In FN·015 I wrote that Singapore doesn't train frontier models. It hosts, deploys and regulates them, and standards set elsewhere arrive here as operating conditions.
I meant it as a point about governance. A year interning in New York has made it a point about economics too. What separates places isn't only where a technology is made. It's how quickly it becomes ordinary work. That question has a name: diffusion.
Invention is not the advantage
Jeffrey Ding's Technology and the Rise of Great Powers (2024) argues that the countries that pulled ahead in past industrial revolutions weren't usually the ones that invented the key technology. They were the ones that spread it widely.
In the second industrial revolution, Britain and Germany had stronger inventors and research centres. The United States pulled ahead anyway, by spreading interchangeable manufacturing through every industry that could use it, with land-grant universities training engineers across the country.
Arvind Narayanan and Sayash Kapoor make the matching point about AI in AI as Normal Technology (2025). The speed limit isn't the models. It's how fast organisations, regulators and people can change around them. Capability moves on a lab's release schedule. Adoption moves on an institution's.
Building the best model is one competition. Absorbing good-enough models is another, and it's one a small country can actually enter.
What Singapore and ASEAN actually are
On at least one measure, Singapore is among the most AI-intensive places in the world. In Anthropic's Economic Index (September 2025), it came second globally on per-capita use of Claude.ai, at about 4.6 times what its working-age population would predict, behind only Israel.
That number counts individuals choosing to open a chatbot. It says nothing about whether their schools and employers had caught up, and when I left, mostly they hadn't.
The state has been building around this for years: AI Verify since 2022, SEA-LION for Southeast Asian languages the frontier models handle badly, and in Budget 2026 a National AI Council chaired by the Prime Minister, regulatory sandboxes, and a 400% tax deduction on up to S$50,000 a year of AI spending.
The region is eleven economies, since Timor-Leste joined in October 2025, with very different starting points. Its shared AI instrument, the ASEAN Guide on AI Governance and Ethics (2024), is voluntary and light enough for Singapore and Laos to both endorse. The physical build-out is moving faster. Singapore paused new data centres from 2019 to 2022 because they were taking too much of its scarce power and land. Johor took the overflow, and Malaysia has since stopped approving data centres that aren't AI-related.
So the region comes in three layers: a high-adoption, high-governance hub; neighbours with the power and land; and a large, young, multilingual population that will meet AI mostly as consumers, and as workers whose jobs are being repriced.
| Layer | Who | What it carries |
|---|---|---|
| Hub | Singapore | Adopts and governs: individual use, frameworks and sandboxes, regional models |
| Host | Johor and Malaysia | Power and land: data centres, rationed approvals, the ASEAN Power Grid |
| Exposed | The wider region | Consumer and worker: Philippine IT-BPM, language coverage, cross-border data |
What diffusion looks like up close
The people I've met here disagree about almost everything to do with AI except one thing: nobody thinks it's going away.
The clearest picture of how it spreads came from comparing notes with my CEO. He uses it to hold a wide job together: meeting prep, agendas, summarising analytics across business units. I use it for code and research, and for the upkeep of my own day: my to-dos, and the Obsidian knowledge base Claude helps me run. It's the same tool doing two different jobs. Nobody rolled it out to either of us. It spread through individuals at both ends of the organisation at once.
The data agrees. In Microsoft's 2024 Work Trend Index, three in four knowledge workers were already using generative AI at work, most of them with their own tools. In May 2026, the US Census Bureau found fewer than one in five American firms, 19.8%, using AI in any business function.
| Measure | Share | Source |
|---|---|---|
| Knowledge workers using it | 75% | Microsoft and LinkedIn, Work Trend Index, 2024 |
| US firms using it | 19.8% | US Census Bureau, BTOS, May 2026 |
The surveys count different things (people in one, firms in the other) in different years, so read the gap as a shape, not a number. Individuals adopt fast, firms adopt slowly, and most of the real change happens in the gap between them, invisible to the official statistics.
I see the other half from the inside, because part of my job is the brake. In GRC and AI compliance, the question isn't whether a model is capable. It's what we're allowed to send it. We work to SOC 2 controls, and data sovereignty comes up more than anything: whose data this is, where it goes, which parts of it a model may see.
