Singapore's AI strategy risks becoming a consumption plan, not a capability plan

There is a version of this story that ends well. Civilisations that extracted fossil fuel cheaply understood, at some point, that the window would not stay open indefinitely.

The ones that used it wisely reinvested the proceeds into durable infrastructure — power grids, industrial capacity, renewable alternatives — building something that would outlast the extraction itself. Those that treated cheap energy as a permanent condition rather than a temporary subsidy burned it for consumption and built nothing that survived the correction. When the window closed, they found themselves dependent on whoever had moved faster.

Singapore now faces a structurally identical problem in a far more compressed timeframe. The fuel is not oil. It is the subsidised window of cheap, abundant access to frontier AI — a window that exists because private capital has not yet corrected to true cost, and which will not remain open indefinitely.

The question the Economic Strategy Review (ESR) — released on 13 May after five committees, eighty engagements, and consultations with more than 7,700 stakeholders — should be answering is what strategic capabilities Singapore must own before that window closes.

It engages that question only partially. Three failures explain the gap. They are strategic, structural, and temporal.

The ESR's AI framing is not wrong. Its second thrust — positioning Singapore as a global leader in AI solutions — correctly identifies the Champions model: leading firms like DBS, Singapore Airlines, and PSA adopt AI end-to-end, generate reference cases, and accelerate economy-wide diffusion.

Singapore as a trusted hub where banks, hospitals, regulators, and researchers solve problems together is a real and defensible advantage.

Workers' Party Member of Parliament Kenneth Tiong's written response, published the day after the ESR's release, sharpens the picture further.

Drawing on remarks he had already made in Parliament, Tiong named the risk the ESR does not: that frontier AI access is not a stable input but a variable one, geoblocked, selectively previewed, and increasingly instrumentalised.

His four-part response — harden security posture to reduce the justification for withholding, build data centres at scale, trade infrastructure access for contractual frontier guarantees, retain some domestic capability as fallback — is grounded and right.

But both the ESR and Tiong's analysis share a frame: secure the access. Neither fully asks the harder question. What will Singapore have built by the time the subsidy ends?

The first failure is strategic. Cheap frontier AI is not a permanent condition to be managed. It is a temporary subsidy to be exploited.

Tiong makes this precise: his subscription to Claude Max — Anthropic's premium AI tier — at $200 USD a month is, in his assessment, visibly underpriced against actual API costs, in the same way Uber rides were underpriced against true operational cost using venture capital funding. Capital captured the market; pricing corrected once consolidation was complete.

If frontier AI pricing normalises toward underlying infrastructure and inference costs over the next several years, Singapore's entire diffusion model — Champions trickling capability downward to small and medium-sized enterprises (SMEs) — prices out everyone below the enterprise tier before the diffusion completes.

A nation that does not control frontier models, compute supply chains, or inference infrastructure remains downstream from the strategic decisions of those that do.

The question is not whether to board the train. The question is what you build while the ticket is cheap.

There are four asset classes Singapore should be converting cheap access into right now: proprietary data infrastructure with Singapore-specific context that cannot be replicated externally; institutional process embedding across government agencies, hospitals, and courts; a critical mass of Singaporeans fluent with agentic tools; and infrastructure leverage that converts the data centre and water management advantage into contractual access guarantees.

Current access conditions are neither guaranteed nor permanent. The ESR’s diffusion-oriented framework appears to assume that current access conditions remain sufficiently stable for diffusion to complete.

The second failure is structural, and it is the one the tripartite consultation process is least equipped to surface.

The document assumes that corporate and worker interests are aligned in the AI transition. Businesses endorse workforce frameworks that cost them nothing and bind them to nothing — the framework relies on alignment incentives and voluntary coordination rather than statutory obligation.

Deputy Prime Minister Gan Kim Yong stated that the government's aim is to protect the worker, not the job. Acting Minister for Transport and Senior Minister of State for Finance Jeffrey Siow linked government support for AI adoption to clear workforce outcomes — job redesign, skills upgrading, career progression.

Both commitments are framed as intentions. Neither comes with a binding obligation on employers, a statutory standard, or a consequence for non-compliance. Without enforcement mechanisms, these commitments remain aspirational rather than enforceable.

