Singapore Wrote the Sovereign AI Playbook. Almost Nobody Funded It.

Ask Singapore government technology leaders whether secure data sovereignty shapes how they invest in AI, and 93.3% will say yes. But when you ask how many are already investing significantly in it, the number falls to a mere 3.3%.

Both figures come from Building a Sovereign AI Foundation for the Singapore Government, an IDC InfoBrief sponsored by Dell Technologies and NVIDIA, based on a December 2025 survey of 30 Singapore government IT decision-makers. 40% are evaluating sovereign AI technologies while another 43.3% are running proofs of concept with no spending plan attached. Nearly nine in 10 agencies are looking and almost nobody has committed.

The stall is regional. Across eight Asia Pacific markets, the same research found sovereign AI has climbed from the seventh to the second-highest government investment priority in a year. Priority moved; procurement did not.

The constraints are also not unique to the government industry. A Singapore bank, hospital or telco faced the same questions about residency, model control and auditability. Government just has to answer them in public.

Sovereignty is a decision, not an address

Asked to define sovereign AI, 43.3% of Singapore government leaders picked strategic competitiveness and innovation capacity. Only 10% picked national control and autonomy.

“Sovereignty is not about saying ‘we own everything,’” one senior Singapore government official told IDC. “It is about whether the nation retains control over the data, the infrastructure, the models, and the decision-making process.”

Andy Sim, vice president and managing director of Dell Technologies Singapore, says the organizations making progress classify workloads before they buy anything. Three questions decide it: What are the consequences if this data is exposed? What happens if this system is manipulated? How much human judgment remains in the loop?

Score high on all three and the workload runs fully on-premises, with no external dependencies at critical points. National security systems sit here. Score high on some, and partners can supply technology and elasticity while the institution keeps governance, residency and oversight. Healthcare analytics, tax systems and citizen services occupy that middle tier. Everything else runs in shared environments with sovereign controls switched on. Singapore’s Government Commercial Cloud already works this way.

“Classify first, then build the architecture around that answer,” Sim says.

The question your compliance map doesn’t ask

Most residency programs track where data sits. But Sim argues that is an incomplete noun.

“A practical test is to assess not only where the data resides, but where prompts are processed and tokens are generated,” he says. “In regulated environments, that matters because sovereignty risk can arise from the movement of context and inference, not just from where data is stored.”

A customer record can sit in a Singapore data center all year and never move, while the prompt describing it, the retrieval context around it, and the model output about it are generated somewhere else. The map shows green but the exposure sits outside the frame.

Autonomous agents make this worse. They generate far more tokens than humans do, and rising volume turns a governance question into a cost and latency question at once. The choices also harden fast. Where models run and how data is routed become governance commitments that are expensive to unwind two years later.

Uniform control is how the budget dies

Security leads Singapore’s concerns at 50%, specifically the vulnerabilities that leave sovereign AI models open to adversarial attack. Premium pricing follows at 40%, then technical complexity at 33.3%.

Some of that price is self-inflicted. Applying full sovereign controls uniformly across every workload drives costs up, Sim says, and it consumes capacity that is already short. “Access to compute is becoming a critical bottleneck for AI development,” one Singapore official told IDC in the survey.

The counter-argument sits inside the same survey. Some 76.7% of leaders agreed sovereign AI may inadvertently restrict innovation by limiting access to global technology ecosystems, advanced research and cutting-edge solutions. These are not skeptics but are the people building sovereign AI systems, and pricing in the cost of their own strategy.

Three layers, never bundled

Ask Sim what to standardize against lock-in and he names three layers. Open model frameworks, so you can run, fine-tune and swap models as the field moves; a portable data layer built on open APIs and standard formats, so changing platform does not mean wholesale reingestion; a governance and security layer owned independently of any vendor’s control plane: identity, access controls, encryption standards, audit frameworks, compliance monitoring.

“When you own governance, you can change technology partners without rebuilding your compliance posture from scratch,” Sim says.

The failure mode is convenience. “The organizations that get locked in are those that standardize on a bundled solution that handles all three layers together conveniently,” he says. “Convenient until the moment you need to change one component and discover everything is tightly coupled.”

Singapore’s buyers already lean that way. Ecosystem participation, meaning open standards, developer communities and local partnerships, is the top criterion for choosing a sovereign AI provider at 60%, ahead of data governance and residency at 56.7%.

The unglamorous part decides it

Half of Singapore agencies now document data definitions and lineage across multiple projects, up from 40% a year earlier. Only 6.7% manage data quality risk enterprise-wide with preventive measures, down from 10%.

Their approach is also changing shape. New investment in modernized data architecture dropped from 66.7% in 2024 to 30% in 2025, while investment in data abstraction layers nearly doubled, from 26.7% to 46.7%. Fewer agencies are rebuilding the foundation; more are putting a governed layer over what they already have.

“Sovereignty risk can arise from the movement of context and inference, not just from where data is stored.” — Andy Sim @ Dell Technologies Singapore

People are the second constraint, and the numbers move in opposite directions. No Singapore agency reported a major skills shortage in 2025. Yet only 3.3% said they can access the skills they need for all initiatives. The shortage did not deepen but spread. Cybersecurity specialists are hardest to hire at 50%, then data architecture and analytics at 43.3%, then sovereign data governance at 40%.

“Customers plan for technology and underbudget for people,” Sim says. A proof of concept runs on specialists. Production runs on 24/7 operations, incident response and model monitoring.

Where it stops being compliance

Almost every Singapore survey respondent expects agentic AI to accelerate adoption, and 76.2% attach a condition: strong guardrails and oversight first. Sim’s list of what belongs in infrastructure rather than the application is short. Tamper-evident audit trails for every agent action; Workload isolation, so an agent cannot escalate privileges or move laterally; real-time observability with a hard stop that fires before the application registers a problem.

The payoff is the argument a CDO can take to the boardroom. Sovereign AI earns its cost when it unlocks data the organization already owns and cannot use.

“Singapore’s banks sit on extraordinary volumes of transaction data and risk signals,” Sim says. Much is too sensitive for shared cloud under current rules, so it stays locked. “The bank that trains on its full data estate under its own governance moves faster than one constrained to the subset it’s comfortable sending to a third party.”

Hospitals tell the same story. Population health data becomes a research asset rather than a compliance liability once the infrastructure makes it safe to use.

Four things to settle this quarter

• Map inference, not just storage. For the top 10 AI use cases, record where prompts are processed and tokens are generated. If that differs from the residency map, there is work to do.
• Price the tiers. Identify which workloads justify full sovereign controls and which are paying for them by accident.
• Check who writes the audit trail. A record generated by the application rather than the platform can be bypassed by the application.
• Name the owner of the governance layer. If the infrastructure partner changed next year, would identity, encryption and audit frameworks survive the move?

“AI is now comparable to identity or payments infrastructure,” another Singapore official told IDC. “Once you depend on it, you must govern it.”

The 3.3% will rise. The organizations that move first will not be the ones with the largest compute contract. Instead, they will be the ones that knew which workloads truly deserved it.

Image credit: iStockphoto/Jacob Wackerhausen