The paradox hiding in Singapore’s own numbers

Singapore’s enterprise AI story looks, on paper, like a triumph. According to ServiceNow’s 2026 enterprise AI maturity research, the country’s AI maturity score climbed 19 points this year, recovering sharply from a 2025 slump, and agentic AI adoption among local enterprises more than doubled year on year. Separate research from IDC, commissioned by Microsoft, found that more than six in ten Singaporeans now use generative AI in some form, ahead of the regional average.

Yet buried in that same IDC dataset is the detail almost nobody led with: manufacturing, the sector where “industrial AI” is supposed to matter most, has the lowest near-term intent to deploy agentic AI of any sector Singapore tracks, with only a small single-digit percentage of manufacturers planning to roll it out within two years. A country celebrated for AI enthusiasm has a factory floor that is, comparatively, dragging its feet.

That is not a contradiction so much as a signal. Consumer and knowledge-work AI can be adopted by an individual opening an app. Industrial AI cannot. It has to be wired into machinery, sensors, safety systems and decades-old operational data before it does anything useful — and that is a fundamentally slower, more expensive kind of adoption than the headline statistics about chatbot usage suggest.

What actually changed this cycle: infrastructure, not intelligence

The most consequential industrial AI news of the past year has had almost nothing to do with model capability. It has been about power and physical footprint. Analysis from BloombergNEF puts the combined capital expenditure of the world’s largest data centre operators at close to $750 billion in 2026, up from under $450 billion the year before, with more than 23 gigawatts of capacity under construction globally. McKinsey’s separate modelling projects that meeting global AI compute demand could require $5.2 trillion of infrastructure investment by 2030.

Individual projects illustrate the scale shift. Energy company Envision has begun commissioning a gigawatt-class “Galaxy Campus” in Ulanqab, Inner Mongolia, explicitly framed as a template for gigawatt-scale AI infrastructure rather than a conventional data centre expansion — a sign that the unit of measurement for industrial AI has moved from server racks to power-plant equivalents. One gigawatt, for context, is roughly the output of a large nuclear reactor, and five facilities of that scale or larger are expected to come online globally in 2026 alone.

Singapore is not exempt from this physical reckoning. The country already holds Southeast Asia’s largest live data centre footprint at an estimated 1.46 gigawatts, but government policy has deliberately throttled new capacity: only a fraction of the near-980-megawatt development pipeline is currently under construction, gated by a Green Data Centre Roadmap that ties new capacity to strict power-efficiency targets. That is a very different posture from markets like the US, where data centre growth has already pushed up household electricity bills in some regions. Singapore is choosing to grow industrial AI infrastructure slower, and greener, than raw demand would otherwise dictate.

The money behind the machines got more creative

Industrial AI’s next phase is also being financed differently. SoftBank recently secured a reported $10 billion loan backed by its stake in OpenAI, using AI equity itself as loan collateral rather than the underlying data centre assets. That is a meaningful shift: it treats a stake in a frontier AI lab the way a bank might once have treated property or plant equipment — as bankable, leverageable collateral. Expect more of this. As industrial AI infrastructure balloons into the hundreds of billions of dollars, the capital increasingly has to come from novel debt structures, not just corporate cash reserves.

Meanwhile, enterprise software providers built around applied AI are starting to show the kind of earnings that give industrial buyers confidence to commit budget. Palantir’s recent results have been read by analysts as evidence that AI-driven software spending is translating into durable revenue for the sector, not just pilot-stage experimentation. For a manufacturing plant manager deciding whether an AI vendor will still exist in three years, that kind of proof point matters more than any product demo.

Singapore’s actual answer: governance before scale

Rather than chase adoption numbers, Singapore’s institutions have spent 2026 building the plumbing that lets industrial AI be trusted, not just tried. Singapore Polytechnic has been named founding training partner for the TIA-Ready Framework, described as the country’s first industry-owned standard for industrial AI readiness, built to help manufacturers assess risk and governance before wider deployment in automation and operational technology settings.

That framework sits alongside A*STAR’s Sectoral AI Centre of Excellence for Manufacturing, which has already paired real factories with research partners — precision plastics manufacturer Sunningdale Tech, for instance, worked with the centre to deploy AI-powered defect detection on its production lines, according to EDB’s own account of the programme. Budget 2026 backs this with real money: S$150 million for an Enterprise Compute Initiative pairing companies with cloud and AI providers, on top of a S$37 billion research and innovation envelope, coordinated by a National AI Council chaired by the Prime Minister with advanced manufacturing named as one of four national AI missions.

The pattern across nearly every credible Singaporean case study is the same: manufacturers are told to fix their data and governance first, deploy narrowly, and only then scale. That is the opposite of the “buy the model, figure it out later” instinct that has characterised a lot of consumer AI adoption — and it is a large part of why manufacturing’s adoption curve looks slower even as the country’s overall AI maturity score is rising.

Redesign beats bolt-on

Cross-referencing Singapore’s own enterprise data makes the underlying lesson explicit. The largest group of Singapore enterprises — roughly a third — are still simply using AI to assist individual employees with day-to-day tasks. Only around one in ten have gone further and redesigned a multi-step business process so that AI completes it end-to-end. But it is that smaller group that captures the disproportionate share of the benefit: enterprises that rebuilt a workflow around AI, rather than layering a tool on top of the old one, were reported to be nearly four times more likely to see meaningful productivity gains.

Applied to a factory floor, this is the difference between adding a chatbot to a maintenance team’s laptops and rebuilding the inspection workflow so that computer vision, sensor data and scheduling are integrated end-to-end — the Sunningdale Tech model, not a generic co-pilot rollout.

What this means if you run a business, not a data centre

Most readers of a piece like this are not building gigawatt campuses or negotiating equity-backed loans. But the underlying shift still applies at a much smaller scale, whether you run a workshop, an e-commerce warehouse, or a small manufacturing line. The next phase of industrial AI rewards three things that have nothing to do with which model you use: clean, standardised data before you automate anything; a governance step — even an informal one — before you plug AI into anything operational; and a willingness to redesign one workflow properly rather than sprinkle AI across many.