To the ESR's credit, it explicitly acknowledges that economic growth may no longer automatically generate good jobs, and that AI disruption may prove more severe than expected. But its proposed response architecture remains overwhelmingly transitional rather than structural.

Human workers are currently necessary to train, validate, and supervise AI systems. The economic logic of augmentation does not necessarily stop at augmentation. Once training sufficiently completes, supervision requirements shrink, and the rationale for retaining those workers weakens materially — not through bad faith, but through fiduciary duty to optimise.

The case of Zhou in China illustrates what that arc looks like in practice.

A quality assurance professional at a Chinese tech firm, Zhou had been responsible for checking the accuracy of large language model outputs — precisely the supervisory role the augmentation phase creates.

When the AI system improved sufficiently to take over that function, he was demoted and offered a 40 percent pay cut. When he refused, he was terminated on grounds of AI-driven staffing reduction.

The Hangzhou Intermediate People's Court ruled on 28 April that this did not meet the legal threshold for contract termination, and ordered compensation. A separate court had established the same principle in December.

China, in the midst of a state-directed push to dominate AI development, has courts issuing rulings that protect workers from AI-driven displacement. Singapore's ESR addresses AI disruption primarily through transition support rather than statutory worker protections.

This arc is already visible across the coding profession, quality assurance roles, and early-stage automated driving — each augmented, then progressively automated, then replaced at the margin where the economics permit.

Overall retrenchments rose from 12,570 in 2024 to 13,790 in 2025, with services accounting for more than three-quarters of cases by year-end. The Ministry of Manpower (MOM), in its quarterly report, attributes the majority to business reorganisation or restructuring. AI-driven displacement is unlikely to appear in official statistics as automation layoffs — it surfaces instead through categories like restructuring and business reorganisation. These are not lagging indicators. They may represent the leading edge of the structural divergence.

The ESR's sixth thrust, on establishing a stronger system for career transitions and worker support, responds to that signal. It does not model where the divergence ends. The energy companies did not plan for when extraction ended. Governments that took them at their word managed collapse rather than transition.

The third failure is temporal, and it may be the most serious because it renders the other two irreversible. The ESR is a five-to-ten-year planning document. The technology is moving on a one-to-three-year cycle. Industrial policy traditionally assumes relatively stable technological baselines. Frontier AI development no longer provides that stability.

The document designs for the augmentation phase — AI complementing workers, lifting productivity, improving job quality — and presents that design as sufficient. The augmentation phase may not be the endpoint. It may function primarily as the data collection period.

There is also a feedback loop problem: Singapore's policy apparatus is evidence-based, which means it moves when the retrenchment data arrives. Certain forms of AI disruption may unfold faster than traditional policy feedback loops are designed to handle. By the time the MOM's quarterly reports reflect the structural shift clearly enough to compel intervention, the window for building may already have closed.

MOM's own Occasional Paper on Overqualification, published in April 2026 — one month before the ESR — acknowledges that with the growing adoption of AI, roles within clerical, administrative, and routine functions may increasingly be augmented or replaced by technology. The overqualification rate has risen from 16.3% in 2015 to 19.4% in 2025.

The paper classifies most overqualification as voluntary. But the same document records that PMET vacancies unfilled for at least six months rose from 14.4% in 2024 to 16.0% in 2025 — a labour market simultaneously pushing qualified workers into below-qualification roles and failing to fill the specialised roles that remain.

A worker who cannot find a role matching their qualifications and cites job stability for taking something lower is coded in the data as a voluntary choice. The structural pressure that produced that choice goes unexamined.

A reactive government sees rising retrenchment numbers and builds better transition support. A visionary government models the endpoint rather than the midpoint, and builds policy for the world that is coming rather than the one that currently exists.

Three questions deserve direct answers from the ESR committees and the government.

What will Singapore have built, not consumed — in proprietary data, embedded institutional capability, and contractual access infrastructure — by the time frontier AI pricing corrects?

What does your model say happens when the transitional phase ends and retention becomes more expensive than replacement?

If you do not have that model, who is building it, and why is it not in this document?

The window is open now. Singapore is generating heat. The question is whether it builds the capacity to generate its own power before the fuel runs out.