whose definitions of economic reality are diverging.

I. From Architecture to Implementation

The industrial world was organised horizontally. Power radiated outward from capitals toward borders. Jurisdiction was the unit of competition, and law and accounting were the paired languages through which states coordinated that competition.

AI shifts that geometry. A model can be trained in one country, deployed in another, and enforced through infrastructure that spans several more. A semiconductor does not register
borders; a data centre routes workloads according to latency, not allegiance. A content-moderation rule written in Brussels shapes deployments in Oregon, Abu Dhabi, and Jakarta — its effects propagating vertically through the stack rather than horizontally across territory.

Authority now operates through this vertical architecture — the AI governance stack — a layered system in which legal texts, economic grammars, translation institutions, and computational infrastructure interact to produce operational results. Territorial law still matters, but its practical effect depends on how it is compiled into procedures, protocols, and technical controls.

Power no longer radiates outward; it descends. A requirement becomes a protocol, a protocol becomes a process, and a process becomes a constraint executed in silicon. In this architecture, authority is exercised not through enactment alone but through the design of the systems that implement it.

II. The AI Governance Machine

The stack has four layers. Law defines the permission language. Accounting defines the value grammar. Translation institutions turn both into protocols, templates, and controls. Compute
executes the result in silicon.

In the industrial era, these layers were loosely coupled. In the AI era, they interlock. A change in one layer propagates through the others. The layers no longer operate as parallel domains; they function as sequential stages in a single compilation pathway.

A statute, standard, or policy does not bind AI directly. Its practical force emerges when language is compiled into structure — when a legal text becomes an audit protocol, a disclosure rule becomes a reporting requirement, and a liability regime becomes a runtime constraint. The transformation is compilational: ambiguity is progressively reduced into operational determinacy, even as each layer introduces its own discretionary choices. The overall movement is downward — toward fewer degrees of freedom.

Symbols become form, form becomes process, and process becomes constraints enforced by the machinery itself.

A conformity requirement under the EU AI Act illustrates the pathway. Statutory text becomes an audit checklist; the checklist becomes a compliance template produced by major assurance firms; the template becomes an automated documentation pipeline; the pipeline becomes a model-level control. The pattern generalises: regulation becomes decisive at the point of execution — when the constraint reaches silicon.

This is the AI governance stack — a vertical system in which law, accounting, institutions, and compute operate as a layered compilation process.

III. The Arrival of Agents

AI is shifting from systems that generate predictions to systems that initiate actions. This is the agentic shift — and it is determinative, because agency reallocates decision rights.

A predictive model can be governed through disclosure, licensing, and risk controls. An AI agent requires constraints embedded in the system’s tools, permissions, and execution pathways — controls that operate at runtime rather than oversight applied after the fact.

As agents proliferate, models become actors, outputs become decisions, errors become harms, and capabilities become power. A demand-forecasting system becomes a supply-chain negotiator. A transaction classifier becomes an executor. A route optimiser becomes a logistics controller. These are not predictions for human review; they are actions that reshape economic life directly.

The agentic shift drives governance downward into the stack. Statutes cannot bind an agent at runtime. Effective control depends on the protocols, templates, and machine-level constraints that monitor and shape an agent’s behaviour — regulation compiled into the same stack through which the agent operates.

IV. The Geopolitics of Vertical Integration

Vertical integration in AI produces a new geopolitical hierarchy — defined not by ideology but by control over the technical and institutional layers through which intelligence operates.

A small number of states and firms now command the pathway from data extraction to model execution, from semiconductor fabrication to the rulemaking machinery that governs deployment. Two countries operate the most complete vertically integrated ecosystems at global scale: the U.S. and China. The EU occupies a structurally different position.^[1]

The U.S. leads through gravitational pull. Its ecosystem combines permissionless innovation, private-sector primacy, capital-driven scaling, and a Translation Layer largely outsourced to
the market. The vertical stack spans domestic chip design, hyperscale compute, frontier model and agentic system development, and the deepest capital markets financing global AI diffusion. The U.S. does not govern AI through a single comprehensive regime; it shapes the global ecosystem through the structural dominance of its platforms, capital markets, and infrastructure — creating a default environment that foreign firms absorb simply by building on it.^[2]

China leads through a different mechanism: policy coherence. Its ecosystem combines state-directed integration, vertically aligned industrial policy, and a regulatory framework fused with political authority. The vertical stack spans a growing domestic semiconductor supply chain, a national energy and compute network, state-aligned data regimes, and coordinated deployment pipelines. China aligns energy, technology, industry, and infrastructure into a single strategic direction — creating a system in which AI governance is not layered onto the economy but embedded within it.^[3]

The EU operates through neither gravitational pull nor policy coherence. It leads through rule projection. It does not operate a fully integrated AI stack; it operates a law-led ecosystem
built from regulation outward rather than silicon downward. Its leverage derives from the ability to define the constraints under which AI built elsewhere must operate — through standards, liability regimes, and disclosure mandates that propagate extraterritorially across supply chains. This creates a distinctive structural position: significant normative authority that depends, in part, on continued access to the compute and models produced by the other two blocs.^[4]

Europe is not without sovereign compute ambitions. EuroHPC, national exascale programmes, and emerging public–private compute alliances represent material investments in domestic capacity and reduced dependence on external hyperscalers. Europe nonetheless remains less concentrated in scale compute and frontier model clusters than the United States or China — a position that continues to evolve as new investments come online. Europe’s comparative strength lies less in compute concentration than in regulatory projection and institutional coherence. The G3 divergence is one of structural orientation, not static deficit.^[5]

These are operational distinctions, not moral ones. Each bloc exercises authority through a different mode — gravitational pull, policy coherence, and rule projection — and each produces a different form of dependency for non-G3 countries. A country building on U.S. infrastructure absorbs market-driven norms. A country integrated into Chinese supply chains absorbs state-directed requirements. A country trading with the EU absorbs compliance obligations. No non-G3 country can engage with AI without navigating all three — and the terms of engagement are set by the blocs, not by the countries that depend on them.

