Winning the AI race without an OpenAI

For three years, the global contest in artificial intelligence (AI) has been reduced to a single race to build ever larger models. The US has OpenAI, Anthropic and Google; China has DeepSeek, Alibaba and ByteDance; Europe has busied itself with rules. In the press, on forums and across capital markets, the question keeps returning to the same point: who can train the more powerful model? 

Computing power, chips, parameters and training costs have become the standard yardstick of a nation’s AI strength. Yet shift the gaze from the laboratory to the real world, and a different picture comes into view. Some institutions that own advanced models have not seen the productivity gains they expected, while some economies that build no foundation models at all are turning AI into practical capability at speed, driving industrial efficiencies and improving public services. The real contest of the AI age may not be confined to the lab.

What decides a country’s future competitiveness might not be who holds the strongest model, but who can embed AI into the real workings of society and the economy fastest and most reliably. The next round of competition may turn less on technical capability than on institutional capability. This judgement is not made in a vacuum. Advanced models are becoming steadily easier to obtain. Today a large corporation and a small firm alike can call on the world’s leading large language models through the cloud. 

Beyond the GPU arms race

Differences between models remain, but the barrier to acquiring the technology is falling quickly. What now separates the leaders from the rest is no longer whether they have AI, but whether they can make AI actually work. Many deployments fail not because the model is not clever enough, but because the organisation is not ready. Processes have not been redesigned, lines of responsibility are blurred, data does not flow and staff are unsure how to work alongside the machine. Advanced technology then stalls at the pilot stage and never reaches core operations. In other words, the greatest challenge AI brings is often not technical innovation — but institutional and organisational innovation.

The capability that underpins such innovation might be called a country’s “AI operating system”. It is not a piece of software but a body of institutional infrastructure that allows AI to be applied widely and reliably: digital infrastructure, systems for training talent, regulatory frameworks, industry standards, organisational governance and mechanisms of social trust. Measuring technological strength once meant counting papers, patents and computing power. At the application stage, the better question is how fast a society can plug a new technology into its own operating system. The gap between nations may come to rest not on who owns the most GPUs, but on who has the most mature AI operating system.

Singapore’s advantage

Seen this way, Singapore’s position may be more favourable than it looks. It has never been the world’s largest technology market, and it has no intention of competing head on with the US and China on foundation models. That does not leave it short of advantages in the AI age. If anything, it holds a set of structural conditions that are often underrated. 

First, Singapore is a highly institutionalised society. In finance, healthcare and public services, most organisations have clear processes, a mature culture of compliance and a strong capacity to execute. Such institutionalisation is sometimes read as conservatism — yet when it comes to putting AI to work, it is a rare asset. AI does not work well in disorder; it is better suited to settings where rules are explicit, responsibility is clear and processes are stable. 

Second, the economy is already deeply digitalised. Digital identity, electronic payments and data infrastructure were laid down years ago, which lowers the most common obstacle to AI adoption, namely systems that cannot connect to the underlying business data. 

Third, an ageing population and limited growth in the workforce make adopting AI not merely a strategic choice but a practical necessity, and history suggests that necessity drives the genuine spread of a technology more powerfully than enthusiasm. 

Fourth, the country’s scale makes cross-agency coordination possible. Governments, universities, industry bodies and training providers that pull in separate directions in many large economies find it easier here to move the same way. Once a new national strategy is announced, the links from policy to training and from industry to education tend to form quickly. This capacity for coordination has not stood out in the past, but in an age where the speed of diffusion decides the outcome, it may prove the decisive variable.

These conditions are not theory. Since 2017, AI Singapore has supported more than 300 applied AI projects locally, with a deployment rate of around 60%, against a widely cited benchmark from MIT of roughly 5%. The gap lies not in the quality of the models, but in the order of work. The projects that succeed complete the mapping of processes, the preparation of the organisation and the definition of responsibility before the first line of code is written. In the partnership with the local power company YTL PowerSeraya, the team embedded the system into existing workflows first and turned to implementation second, which delivered faster decisions and greater operational accuracy.

Integration before innovation

The distance between 60% and 5% points to a fact that is easily missed. What decides whether AI succeeds is often not the model but a repeatable method of deployment: define the problem and the responsibility, embed the work into the process and only then turn to the technology. This “integration first” sequence is an institutional capability made concrete. It cannot be bought with a stronger model; it can only be built up through the organisational experience a society accumulates over time. 

The same absorptive capacity shows in how talent is trained. In AI Singapore’s apprenticeship programme, about 80% of trainees do not come from a computer science background, but are practitioners from other fields onto whom AI skills are grafted. The aim is not to produce a handful of elite researchers, but to widen — through national reskilling — the capacity of the whole society to absorb AI.

At the same time, Singapore has kept investing in another piece of infrastructure that is easy to overlook, which is trust. The Model AI Governance Framework, the AI Verify testing tool and the alignment with international standards such as ISO/IEC 42001, are often misread as an added regulatory burden. Their real function is to give heavily regulated sectors such as finance and healthcare the confidence to use AI in decisions that matter.

The earlier conditions concern whether a society can absorb AI; governance and trust are core factors when it comes to willing AI adoption, and both are crucial. However, none of this guarantees success. The same institutionalisation that can give AI structure can also obstruct it, once regulators treat every deployment as a risk rather than a capability. Whether the advantage is realised depends on Singapore recognising that its comparative advantage in the AI age lies not in building models, but in using them well.

The logic reaches beyond Singapore. Across Southeast Asia, most economies will not become centres for building foundation models, yet they face the same difficulty of turning AI into productivity. If Singapore can be the first to make the institutional capability of “using AI well” into a model others can learn from, its value in the region will no longer rest on how large its models are, but on whether it can export a body of governance and organisational experience that lets AI land reliably.

Many still measure AI strength by the number of patents, the level of research spending and the size of models. Those measures matter, but they mainly capture the supply side of innovation. The real difficulty of the AI age lies in the distance between supply and absorption: once a technology appears, can it be adopted widely across society, used effectively by organisations and turned into higher productivity and public value? In the end these questions test not the laboratory, but the institutions.

Escaping the pilot loop

The real dividing line becomes clear from here. It is not which countries own GPUs and which do not, but which organisations and countries already have the institutional capability to make AI truly work, and which remain caught in round after round of pilots. Pilots can run forever while value never arrives — the ability to embed AI into core processes is the gate that turns technology into productivity. The laboratory sets the ceiling of a technology, the organisation decides whether it reaches reality and institutions decide how far a country can finally carry it. 

People once assumed that the AI contest was a contest between models. In time we may find that what truly decides the outcome is how a country organises itself.

Related: Beyond Elon Musk: Who benefits from the trillionaire era? | Mythos vs DeepSeek: Why cheap AI could tip the US-China tech race