The Asia Pacific will account for about a third of global data centre demand by 2030, according to McKinsey’s research, driven by local AI adoption and a broader enterprise digitisation cycle.
However, across Asia Pacific (APAC), more than just building capacity, the AI conversation has revealed the limits of current data infrastructure.
While we are seeing how quickly AI and enterprise digitisation are reshaping infrastructure demand across the region, it is important to note that the roadblock now is not just about building more capacity. Power availability, hardware constraints, rising costs and deployment inefficiencies are already shaping what enterprises can realistically scale.
In a region where markets differ widely in digital maturity, regulation, connectivity and cost structures, a single infrastructure model will rarely be enough. Enterprises need to look beyond capacity and ask whether their data is governed and accessible enough to support production workloads.
Production demands are forcing enterprises beyond cloud-only architectures
When AI is embedded into different business functions, costs and reliability concerns turn what were primarily IT decisions into core business considerations. Exclusively depending on centralised hyperscale cloud for AI tasks is no longer viable. Instead, workloads should be executed where they provide the greatest operational benefit, whether that is in a private setting, on-premises, within the public cloud, or at the edge.
This is why enterprises across APAC are moving toward more distributed architectures. IDC predicts that by 2027, three out of four of enterprise AI workloads in Asia Pacific (excluding Japan) will run on hybrid, fit-for-purpose infrastructure to accelerate time to value while optimising performance, cost and compliance.
Hybrid and edge approaches allow organisations to bring compute closer to where data is generated, reducing latency and bandwidth costs while improving responsiveness for real-time applications.
However, this shift also signals that the challenge is about redesigning infrastructure to operate more efficiently across environments. Centralised systems built for training large models are not always suited for the complexity and scale of enterprise inference, where workloads must run reliably across multiple business functions, data environments, and operational requirements.
Infrastructure performance depends on the entire technology stack
A common misconception is that AI performance can be improved simply by adding more compute. More compute may help, but it will not fix a poorly designed stack. AI infrastructure is an interconnected system: compute, storage, networking, data pipelines, orchestration, governance and integration all need to work together.
This is why infrastructure performance issues can persist even after organisations increase cloud capacity or invest in GPUs. According to Cloudera’s Data Readiness Index 2026, more than a quarter of respondents across APAC indicated that their operational initiatives are often hindered by infrastructure performance issues.
These issues often emerge outside the compute layer, particularly when data cannot move efficiently, pipelines are poorly optimised, or systems cannot communicate across environments. In practice, AI performance is limited by the weakest part of the stack. To scale AI sustainably, enterprises must adopt optimisation strategies such as rebalancing workloads across different compute types or redesigning architectures to improve efficiency.
Poor data architecture is the biggest obstacle to scaling AI
The most important AI infrastructure issue is often the least visible: the data architecture.
- Remus Lim is Senior Vice President, Asia Pacific & Japan, Cloudera
Many organisations believe they have full control over their data because it is stored somewhere within their environment. In reality, control does not always mean usability. Data may sit across multiple silos, systems and formats. It may be duplicated, inconsistently governed or difficult for AI models to access in the right context of the organisation.
When this happens, AI systems spend more time reconciling and moving data than generating useful insights. What looks like a compute problem can actually be a data architecture problem. Organisations end up burning expensive infrastructure resources to compensate for fragmented data environments.
Without a consistent data architecture as a robust foundation, each new AI use case risks becoming another isolated project, rather than part of a reusable enterprise capability.
Treating data architecture as core infrastructure means to focus on open formats, common governance, trusted metadata and architectures that allow data to be accessed where it resides. This is the key to moving AI beyond individual pilots to become a repeatable source of enterprise value.
Data sovereignty requirements are redefining infrastructure decisions
At the same time, organisations in APAC must navigate an increasingly complex regulatory landscape. IDC notes that more enterprises are already experiencing moderate to significant disruption to their IT operations due to evolving data privacy, cybersecurity and AI regulations. This is pushing organisations toward what IDC describes as an era of “compliance by design”, where regulatory foresight is integrated into core technology strategies.
These concerns force enterprises to answer practical questions. Where does the data reside? Who can access it? Can it cross borders? Can it be audited? Can AI models use it without creating compliance or security risks?
These are no longer questions that can be answered after an AI system is deployed. They need to be designed into the architecture from the outset. This is why hybrid and distributed infrastructure models are becoming more important. By enabling data to be accessed where it resides, rather than moved unnecessarily, organisations can better align with sovereignty requirements while improving efficiency.
A rigid infrastructure model may work in one market but pose risks in another. A distributed and well-governed model, on the other hand, gives organisations more control over where workloads run, how data is used and how AI can be scaled across jurisdictions.
The enterprises that succeed will be those that fix data fragmentation, optimise the full technology stack and build for a distributed, regulated region. They will understand that AI success is not defined by who has the most compute, but by who builds the most efficient and trusted data foundations.
Remus Lim is Senior Vice President, Asia Pacific & Japan, Cloudera.