That's where Narayanan and Kapoor stopped being abstract for me. The model was ready months ago, and the organisation is still working out what it's allowed to do with it. That isn't a failure. That work is the diffusion.
Outside work, I expected to sort my friends in consulting, finance, AI safety research and public service by who used AI most. That turned out to be the wrong question.
Nobody I met was asking whether. They were asking how: how to get the most out of it without being careless about what it costs.
Two engines
Put New York next to home and you get two engines driving the same diffusion.
In the American one, individuals move first, firms add the permission layer afterwards, and the state arrives last and lightly. It's fast and uneven, and the costs land wherever they land: on the workers whose tasks go first, and on questions of data and liability nobody settled in advance.
The Singaporean engine runs the other way. The state sets direction, publishes frameworks, funds the upskilling and brings firms in through incentives. SkillsFuture is a country treating diffusion as a public good.
From the inside, it felt like caution. When I left for New York, my university, Singapore Institute of Technology (SIT), was split. One camp of students and professors was all in. The other worried that leaning on AI would wear down the abilities a degree is meant to build. Neither camp was wrong, but Singapore was clearly the more cautious place.
That's Singapore's habit: move deliberately, then move at scale. It worked for the internet and the smartphone, which took the better part of a decade to become ordinary. AI hasn't offered that time. By UBS's widely cited estimate, ChatGPT reached 100 million monthly users about two months after launch.
The worry I've brought with me isn't whether Singapore is careful enough. It's whether careful still works at this speed.
Can we keep our edge and spread the technology safely at the same time? The lesson isn't to become more American. The discovery is already happening at home, in the usage numbers and in the students who went all in before their professors had decided anything. What the state-led engine needs is a way to learn from it, fast enough that the framework follows the practice instead of guessing at it.
What it means past the Causeway
For most of ASEAN, diffusion arrives first as a question about work.
The Philippines is the sharpest case. Its IT and outsourcing sector employed about 1.9 million people in 2025 and earned around US$40 billion, much of it from work language models are getting good at first: customer support, transcription, back-office processing. These economies didn't invent the technology, and they're among the first to have their work repriced by it.
But exposure and absorption are the same position seen from two sides. A workforce that size, already inside other companies' workflows, is also among the best placed to learn the tools first and sell that skill. Which way it goes depends less on the models than on how fast its firms and workers absorb them.
Language matters too. A model that handles English well and Javanese or Tagalog badly spreads unevenly by design. That's why SEA-LION, Thailand's Typhoon and Indonesia's Sahabat-AI matter, and AI Singapore helped train the last of those.
So does data. The question my team asks every day, what we can send the model, has a regional version: what data can cross which border. The Digital Economy Framework Agreement, due to be signed at the ASEAN Summit in the Philippines this November, is ASEAN's most ambitious attempt to line up eleven answers.
ASEAN won't produce a single strategy. Singapore's part is to make sure the lessons it learns from deployment don't stay inside Singapore.
Where this points
The thing I'll take home is a change in what I look at.
Before New York, I'd have judged a country's AI position by its compute, its frameworks and how close it sat to the labs. Now I'd ask three other things. How much informal adoption is nobody counting? How long does an institution take to catch up with its own people? And who absorbs the costs while that gap stays open?
None of those has a good number yet, and that's the work I want to keep doing: better evaluations and measurement tools for how AI is actually landing in Singapore and across ASEAN, so the people drafting governance and safety standards work from evidence. Singapore is small enough to measure this properly and trusted enough to be cited. That's how a country that doesn't build frontier models becomes a stakeholder in how they're governed: as a responsible adopter whose evidence others use.
It's the same instinct that drew me to carbon accounting. You can't govern what nobody reports.
Open questions
- How do you measure diffusion that nobody declares?
- Firm surveys miss what workers do on their own, and nobody has a good unit yet.
- Can a university set norms at the speed of its own students?
- At SIT, the students had decided long before any policy did.
- Can a state-led system make room for permissionless discovery?
- Is a sandbox permission-free enough to find what nobody planned?
- Who reprices the first jobs, and who has a say?
- In the Philippines, the repricing is already under way.
- What does Singapore owe the region?
- Is absorbing AI well a national advantage to keep, or a regional good to share?