A company can run the same model in Virginia, Vienna and Shenzhen. It will still face materially different operating conditions.
The difference starts below the interface. It includes who supplies the compute, which data can move, how a model reaches the market, what evidence a provider must retain, and who can stop the system when it causes harm. Governments are shaping each of those layers now.
I use stack as shorthand for that joined system: compute, models, data, rules, institutions and routes to market. It is an analytical map. There are no clean borders, and the systems still trade, borrow standards and share research. But their default settings are moving apart.
This article compares three centres that already shape widely used defaults. The United States is organising around capability, private capital and the export of American technology. The European Union is using rules, interoperability and public investment to reduce dependency without closing itself off. China is coordinating industrial development, public-service controls and domestic capability under a single state strategy.
Other countries combine parts of these approaches, as the UK and India examples show.
Consider one provider trying to offer the same public generative service across all three. Policy instruments in each system have affected what providers could deploy and which evidence they had to preserve. The model may be identical while rules, infrastructure and authority differ by jurisdiction.
This comparison is current to 2 October 2026 and describes structural direction, not legal applicability. Exact licensing, transition, filing and security-assessment requirements can change and must be checked against current official text for the provider, system and use case.
The model layer gets most of the attention. The surrounding system will decide how much of the model’s power can actually be used.
The American stack: scale first
The American system begins with private firms that can spend at a scale few states can match. Frontier labs, cloud providers, chip designers and capital markets reinforce one another. The result is fast capability growth and concentrated control over the infrastructure needed to train and serve large models.
Federal policy now makes that structure explicit. The White House’s 2025 AI Action Plan is organised around innovation, infrastructure and international leadership. A companion executive order calls for full-stack export packages covering hardware, models, software, applications and standards.
This is more than support for a successful industry. It is an attempt to make American infrastructure the default substrate for AI elsewhere.
The governance around it is less unified. Federal agencies apply existing authority in their own domains. States add rules of their own. The administration itself described the result as a patchwork when it proposed a national legislative framework in March 2026. A proposed framework is not a single enacted federal regime. The fragmentation remains part of the stack.
Export controls have formed a harder boundary. United States measures have restricted advanced compute, particular end users and diversion routes at different times. Their scope has changed with security and trade policy. This article does not state what currently requires a licence; a deployment decision needs a current case-specific check.
The American stack therefore combines speed with uncertainty. Firms can build quickly inside a deep commercial ecosystem. They also face changing state rules, sector-specific enforcement and strategic controls that can alter which customers and regions remain accessible.
Its central contradiction is concentration. The system celebrates open competition while the cost of frontier compute pushes power toward a small group of providers. An application can change model suppliers. Moving its data, evaluations, permissions and operating history is harder.
The European stack: rules-led sovereignty
Europe starts from a different question. What conditions should apply before a system enters the market or affects a person’s rights?
The AI Act supplies the most visible answer, but the European stack is wider than one regulation. Data protection, competition rules, platform duties, cloud switching, digital identity and public compute investment all shape the environment in which an AI system operates.
The 2024 AI Act established a risk-based framework, and a later official amendment changed parts of its implementation schedule. This article does not state the current dates or decide which provision applies to a reader’s system. Teams need to verify the operative text, transition rule, role and use case.
That change exposes the European tension. Rights and safety remain the declared foundation. Policymakers also know that compliance cost, limited compute and slower capital formation can weaken the firms expected to comply.
Sovereignty is the response, although the word is often misunderstood outside Europe. It does not automatically mean enclosure. The EU’s Data Union strategy couples sovereignty with cross-border exchange and access for trusted partners. The Data Act promotes interoperability and easier switching between data-processing services. The European Digital Identity framework aims for wallets that can be recognised across Member States without creating a single central identity database.
The infrastructure programme follows the same logic. EuroHPC AI Factories and planned gigafactories are intended to give researchers, startups and industry more access to European compute. Nineteen AI Factories show real public commitment. Planned capacity is still planned capacity, and European systems continue to depend on hardware and cloud technology produced elsewhere.
