Securing the sovereign foundations of Singapore’s AI economy

SINGAPORE’S Economic Strategy Review (ESR) – 32 recommendations across eight thrusts, drawn from more than 80 consultation rounds with over 7,700 participants – is a serious piece of work.

Acting Minister for Transport and Senior Minister of State for Finance Jeffrey Siow has described it as an effort by a new cohort of political office-holders, working with stakeholders across society and a new generation of business leaders, to chart how Singapore can stay competitive and resilient in a changed world.

Understandably, much of the public debate has centred on its artificial intelligence recommendations and the future of jobs. In every major wave of industrialisation, labour disruption has been real and painful.

Yet, over time, it has also proved cyclical. Research firm Forrester found that 55 per cent of employers regret their AI-driven layoffs.

IBM, Salesforce, Google and Meta have added back workers in redefined roles since late 2025. Research advisory Gartner now projects that by 2027, half of companies that cut for AI will rehire for similar roles.

While the churn is real, it is not destiny. Workers and businesses still have room to reposition themselves in an AI-enabled economy.

The ESR’s AI recommendations – on skills, company-level adoption and innovation – are designed to support exactly that.

Where the ESR is less clear is on the structural preconditions that make these recommendations possible.

Even the best skills and adoption programmes will falter if Singapore cannot reliably secure three inputs: access to advanced AI models; the tokens that ration the use of model capacity; and the semiconductor chips that power them.

These are the reserves of statecraft.

What an AI-enabled economy runs on

AI model access is the first structural constraint. More than 70 per cent of companies in Singapore already report some form of AI use.

Much of it is built on closed-weight frontier systems such as the GPT, Claude and Gemini families, said financial services company Morgan Stanley. These systems are currently best-in-class.

The performance edge, however, comes with a sovereignty trade-off – the models are proprietary and centrally controlled, so the terms of access are ultimately set abroad. This is regardless of how carefully Singapore regulates domestic use.

Project Glasswing and Anthropic’s Mythos preview demonstrated how a single frontier model could expose thousands of severe flaws across energy, health, logistics and enterprise systems.

It did so faster than state institutions could respond, forcing governments to confront the problem of “sovereign recovery” for infrastructure they do not control.

In an AI-enabled economy, a policy alignment or access change by a few foreign providers can therefore ripple through banks, ports, education, media, and small and medium enterprise (SME) tools at the same time.

The ESR rightly supports AI innovation and encourages businesses to experiment with leading AI tools. But it should also spell out how Singapore will secure long-term, predictable access to these frontier models on terms that cannot be changed unilaterally abroad.

That demands a state-led model access strategy comprising diversifying providers, anchor-tenant deals for critical sectors and, where feasible, sovereign control layers for core public services shielded from anticipated geopolitical pressures.

The next constraint is the adequacy of tokens. If the ESR succeeds in mainstreaming AI across sectors, Singapore’s consumption of AI tokens will rise sharply.

Tokens are the operative unit of AI consumption because they determine how much frontier capability Singapore can actually deploy.

But unlike oil, the price of which can be hedged across suppliers and contracts, token pricing is concentrated and structurally biased towards rising as models become more capable and workloads more complex.

An OpenRouter analysis of the April 2026 GPT-5.5 launch found that list prices doubled from GPT-5 to GPT-5.5.

And once larger context windows, reasoning modes and tool-use premiums were included, real-world per-token costs for many workloads rose by roughly 50 to 90 per cent.

At the same time, agentic workflows are expected to consume five to 30 times more tokens per task than today’s simple queries, indicated Gartner research.

Even if raw inference gets cheaper, total spend is expected to rise as AI becomes more deeply embedded in day-to-day operations of an AI-fluent workforce.

Executives at companies such as Uber have already admitted that they exhausted their planned 2026 AI budgets within months as engineers leaned heavily on Anthropic’s Claude models.

Bryan Catanzaro, vice-president of applied deep learning research at Nvidia, similarly noted that for some teams, compute now costs more than the salaries of the people using it.

Meanwhile, China’s daily token consumption has already exceeded 140 trillion – more than a thousandfold increase in two years.

At the national scale, token access and pricing should therefore start to look like fuel for the digital economy.

The ESR should also have recommended treating tokens the way Singapore already treats oil, water and strategic food reserves – through pooled procurement for key sectors; long-term contracts where justified; and for essential public digital services, the equivalent of a “strategic token reserve”.

That requires coordination and bargaining power that only the state can muster.

The third constraint: chips on which AI compute runs.

Every AI prompt from Singapore ultimately runs on a small number of advanced semiconductor chips. Most of them are designed in the US and subject to tightening export controls.

Given the dominance of US AI infrastructure here, Singapore’s AI economy is expected to run primarily on US-designed, export-controlled graphics processing units (GPUs) deployed in local and regional data centres.

This is whether the workloads belong to Singapore banks, global tech firms or foreign customers.

On Jan 12, the US House passed HR 2683, the Remote Access Security Act. It extends export-control rules so that remote or cloud access to advanced AI chips can be regulated like physical exports.

In practice, this allows Washington to treat foreign access to US GPUs via cloud services as an export that can be licensed, limited or cut off.

Against that backdrop, Singapore’s legislative assurances for chip supply access remain thin.

For example, the US-Singapore Critical and Emerging Technology Dialogue, as set out in the October 2023 joint statement, focuses on AI governance, standards and cooperation, but does not include binding commitments on chip access.

Deputy Prime Minister Gan Kim Yong’s April 2025 visit to Washington, during which he discussed AI chips, tariffs and critical exports with US Commerce Secretary Howard Lutnick, likewise produced no public offtake guarantees or priority-access arrangements.

None of this features prominently in the ESR’s recommendations, which naturally emphasise domestic skills and industry support.

Yet, if Singapore’s AI workloads sit on hardware that can be throttled by the decisions of foreign governments, that strategic dependence could become strategic vulnerability for our AI-enabled economy.

What only the state can do

For Singapore, AI sovereignty is therefore not an abstract ideal. It is the practical task of ensuring that dependence on others cannot be turned against us, while preserving our ability to run the economy on our own interests.

In the AI era, these dependencies include – among other key inputs – an assured access to AI models, tokens and compute capacity.

As the government acts on the ESR, it should therefore approach our AI-enabled economic transformation as two parts: upgrading skills and businesses on the surface; and, beneath that, building a sovereign foundation that cannot be arbitrarily cut off or priced out of reach.

That is a structural mandate for national leadership – and not a problem that individual companies or workers alone can solve.

The writer is chairman of the public affairs group at the Public Relations and Communications Association Asia Pacific, and is currently reading war studies at King’s College London