V. The Vertical Geometry of Global Influence

The G3 do not merely compete over AI — they define the terms under which the rest of the world engages with it. Compute becomes a strategic asset, standards become instruments of
projection, translation institutions become geopolitical actors, and the stack becomes the terrain on which global competition is conducted.

This creates a new dependency structure. Third-country firms building on U.S. or Chinese
AI foundations inherit constraints they did not author — terms of access that can be redefined at any moment through export controls, sanctions, or platform-policy changes. Countries that control none of the stack’s critical layers rely on those who do. Territory still matters, but it no longer alone determines who sets the terms. The stack’s control points do.

National autonomy in the AI era is not declared but constructed — encoded in technical standards, operationalised through private firms, and anchored in the computational physics
that shape access and control. The question for every non-G3 country is not whether to engage with AI, but on whose terms — and at what cost to sovereign discretion. Any viable response must operate at the level of machine execution, where behaviour is actually produced, constrained, and steered. Part V returns to this challenge directly, asking whether a governance architecture adequate to these constraints can be built.

This diagnosis sets the stage for the fracture that follows: the point at which the G3 no longer share a common definition of economic reality.

VI. The G3 Fracture

Vertical AI power now intersects with the legal, accounting, and classificatory fractures examined in Part II. The U.S., China, and the EU no longer interpret AI through a shared conceptual frame. Each operates a distinct definition of economic reality, encoded in a different AI worldview.^[6]

The U.S. projects a GAAP-aligned, market-driven valuation framework: venture-capitalised,
Level 3 fair-value oriented, and optimised for capital formation. Foreign firms building on
U.S. AI foundations absorb this financial worldview by default — a translation that shapes how their models are priced, deployed, and scaled.

China embeds a CAS-aligned, state-directed industrial model: integrated into national planning and oriented toward long-horizon capability accumulation. Foreign firms leveraging Chinese AI supply chains face different requirements: localised data control, security obligations, and algorithmic-transparency standards that may conflict with commitments under other G3 regimes.

The EU codifies an IFRS-aligned, law-led governance regime: prudence-constrained, disclosure-intensive, and normatively anchored. Its compliance obligations reach foreign AI deployments extraterritorially, even as its regulatory authority rests on access to compute and models it does not produce.

The same model. The same weights. The same compute. Three distinct realities — because value, risk, and agency are stack-relative properties. A frontier model trained in California,
fine-tuned in Shenzhen, and deployed in Frankfurt is not a single asset moving across borders. It becomes three different objects: a financial instrument, a strategic capability, and a regulated risk — depending on the interpretive frame through which it is read.

What is emerging is not a policy divergence but a fracture in the underlying logics that define value, risk, and agency. As these foundations separate, the consequences compound: supply chains that cannot be synchronised, liability regimes that resist reconciliation, measurement systems that refuse normalisation, and constraints on AI agents that defy interoperability.

These are not disagreements over regulatory detail. They are collisions between incompatible accounts of what AI is — three ecosystems offering different answers to what machine intelligence constitutes, how it should be evaluated, and who bears responsibility for the actions of AI agents. The fracture widens not because nations disagree on the need for AI governance, but because they begin from different assumptions about the nature of AI itself.

VII. The Offramp from Dependency

The fracture mapped in this Part forces a structural choice on every non-G3 jurisdiction. No country can engage with AI without inheriting dependencies — on U.S. platforms, Chinese supply chains, or EU compliance regimes, and often on all three simultaneously. The question is not whether to accept external constraints but how to manage them: align with one stack, borrow selectively from several, or construct an independent hybrid.

Part IV turns to the mechanisms that mediate across this divide. It begins with the institutional infrastructure of translation — examining Hong Kong as a reference implementation of how
incompatible governance stacks currently communicate — and then develops the Incorporation Heuristic: an analytical model describing which rules propagate into global practice and why, as a function of visibility, workability, and necessity. Part IV also introduces the Sovereignty Paradox, the Gateway Rules framework, and the AI Power Hierarchy as structural consequences of translation under conditions of opacity and national competition. Part III identified where power resides in the AI stack. Part IV examines how that power is exercised — and where agency remains for countries navigating between the dominant blocs.

1. U.S. CHIPS and Science Act (2022); EU Chips Act (2023). China's semiconductor self-sufficiency and national compute-network initiatives reflect parallel strategic responses.↩︎
2. Empirical data on U.S. cloud-market concentration, GPU design leadership, and AI venture-capital flows will be documented in the v1.0 edition.↩︎
3. New Generation Artificial Intelligence Development Plan (2017); 15th Five-Year Plan for National Economic and Social Development (2026–2030), adopted March 2026. Subsequent compute-infrastructure and industrial-policy initiatives will be documented in the v1.0 edition.↩︎
4. A. Bradford, The Brussels Effect (2020); Regulation (EU) 2024/1689 (EU AI Act).↩︎
5. Council Regulation (EU) 2021/1173, establishing the European High Performance Computing Joint Undertaking (EuroHPC JU).↩︎
6. The G3 fracture analysis is an original analytical contribution of this manuscript, building on the G2/G3 categories introduced in Part I.↩︎