Every enterprise now claims to be "doing AI." Almost none can prove it is paying off. McKinsey research puts AI adoption at 88% of organisations globally, yet fewer than 10% have scaled agentic AI to deliver measurable value in any single function. A 2026 Forrester study commissioned by FPT found that only 26% of enterprises consider themselves advanced in operationalising AI — even as most have already committed real budget to it. The pattern holds across every major research house tracking the space this year: adoption is universal, but AI at scale remains rare.

At ATxEnterprise 2026 in Singapore, five sessions dissected exactly why that gap exists — and what closing it will require. Drawing on speakers from ST Telemedia Global Data Centres, Pfizer, DBS Bank, TuringData, NTT DOCOMO, Singtel, DayOne Data Centres and BDx Data Centres, a clear picture emerged: the barrier to enterprise AI scale is not ambition, and it is rarely the model. It is AI infrastructure — the compute, data infrastructure, connectivity, hyperscaler partnerships and talent that AI needs to run reliably at production volume, day after day.

Want the full picture from this year's event? Download the ATxEnterprise 2026 post-show report.

The scale gap, by the numbers

The session "Powering the AI Surge: Are We Infrastructure-Ready?", led by MC Lim of ST Telemedia Global Data Centres, put a number on the disparity that will sound familiar to anyone reading the external research: 88% of organisations globally have integrated AI into at least one business function, but only 32% have successfully scaled it. Research spanning nine Asian markets found that AI readiness rests on five dimensions — strategic alignment, organisational readiness, data governance, existing digital infrastructure, and future planning — and that most enterprises are underfunding exponential AI demand with incremental, business-as-usual budgets.

The infrastructure shortfall is stark. Ninety-nine per cent of organisations in that research reported insufficient compute capacity for AI workloads, and half said fewer than a quarter of their IT locations were AI-ready. On the talent side, 52% lacked in-house expertise to manage AI infrastructure, and 62% faced broader AI talent shortages. This lines up closely with a 2026 UST enterprise survey, which found 86% of leaders feel ready to scale AI enterprise-wide, while 44% are simultaneously blocked by data quality — a confidence-capability gap that mirrors what was heard on the ATxEnterprise stage.

Singapore was repeatedly cited as the regional exception rather than the rule — an example of what AI-ready infrastructure looks like when power, policy, and partnerships align. Singapore's AI infrastructure build-out includes multi-operator AI testbeds and direct partnerships with global AI leaders, and CBRE's 2026 Asia Pacific Data Centre Trends report shows why: the city-state has two new tranches of data centre capacity totalling 1.2GW coming online, positioning Singapore's AI infrastructure as a premium hub for high-density AI workloads even as larger volume requirements push into neighbouring Southeast Asian markets.

Data infrastructure is the real bottleneck

If compute is the headline constraint, data infrastructure is the one enterprises consistently underestimate. In "Fixing the Data Bottleneck," Gautam Gupta of Pfizer, Dr Puay Guan Goh of the National University of Singapore, and Garry Steedman of DBS Bank argued that the definition of "clean data" has fundamentally shifted — it now has to include metadata, semantic layers and business context, not just structure, in order to feed AI systems at all. Without that layer of data infrastructure, enterprises cannot build the observability and governance needed for traceable, trustworthy AI, particularly in regulated sectors like banking and pharmaceuticals where data privacy is non-negotiable.

That same data infrastructure gap showed up from the vendor side in "TuringData: The Unified Data Platform Powering the AI Factory Era." Nikhil Madan, VP of AI Infrastructure at TuringData, described an industry so fixated on GPUs that it has neglected the data infrastructure needed to feed them — up to 70% of GPU capacity sits idle simply waiting for data. TuringData's answer is lightweight, scalable data infrastructure built for the AI factory era: platforms starting at half a terabyte per second and scaling to exabytes, built in partnership with Nvidia, Dell and Lenovo. Southeast Asia, Madan noted, is one of the fastest-growing adopters of this AI infrastructure globally, but is also driving demand for sovereign AI shaped by local regulatory and linguistic needs — reinforcing that data infrastructure decisions in Asia can't simply be copied from the US or Europe.

