Sovereign AI & Strategic Infrastructure

Institutions need to decide which parts of the AI stack they must control, what they can procure, and how they will keep essential services operating when a dependency changes.

Why this matters

AI services depend on more than models. Data, cloud services, chips, identity, software and trained operators determine whether a system can perform its intended work. A locally developed model can support language capability while relying on internationally supplied infrastructure. A globally supplied service can offer useful operating support while leaving the customer with responsibility for permissions, data and the business process.

The Singapore National AI Strategy update of May 2026 connects compute availability and efficiency with governed datasets, research, talent and partnerships. These priorities make infrastructure a question of capability and resources as well as ownership. The strategy describes intended actions, rather than certifying any institution's readiness or independence.

Dependence can also be costly to change. The UK CMA's cloud investigation identified commercial and technical barriers to switching, including egress charges, service differences and skills constraints. Its geographic market is UK-plus-EEA; it is not a measurement of ASEAN market concentration. It provides mechanisms a regional buyer can investigate in its own architecture and contracts.

Our starting position is that control should be defined through specific rights and operating capabilities. The important evidence is whether an institution can assess the system, govern its use, retain its records and continue an agreed service during disruption. Ownership and partnerships are alternative means of meeting those requirements, with different costs and responsibilities.

Questions we examine

• What must remain under local authority? Separate the protection of sensitive records, control of consequential actions, local-language capability and continuity of essential services.
• Where is the material dependency? Trace cloud, model, chip, identity and software relationships below the supplier's name, including dependencies shared by nominal alternatives.
• What does an open model actually allow? Inspect licences, weights, code and data information, then assess whether the institution can operate and maintain the deployment.
• Can the workload move? Test the export of data and operating records, replacement quality, permissions and recovery time before relying on an exit clause.
• What capacity can be sustained? Examine compute availability, support, resource efficiency and the cost of retaining an alternative or fallback.
• Which assessment belongs to which entity? Make processing routes and country-specific acceptance explicit when a Singapore-based operation serves other ASEAN markets.

These questions apply at different scales. A national programme may seek research capability and public benefit across an economy. A company needs evidence for its own critical services. The distinction prevents a national ambition from becoming an unsupported assurance about a particular deployment.

Analysis and working resources

Sovereign AI does not necessarily mean sovereign models explains how to distinguish national capability from control over an individual service. It provides a decision record for defining the requirement, identifying the dependency and comparing investments.

Who controls the AI stack? addresses procurement and operating resilience. It compares deployment choices and sets out an exit exercise that tests the records, quality, access and people required for the replacement service.

For a regional deployment, the ASEAN rollout workbook records acceptance separately for each entity. Its purpose is to make data flows, local requirements, workload tests and approving owners inspectable. It does not provide a blanket legal opinion across ASEAN.

The release and operating pack helps define permitted actions, monitoring, recovery and responsibility. The investment and portfolio memo connects architecture choices to workload value, full operating cost and the evidence needed for a funding decision.

These resources are decision aids. They do not establish a measured sovereignty score, certify compliance or prove the performance of a particular supplier. Downloadable assets and copyable records should be used with the institution's own evidence.

From research to an infrastructure decision

A private executive briefing can examine one defined choice: an AI platform purchase, a local deployment, a regional architecture or an essential service's fallback. A useful starting packet contains the current service map, data categories and processing destinations, supplier terms, workload requirements and the disruption the business can tolerate.

The resulting discussion should identify what the institution needs to retain, the evidence missing from the proposed arrangement, and the next test that would resolve the decision. It can inform a build, procurement or operating plan without assuming that more ownership, more suppliers or a national model is always the best answer.