Localised AI Models and Centralised Diffusion

Executive Summary

On June 12, 2026, the US government issued an export control directive requiring the immediate suspension of access to Anthropic's Fable 5 and Mythos 5 models for all foreign nationals, everywhere. The stated reason was a narrow jailbreak technique. The mechanism was simple: disable the API globally, revoking access from hundreds of millions of users based on a single jurisdiction's regulatory decision.

This event demonstrates centralised diffusion—the emerging architecture of AI access in which models are bound to jurisdictions, deployable only within defined borders, and subject to control by regulatory bodies.

Centralised diffusion is distinct from the "international diffusion" framework that has dominated policy discussion since 2023. International diffusion assumes models will be available globally, with variation in access terms, pricing, or capability. Centralised diffusion inverts this: models are default-local, with access to other jurisdictions as an exception requiring explicit permission.

This report examines several observations. First, localisation is now the baseline architecture—the Fable/Mythos suspension revealed that models can be instantly geographically restricted without technical constraint. Second, regulatory authority now precedes technical deployment. Anthropic built Fable with extensive safeguards, red-teamed them with government participation, and documented them transparently; none of this prevented suspension. Third, because models are deployed as metered API services rather than distributed weights, control happens at the endpoint layer—governments restrict API access. Fourth, centralised diffusion is economically self-reinforcing: a model available in the US serves US customers at low per-token cost due to scale, while the same model restricted to a single jurisdiction serves fewer customers at higher cost. Fifth, developing countries face constraints in building autonomous AI infrastructure when deployed models may be subject to suspension.

The central finding is that centralised diffusion emerges naturally from the current infrastructure and governance environment.

Definitions

Three concepts need separating, because policy discussion has blurred them:

The "AI diffusion debate" of 2023-2026 largely assumed an international diffusion model. Some proposals aimed to expand access (democratise training compute, open weights); others to restrict it (export controls, capability tiers). Centralised diffusion asks why models should diffuse globally at all. The answer from multiple governments, simultaneously, is increasingly that they should not.

(From a centralisation perspective, this looks like democratisation—more actors have model access. From a governance perspective, it multiplies the regulatory points at which AI access can be restricted.)

The Fable/Mythos Suspension

The Fable/Mythos suspension did not emerge from a vacuum. It is the culmination of a regulatory escalation extending back several years:

The suspension reveals three technical facts:

What enabled this to happen with minimal procedural obstacle? The absence of a governing framework for model suspension. Anthropic had no formal recourse: no requirement for due process, technical review of the jailbreak's severity, proportionality assessment, transparency about the government's specific concern, or notice period for compliance. The company complied because non-compliance risked sanctions; the government issued the directive because no statute prohibited it.

(Export control of semiconductors affects supply chains; affected parties have months or years to adapt. Export control of a deployed API affects users instantly and globally.)

Before June 2026, the question was whether a government could restrict global access to a model. The suspension answers it: yes, instantly, under existing legal authority. The question now becomes whether governments should restrict global access. Major powers are converging on yes.

Evidence from major powers:

These are parallel moves toward the same architecture.

The Architecture

Localised model deployment operates across three technical layers:

Localised diffusion requires supporting regulatory infrastructure:

Together, these create incentives for localisation: a company deploying globally must comply with all these simultaneously; a company deploying locally complies with one jurisdiction.

The supply chain implications are significant. Pre-localisation, Anthropic/OpenAI train models in California and deploy globally via API. Post-localisation, US companies train for US deployment and export-permitted allies; China trains for China and Belt & Road partners; the EU trains for EU citizens; smaller countries license or adopt open-source models. Each supply chain develops different architectures, safety practices, and governance. Interoperability becomes difficult; standards diverge. This efficiency loss is intentional—regional control takes priority over global efficiency.

Democratisation vs. Localisation

The diffusion policy debate of 2023-2026 framed access expansion as democratisation. Democratisation aimed to expand capacity for training and inference globally by reducing barriers to entry—lower chip costs, open-source weights, tutorials. Localisation contains models within jurisdictions by mandating local deployment, restricting cross-border access, and applying export control.

Both produce a landscape with more actors having model access, though the governance structure differs. The outcome is jurisdictional compartmentalisation.

Autonomous Capacity

The foundational premise of the diffusion debate was that if developing countries can train models, they can build autonomous AI capability. This premise depends on deployment remaining coupled with training. A country can train a frontier model while being prevented from deploying it across borders; training and inference are now decoupled.

Consider: India trains a world-class language model competitive with GPT-5. Technically, this is autonomous capability. India cannot deploy it globally—export control restricts cross-border access. Indian users can use it; users elsewhere cannot. India has trained a model but cannot monetise it, scale its capabilities through user feedback, or establish market dominance. It operates at regional scale.

True autonomy requires both training capacity (ability to build models) and deployment capacity (ability to serve models globally at scale). Localisation decouples these: training is distributed (many countries can train), deployment is centralised (only major powers can serve globally). The inference economy recreates the asymmetry it was supposed to resolve.

The Geopolitical Logic

Why are major powers moving toward localisation? The answer is geopolitical.

(From a purely rational standpoint, localisation is geopolitically optimal—it gives each jurisdiction control and prevents dependencies on foreign infrastructure. This assumes jurisdictions view AI deployment as zero-sum. Geopolitical logic prevails over economic logic when security and autonomy are perceived as threatened.)

The Institutional Vacuum

Localisation is possible because there is no international framework preventing it. There is no UN Convention on AI Access, no treaty establishing model deployment as a right, no dispute resolution mechanism for export control. When the US suspended Fable/Mythos, Anthropic had no recourse—it could not appeal to an international body, challenge the government's reasoning, or request procedural safeguards.

The UN's AI governance initiatives have produced no binding instruments; the Global Compact on Digital Rights remains voluntary; the proposed AI Neutrality frameworks have no enforcement mechanism; MAITRI and other regional diffusion initiatives focus on building infrastructure, governing access only indirectly. In the absence of binding international law, unilateral action by major powers fills the space.

What Autonomous Deployment Would Require

To resist localisation and maintain genuine autonomous capacity, developing countries would need frameworks preventing unilateral suspension. This is technically and institutionally difficult.

This is achievable—it requires will and capital, though developing countries must coordinate in ways they have not yet demonstrated willingness to do.

The Path Forward

Localised diffusion is now the baseline assumption for major powers. The question is whether developing countries will accept it or resist it.

Acceptance means building local models and accepting local deployment, dependent on local infrastructure and subject to local government control. This is lower-cost and easier to implement, though lower-autonomy and creates rental dependencies.

Resistance requires five things:

Centralised diffusion—localised models bound to jurisdictions—is now the baseline for major powers. The Fable/Mythos suspension was a demonstration. The question developing countries face is whether they will be active participants in shaping it or passive consumers of it. Passive participation means adopting local models as dictated by major powers or renting inference from regional providers, accepting that models are jurisdiction-bound. Active participation means building regional infrastructure, coordinating on governance, and establishing frameworks that prevent unilateral suspension.

The infrastructure and governance for active participation are in progress. Building them is the work of the next 2-3 years. Every month that major powers consolidate localised deployment increases the cost of establishing alternatives. For developing countries, the question is how to build autonomous AI infrastructure that cannot be unilaterally suspended by foreign governments—a question requiring different investments, coordination, and governance frameworks.