Trust, identity and the AI-native network layer

Infrastructure readiness isn't only physical. In "Building AI Native Digital Infrastructure," Hiroki Kuriyama, President and CEO of NTT DOCOMO GLOBAL, argued that as AI shifts from an augmentative tool to an autonomous actor, the next layer of AI infrastructure has to be trust itself. His proposed Universal Wallet Infrastructure — a cross-border, cross-industry platform for verifiable identity, skills and asset credentials — is a direct response to fragmented digital identity systems that can't keep pace with AI-native operations. Practical deployments already underway include 5G infrastructure partnerships expanding broadband access in Indonesia and a workforce-credentialing pilot in Japan with Pearson. Kuriyama's core point applies well beyond telecoms: enterprise AI at scale requires infrastructure that both humans and AI agents can trust, not just infrastructure that is fast.

Hyperscaler partnerships and the physical backbone

None of this AI infrastructure gets built without the data centre operators and hyperscaler partnerships underpinning it. "Collaborating with Hyperscalers: Building the Infrastructure Backbone for AI and 5G" brought together Jamie Khoo (DayOne Data Centres), Manoj Prasanna Kumar (Singtel Digital InfraCo), Virat Patel (Pioneer Consulting Asia) and Sujit Panda (BDx Data Centres) to unpack what it actually takes to build AI infrastructure at hyperscale in Asia. Singtel described a full-stack approach spanning data centres, sovereign AI cloud, and connectivity via satellite, subsea cable and fibre — positioning itself as a fixed-mobile convergent operator built for both AI and 5G infrastructure demand. DayOne Data Centres, operating across eight markets including Singapore, Johor and Batam, pointed to a broader industry shift towards 100MW-plus facilities and, increasingly, gigawatt-scale campuses.

The panel's numbers illustrate how fast density requirements are moving: rack power draw is evolving from 20kW to 500kW within a few years, forcing a shift to advanced liquid cooling and new electrical and safety standards, with hyperscalers now demanding 100MW facilities delivered within 12 months. External data backs up the urgency — CBRE's 2026 outlook puts Asia Pacific data centre investment at a record US$11.6 billion in 2025, with global hyperscaler capex on AI infrastructure rising 61% in 2026 alone. Land, power and water scarcity, plus a persistent shortage of skilled construction and modular-design talent, were flagged across the panel as the binding constraints on how fast that hyperscaler-backed AI infrastructure can actually be delivered — echoing the same talent gap raised independently in the data bottleneck and infrastructure-readiness sessions.

Curious what else came out of the show? Browse all ATxEnterprise 2026 session summaries.

What leaders need to fix before 2027

Read together, the five sessions point to the same conclusion from five different vantage points. Enterprise AI scale is not primarily a model problem, a budget problem, or even an ambition problem — leaders across every function say they want to scale AI. It's an AI infrastructure problem, made up of four compounding gaps:

1. Compute and physical capacity. Insufficient AI-ready compute and data centre capacity remains the single most cited constraint, and hyperscaler partnerships building 100MW-plus, liquid-cooled facilities are racing to close it.
2. Data infrastructure. "Clean data" now means governed, contextualised, AI-ready data infrastructure — not just structured records — and most enterprises haven't rebuilt for that standard.
3. Trust and identity infrastructure. As AI moves from augmentative to autonomous, AI infrastructure has to extend into verifiable, cross-border trust frameworks for both human and AI identity.
4. Talent. Every single session, independently, flagged a shortage of people who can build, govern or operate AI infrastructure at scale — a gap that shows up as clearly in Singapore's hyperscale data centres as it does in enterprise IT departments.

Singapore's position as a regional AI infrastructure hub — underpinned by national investment, hyperscaler partnerships, and Southeast Asia's fastest-growing data infrastructure pipeline — gives Asian enterprises a genuine advantage heading into 2027, provided they treat infrastructure as a strategic determinant of AI leadership rather than a line item behind the model. For enterprises operating in or through Singapore especially, the organisations that move from experimentation to execution on AI infrastructure now will be the ones with a real answer, next year, to the question every board is already asking: we've adopted AI — so why hasn't it scaled?

See where AI infrastructure is heading next. Pre-register for ATxEnterprise 2027.

This article draws on official AI-generated session summaries from ATxEnterprise 2026, Asia Tech x Singapore's enterprise tech conference held at Singapore EXPO.