The European stack is therefore rules-led and sovereignty-seeking, but still open by design and dependent in practice. Its power comes from market access. A company that wants European customers must often change how it documents, tests and operates a system. Its weakness appears when those requirements arrive faster than the infrastructure and capital needed to build European alternatives.
The Chinese stack: coordinated development and control
China joins industrial policy and operational control more directly.
Published in 2023, China’s official Interim Measures for Generative AI Services described provider, data and content duties for public-facing services. That is a historical description of the instrument, not a claim about current applicability. Any present filing or security-assessment requirement needs to be checked against the current official text.
The industrial side is equally important. The State Council’s 2025 AI Plus guideline calls for stronger model foundations, intelligent compute, data supply, domestic innovation and shared infrastructure. It also promotes open-source ecosystems and international cooperation.
Development and security have appeared in the same policy frame. Chinese policy documents published in 2025 treated synthetic-content provenance as a governance topic rather than only a voluntary product feature. This article does not assert their current effective duties; those require a current official-text check.
Calling this a sealed stack would be convenient and wrong. Chinese policy supports domestic chips, frameworks, models and data infrastructure, but policy ambition is not self-sufficiency. Foreign hardware still matters. Open models cross borders. Research and supply chains continue to connect the system to the rest of the world.
The tighter description is state-coordinated and increasingly domestically oriented. Access can be constrained at both the service and infrastructure layers. At the same time, open-source models and international technical work give Chinese systems routes outward.
Its central contradiction is familiar. A state can coordinate investment and adoption at speed. It can also make the boundary between safety, political control and commercial permission difficult for outsiders to separate. That ambiguity becomes part of the operating cost.
The map has more than three colours
The United Kingdom and India show why the taxonomy needs soft edges.
The UK combines sectoral regulation with unusually deep work on frontier-model evaluation. Its AI Security Institute evaluates advanced systems and researches severe risks. The country is also investing in public compute and pursuing greater sovereign capacity. Yet its own Compute Roadmap describes a fully British chip-to-software stack as an ideal end state, not present reality. British labs, regulators and infrastructure remain closely integrated with allies and global providers.
India is building a different kind of public layer. The IndiaAI Mission offers subsidised common compute to researchers, startups, universities, smaller firms and government. It also funds Indian-language and multimodal foundation models, with an expectation that supported work contributes to the open-source ecosystem.
This is serious infrastructure policy. It is also a mixed system. Much of the compute is provided through commercial partners, and advanced accelerators are still produced abroad. India is building sovereign capability through global hardware rather than waiting for a completely domestic supply chain.
Neither country fits neatly inside one of the three larger systems. That is the point. National AI strategies can borrow American infrastructure, European rules, Chinese open models and domestic public investment at the same time.
The edges are where the strategy lives
A multinational organisation does not choose one stack once. It reconciles them operation by operation.
A model may be trained under one legal regime, hosted by a provider headquartered in another, adapted with local data in a third, and used to make decisions about people in all three. Identity credentials, evaluation results and incident records may not carry the same legal meaning at every step. A feature that is permitted in one market may require disclosure, human review or a different data flow in another.
That friction creates cost. It also creates bargaining power.
Interoperability rules can make it easier to leave a provider. Export packages can pull an allied country toward one technical ecosystem. Shared evaluations can turn a national research institution into an international standard setter. Subsidised compute can give local firms room to build models for languages and public services that global providers underserve.
The likely outcome is partial alignment, not clean separation. Trusted partners will exchange technology. Firms will maintain regional variants. Standards will converge in some layers and split in others. Compute, identity, content provenance and model evaluation may each follow a different political map.
The evidence in this comparison supports an expectation of regional and layer-specific divergence. Broad adoption of portable identity, evidence and market-access rules across these regions would overturn it. The phrase “AI regulation” is still too small for the current contest, which covers infrastructure, market access, institutional authority and the evidence a system must produce while it operates.
The model can travel. The authority around it does not travel cleanly.