Who should own AI? The case against – and for – nationalisation

In May, DBS’ chief economist Taimur Baig and former International Monetary Fund economist and author Anthony Annett made a case in The Business Times for nationalising parts of the AI industry – the same treatment reserved for nuclear weapons and central banks.

It was a well-argued piece and deserves a careful answer.

The question it raises is not really “should AI be nationalised” because that framing is too blunt. AI is not one thing. It is foundational models trained on the world’s text, chip fabrication plants, data centres, software firms creating AI agents and thousands of start-ups building applications. The real question is: nationalise what, exactly, and how much?

The case for nationalisation

Baig and Annett’s argument deserves to be heard in full because it addresses some legitimate concerns.

Their first concern is safety. Advanced AI systems could, in the wrong hands, disrupt financial systems, erode or destroy wealth, or interfere with utilities such as power grids, water systems and transport networks. Rogue actors could use AI to spread disinformation at industrial scale, run scams that are indistinguishable from genuine communication or, worse, design a lethal pathogen.

These are not science-fiction scenarios; they are the stated concerns of the very labs behind the technology.

To build on their case, Baig and Annett reach for an analogy: nuclear power. It can provide a clean, reliable source of energy, but also build weapons of mass destruction. That dual-use nature is precisely why nuclear technology sits under state control everywhere, bound by international non-proliferation treaties.

If AI carries a comparable dual-use risk, the argument goes, perhaps it deserves comparable treatment.

Their second concern is about power. A handful of firms – OpenAI, Anthropic, Google and a few others – are building the foundational systems on which entire economies may soon run. Besides worsening inequality, that kind of concentration does not stay economic; it can translate into political power and public ownership would blunt that conversion.

There is also a fairness argument: The breakthroughs underpinning today’s AI trace back to decades of publicly funded research, and the models themselves are trained on data effectively generated by all of us.

If the public funded the seed and supplied the soil, perhaps it deserves a share of the harvest – a capital dividend, distributed the way a sovereign wealth fund distributes returns, rather than a windfall captured entirely by a handful of shareholders.

It is a persuasive case. But it is also too broad to survive contact with how AI actually works.

The nuclear analogy does not really fit. Uranium is a physical substance. You can fence a reactor, guard an enrichment facility and inspect nuclear sites to detect diversion to weapons-grade uranium, which is roughly what the non-proliferation regime does.

AI is not like this. It is code, learning models and know-how. It has already diffused across thousands of researchers, universities, open-source repositories and, now, as we have seen with the rise of freely downloadable Chinese models, across national borders that no treaty currently governs. You cannot ring-fence what has already leaked into millions of laptops.

Even narrowing the target to the handful of firms building frontier models does not make state ownership costless.

Governments are rarely the fastest or most creative builders of cutting-edge technology – bureaucracies optimise for caution and accountability, not the rapid, hit-and-miss experimentation that produces breakthroughs. Nationalise the labs building frontier models and you may get safer AI, but you will very likely also get slower and less capable AI.

Why nationalisation cannot be a blanket policy

The AI ecosystem has layers, and the safety and power arguments apply to them unevenly.

At the very top sit the foundational models – the GPT-, Claude-, and Gemini-class systems capable of the catastrophic misuse Baig and Annett describe. This is the layer where their argument is strongest; it is where the nuclear analogy actually holds some water.

Beneath that sits infrastructure – chips, power supply, data centres. This is a resource-allocation and supply-chain problem, not a control-of-catastrophic-capability problem. Taiwan does not need to nationalise TSMC to manage the risk of a rogue chip; it needs export controls and standards for testing and verification.

And at the widest layer sit start-ups and application developers – the thousands of companies building narrow, specific tools on top of models they do not own or train themselves. This is where nationalisation makes the least sense because this is precisely the layer where competition, speed and trial-and-error matter most, and where the catastrophic-risk argument barely applies.

Collapsing all three layers into a single policy – “nationalise AI” – is a category error. The debate should really be about the top layer alone, and even there, cautiously.

The Singapore case

For a country like Singapore, nationalisation makes even less sense.

Singapore’s AI strategy has never been about building the next frontier model; it is about applications – tools built by private firms, often small ones, solving specific problems in logistics, finance, healthcare and language.

The country’s economic model runs on foreign investment, open competition and a reputation for being an easy, predictable place to build a business.

A state-dominated AI industry would cut directly against all three.

It would signal to foreign investors that the state intends to dominate the sector in which they plan to invest. It would narrow competition in exactly the layer – applications – where competition drives quality. And it would very likely produce worse products, not safer ones, since the safety case for nationalisation was always about frontier models, not about the delivery app bolting a chatbot onto its customer service line.

Singapore’s interest is not in owning AI. It is in making sure AI, mostly owned and built elsewhere, is governed well when it arrives.

The better role: referee and occasional coach

If outright ownership is the wrong tool for nearly the whole AI economy, that does not leave governments with nothing to do. It leaves them with the more difficult job of designing the rules of the game rather than becoming the main player – and helping the game move faster in the right direction.

The first lever is guard rails around sensitive sectors. National security, finance and critical utilities deserve tight, specific regulation because they are exposed to systemic risk.

The second is incentive design. Tax breaks and subsidies can be effective instruments, and governments can use them to nudge AI development towards job-augmenting tools rather than job-displacing ones; to slow the roll-out of AI in elementary education such that it does not blunt children’s cognitive development; and to reward innovation aimed at public good in areas such as public health, urban planning and disaster response, rather than purely private profit.

The third is governance standards with real teeth. This is where regulation earns its keep: mandatory testing and certification before high-risk AI systems go live, just like how new drugs are approved before they hit the market; explainability requirements for decisions that materially affect people’s lives, in health, credit, hiring and insurance; independent audits before deployment, not after a failure; mandatory incident reporting when AI systems cause harm; and provenance standards – both for the data used to train a model, and for labelling AI-generated content, so disinformation carries a fingerprint.

Singapore has already begun building some of this through AI Verify, the world’s first government-built AI testing toolkit, which runs technical audits for fairness, robustness and explainability.

The fourth lever is less often discussed and it deserves equal billing: government as accelerator, not just referee.

Singapore already has a working template for this in a different industry. The Monetary Authority of Singapore’s regulatory sandbox lets fintech firms trial novel financial products in a controlled environment, with relaxed rules and close supervision, before those products are cleared for the wider market.

The same logic applies neatly to AI: a sandbox lets a healthcare AI tool, an autonomous vehicle system or a new financial advisory bot be tested against real conditions with oversight, without either strangling it in pre-emptive regulation or unleashing it on the public untested.

Done well, sandboxes let the state do something nationalisation never could: speed up safe innovation instead of slowing it down.

Baig and Annett are right that AI is no ordinary technology, and that the state cannot simply watch from the sidelines while a handful of firms build systems capable of systemic harm. But the nuclear analogy that makes their argument so persuasive is also what limits it: Uranium can be fenced, and models and applications cannot.

Full nationalisation is the wrong tool for almost the entire AI economy – and actively harmful for an economy like Singapore’s, built on the private, competitive, applications-driven layer of the stack.

What the state can and must do is hold the keys to the small number of systems capable of a meltdown, referee everyone else with real standards and real teeth, and – through tools like the sandbox – help the good ideas move faster than the bad ones. Ownership was never really the point.