Investor scrutiny of a potential AI bubble has intensified over the past twelve months. Yet much of this debate remains deeply United States (US)-centric, overlooking how materially different the underlying dynamics are in other markets. Headlines have focused on the rapid expansion of large-scale AI training facilities, the sheer scale of capital being deployed, and growing questions over whether this capacity will be fully and profitably used. That framing, however, obscures important regional differences, with these concerns most pertinent in the US.

Applying this narrative to Asia-Pacific (APAC) risks obscuring the region’s fundamentally different market structure. In APAC, the AI and data centre story is less defined by massive AI Training capacity creation but by significantly dispersed AI Inferencing. This distinction matters. Understanding APAC’s growth trajectory requires a reframing of assumptions, away from US training-led cycles and toward the region’s real demand drivers and utilisation-led fundamentals.
 

The US Market: Signs Of Hype-Cycle Fatigue

In the US, AI continues to be viewed as the single strongest growth driver for data centres, with substantial capital being deployed into large-scale model training facilities, amid broadly bullish investor sentiment. Over the past year, however, signs of strain have begun to emerge, particularly around whether the pace of infrastructure build-out is aligned with near-term demand.

GPUs – the specialised chips used to train AI models and the most capital-intensive component of AI-focused data centres – sit at the centre of this debate. As expectations for AI growth accelerated, projected demand for GPUs rose sharply, triggering a rapid expansion of training-oriented capacity.

This build-out was enabled by abundant land availability, relatively permissive planning regimes and fast construction timelines, allowing large volumes of capacity to come online within nine to twelve months. While this supported rapid growth, it also meant that speculative capacity has been delivered faster than demand could absorb it.

As a result, some large training environments are operating below peak utilisation. Usage levels are below expectations and GPU-as-a-service lease pricing has declined by over 50 percent over the past two years. This reflects the inherently cyclical nature of training demand, which tends to peak during model development phases and taper once training is complete.

Capacity, however, was scaled on the assumption of sustained exponential growth, driven by a competitive race for marginal improvements in foundation models. Too many operators expanded simultaneously.

These dynamics have been amplified by a concentrated ecosystem in which a small number of dominant players often act simultaneously as customers, capital providers and suppliers.

Together, these factors have contributed to growing concerns about an AI “bubble”. But they are rooted in conditions that are highly specific to the US. Elsewhere, the picture is more nuanced.

 

The Fundamental Advantages For APAC Data Centres

The volatility now seen in parts of the US data centre market is driven by the stop-start nature of model training, not by any loss of confidence in AI’s long-term importance. This is not a useful comparison for Asia-Pacific, where demand is instead tied to inference; the real-world application of these trained AI models. This is the revenue-generating segment of the industry, focused on real-world deployment rather than speculative training capacity.

APAC operates under a fundamentally different set of demand drivers. While risks associated with AI training will persist, the requirement for data storage and processing is only set to grow alongside AI adoption because of inference. The core story in the region is how AI is commercialised and embedded into day-to-day enterprise activity.

APAC is home to 60 percent of the world’s population, with small and medium enterprises (SMEs) comprising 98-99 percent of firms in the region, employing around 68 percent of APAC labour, and generating around 40 percent of APAC GDP.  These enterprises are the primary end users of AI-enabled tools and the most immediate beneficiaries of productivity gains, underpinning sustained demand for inferencing capacity and data management. In contrast to the US, these tailwinds are structural rather than cyclical.

Investors should distinguish between the building of large language models and monetising them. Monetisation flows from inference and the deployment of AI to drive productivity gains, both of which are occurring first and at scale in APAC. These workloads generate consistent demand for capacity, led by SME requirements and government sponsored programmes; supporting lower vacancy risk and more stable utilisation.

APAC-based data centres also play a critical role in the global demand curve for machine learning and inferencing workloads. Reflecting a combination of local enterprise requirements, alongside global workloads, this demand is increasingly being offshored to lower-cost locations within the region. Over the medium to long term, as the data strain on SMEs across APAC continue to grow, AI inferencing is expected to represent the fastest-growing demand driver for data centres,  as large language models become increasingly available and the focus shifts towards practical productivity applications.

Beyond demand fundamentals, APAC is structurally insulated from many of the oversupply concerns evident in the US. Capacity across the region consistently lags demand, resulting in tighter utilisation rather than excess supply. This is driven in part by physical infrastructure realities. These facilities only operate efficiently when built in ‘cloud environments’, meaning demand is concentrated in established data centre clusters, which are often in major metropolitan areas like Singapore, Osaka, Mumbai and Seoul. As a result, development in these locations typically requires vertical, multi-storey facilities with longer construction timelines and complex regulatory processes. Persistent constraints on land, power and water act as natural barriers to the speculative overbuilding seen in the US, supporting more resilient utilisation and returns over the long term.

For investors, regional perspectives are central to identifying resilient, long-term strategies in the data centre sector and US narratives should not dictate global investment frameworks. Rather than viewing APAC through the prism of a global AI “bubble,” the region should be assessed on its own fundamentals: sustained, utilisation-led growth driven by monetisation and inference.