Three Paths to 2027: Convergence, Two Stacks, or AI Non-Alignment
FUTURES · FORESIGHT. Horizon: 19 July 2027. These are structured scenarios and observable signposts, not events the Institute claims will occur. The factual baseline is 19 July 2026.
The factual baseline is narrow.
Chinese models have moved close to the global frontier. China has established WAICO with 29 founding signatories and has presented open-source ecosystems and capacity-building as elements of a just global AI order. The United States and partners are expanding Pax Silica around secure, trusted supply chains. The United Nations has opened a universal Global Dialogue on AI Governance.
What these structures become is not yet known.
The purpose of foresight is not to convert that uncertainty into confident prediction. It is to identify distinct paths, the mechanisms that would produce them, and the evidence that would tell us which path is forming.
Scenario 1: Conditional convergence
In this path, geopolitical competition remains intense but does not divide the technical ecosystem into incompatible stacks.
Chinese open-weight models are widely used through Western and regional serving infrastructure. US and allied chips remain important across multiple political alignments. Shared evaluation methods, agent-identity conventions, provenance formats, and security reporting create a thin but workable interoperability layer. WAICO, Pax Silica participants, and the UN Global Dialogue coordinate on some standards without sharing a single political model.
States retain different content, data, and procurement rules. The systems remain portable enough that organisations can test and migrate among models from different origins.
Mechanism
Economic interdependence and developer demand make hard separation too costly. Open weights reduce model-level lock-in. Common safety incidents create pressure for shared technical records even where policy diverges.
Signposts
Chinese model weights run broadly on US, European, Gulf, and domestic infrastructure.
Major serving frameworks support models across jurisdictions without separate code bases.
WAICO and UN processes reference compatible evaluation or reporting standards.
Decision-authority, identity, and provenance records map across regulatory regimes.
Export controls remain targeted rather than becoming a comprehensive model and software embargo.
Disconfirming evidence
broad bans on foreign-origin models;
mandatory nationally isolated serving stacks;
incompatible identity and trace standards;
systematic withdrawal of cross-border model access.
Scenario 2: Two stacks
In this path, WAICO and Pax Silica become centres of increasingly separate ecosystems.
The US-led stack integrates trusted semiconductor, cloud, model, security, and procurement relationships. The China-led stack integrates domestic and partner hardware, open-weight models, capacity-building, standards, and infrastructure agreements. Access to advanced components becomes conditional on alignment. Cross-stack models remain technically possible but are legally, contractually, or operationally disfavoured.
The split is not only about model content. It reaches accelerators, compilers, serving engines, cloud attestations, agent identity, evaluation, data flows, and certification.
Mechanism
National-security policy expands from chips to the full AI stack. Each side treats dependence on the other as an unacceptable strategic vulnerability. Procurement and standards convert political alignment into technical incompatibility.
Signposts
participation in one initiative limits participation or trusted status in the other;
model and cloud procurement rules use country-of-origin exclusions;
distinct hardware-software optimisation ecosystems deepen;
evaluation and certification become bloc-specific;
states must choose standards or identity systems that are difficult to recognise across stacks;
model access is withdrawn during political disputes.
Disconfirming evidence
continued cross-stack commercial adoption at scale;
mutual recognition of technical assurance;
major states maintaining material roles in both ecosystems.
Scenario 3: AI non-alignment
In this path, middle powers and Global South states use competition to assemble mixed stacks rather than choose one.
A country may use Chinese open-weight models, allied or domestically acquired accelerators, European data and rights rules, Gulf-funded data centres, local-language datasets, and UN scientific assessments. It negotiates training, finance, infrastructure, and technology transfer from multiple sources.
The result is not neutrality. It is active portfolio strategy.
Mechanism
Neither bloc can offer every layer at the best price or on acceptable terms. States gain bargaining power by preserving alternatives. Open weights make the model layer more portable, while energy, capital, language, and regional infrastructure create new centres of leverage.
Signposts
countries sign infrastructure and governance agreements across supposed bloc lines;
procurement requires multi-model portability and local operation;
regional clouds and evaluation institutes grow;
local-language models are built from multiple foundation-model families;
states demand training, code, or joint laboratories rather than API credits alone;
UN forums become venues for coordinating mixed-stack rules.
Disconfirming evidence
exclusive supply conditions become normal;
the cost of interoperability overwhelms the benefit of diversification;
middle powers lack the institutional capacity to manage multi-stack risk.
The variables that decide the path
Weight releases
If frontier-class weights are repeatedly released and practically deployable, the model layer becomes more portable. If governments or developers restrict high-capability weight distribution, hosted and trusted access becomes dominant.
The first short-term signpost is Kimi K3. The announced 27 July release should be recorded as delivered, delayed, partial, or changed, together with its licence and deployment reality.
Compute substitutability
Model portability matters less if only a few actors can supply the accelerators, networking, energy, and software required to run the system.
Watch whether alternative hardware ecosystems support leading models at competitive cost and whether states build usable domestic or regional compute rather than announce capacity alone.
Institutional delivery
WAICO becomes consequential if it moves resources, develops standards, funds capacity, or changes member-state options. Pax Silica becomes consequential if participation produces resilient supply, investment, and operational preference. The UN layer becomes consequential if scientific assessment and dialogue affect procurement, standards, or cross-border practice.
Count projects, budgets, rules, and changed decisions, not speeches.
Access contestability
An AI order becomes governable when access decisions can be reconstructed and challenged.
Watch for public criteria, notice, reasons, review dates, exceptions, appeals, and correction mechanisms in export controls, model-provider restrictions, cloud eligibility, trusted-partner status, repository releases, and public procurement.
If consequential access continues to be changed by private terms or state letters without a decision record, the order will be powerful but weakly accountable under every scenario.
The Synthocracy watchlist
| Signpost | What to record by July 2027 |
| Frontier open weights | releases, licences, independent replication, real hosting cost |
| WAICO | ratification, membership, budget, secretariat, projects, standards |
| Pax Silica | membership, conditions, finance, infrastructure, access consequences |
| UN governance | scientific reports, dialogue outcomes, uptake in national decisions |
| Export controls | object, recipient class, criteria, duration, exceptions, remedy |
| Model restrictions | provider, jurisdiction, affected users, notice, reversibility |
| Compute | regional capacity, hardware diversity, energy constraints, utilisation |
| Agent standards | identity, delegation, tracing, authority and mutual recognition |
| Multi-stack states | procurement portability, mixed suppliers, local evaluation capacity |
What remains invariant
The scenarios differ in openness, alignment, and interoperability. One governance question persists across all three:
Who can open or close access to a consequential AI capability, under whose authority, on which evidence, for how long, and with what route of challenge?
In conditional convergence, that question prevents interoperability from becoming unaccountable diffusion.
In two stacks, it prevents trusted status and exclusion from becoming invisible exercises of power.
In AI non-alignment, it prevents a diverse supplier portfolio from becoming a dependency web no institution can reconstruct.
The global AI order may converge, split, or become multipolar. Under every path, access remains a decision.
Sources for the factual baseline
Ministry of Foreign Affairs of the PRC, Signing Ceremony of the Agreement Establishing WAICO, 16 July 2026.
Chair’s Statement of the 2026 WAIC, 17 July 2026.
European Commission, Commission signs Pax Silica declaration, 25 June 2026.
United Nations, Global Dialogue on AI Governance.
Stanford HAI, The 2026 AI Index Report.
Artificial Analysis, Kimi K3 evaluation, 17 July 2026.
This is Article 7 of Open-Weight Power: Models, Access, and the Emerging AI Order.
