Asia Pacific enterprises are increasing artificial intelligence spending, but infrastructure integration, data location and connectivity remain constraints on turning those budgets into repeatable business value.
In this emailed TNGlobal Q&A, Serene Nah, Managing Director and Head of APAC at Digital Realty, discusses workload placement, hybrid infrastructure, data sovereignty, interconnection and the metrics enterprises should use to judge whether AI infrastructure investment is improving economics rather than simply adding capacity.
Nearly six in ten APAC respondents expect AI investment to rise by more than 25 percent, yet only 16 percent report currently realizing AI ROI. What is creating the biggest gap between spending and measurable returns?
The findings point to a growing infrastructure challenge as businesses put more investment behind AI.
Across APAC, 92 percent of respondents now tie their data-location strategy to their AI plans and 88 percent already operate with a distributed approach to data. Yet only 22 percent describe their hybrid infrastructure as fully integrated.
That matters because AI does not operate in isolation. Workloads need access to data, compute, clouds and networks that may sit across different environments and markets. The more distributed that architecture becomes, the more important decisions around data location, connectivity and workload placement become.
For businesses, the question is therefore not only how much they invest in AI, but whether the infrastructure behind that investment allows workloads to access the data and services they need reliably, securely and efficiently. Closing that gap will be an important part of turning greater AI investment into repeatable business value.
Australia appears further ahead on AI ROI than the regional average. What infrastructure or organizational differences do you think explain that gap, and how cautious should we be about comparing markets directly?
While we can’t attribute Australia’s position to a single factor, the findings do show quite clearly that APAC markets are at different stages of their AI journeys. In Singapore, infrastructure readiness features more prominently as a challenge. In Japan, compliance is a greater consideration, while South Korean respondents place more emphasis on interconnection.
For businesses, this means that instead of looking for a single APAC model, infrastructure decisions need to reflect an organization’s AI maturity, the requirements of the workload and the operating environment in each market. What works in Australia may not be the right model for Japan, Singapore or South Korea.
Half of Singapore respondents cite a lack of specialized AI infrastructure as a challenge. What is actually missing most often: power and capacity, accelerators, cooling, interconnection, cloud access, skills, or something else?
AI infrastructure readiness is less about any one component and more about whether the full environment can support the workload.
For example, an enterprise may have access to compute, but that compute still needs sufficient power and cooling, reliable connectivity to enterprise data and cloud environments, and the resilience and security needed to operate reliably. As AI workloads become denser and more complex, these requirements become increasingly interdependent.
That is why the starting point should be what the workload actually requires. An inference application that needs to respond in real time will have different infrastructure requirements from a large-scale training workload.
We are also seeing customers want greater confidence that these different components will work together before committing to deployment. This is part of the thinking behind the Digital Realty Innovation Lab, where customers can validate AI and hybrid-cloud designs in a live environment before scaling them more widely.
Across APAC, 88 percent say they use a distributed data approach, but only 22 percent describe hybrid infrastructure as fully integrated. Where do those integrations most often break down?
The challenge is connecting where the data sits with where the workload needs to run.
An enterprise may have data across its own infrastructure, multiple clouds and different markets. An AI workload then needs to access that data securely and efficiently, sometimes across jurisdictions with different requirements.
That creates several potential points of friction: connectivity between environments, latency, security controls, data movement and governance. As AI becomes more distributed, these are increasingly infrastructure questions rather than simply application questions.
The aim should be to give enterprises flexibility over where they place workloads without creating separate infrastructure islands every time they enter another cloud or market.
How should enterprises decide where to place training, inference and data-intensive workloads when latency, cost, data residency and resilience point toward different locations?
The starting point should be the workload. Training and inference illustrate why. Training generally places a greater premium on access to large amounts of compute and power, giving enterprises more flexibility over location. Inference can have a different profile. When an AI application needs to respond in real time, proximity to users, enterprise data and networks becomes much more important.
Data location also needs to be considered alongside those infrastructure requirements: where can the data legally reside, how frequently does the workload need to access it, and what are the latency, resilience and security requirements?
Those questions may point to different locations for different workloads. For a regional enterprise, the right answer is unlikely to be putting everything in one place. It is building an architecture where workloads can run where their requirements are best met, while remaining connected to the wider enterprise environment.
What changes when a company operates across several APAC jurisdictions with different data-sovereignty and compliance requirements? How can it avoid building isolated infrastructure in every market?
Operating across multiple jurisdictions means enterprises need to be more deliberate about where data and workloads reside. Performance and cost remain important, but businesses also have to consider whether particular data can move across borders and what local governance requirements apply.
This does not mean every market needs to become a separate infrastructure island. A more practical model is to localize the data and workloads that need to remain in-market, while connecting them into a wider regional architecture.
We are building around this model across Asia Pacific. In Malaysia, for example, customers can deploy infrastructure locally in Cyberjaya while connecting into other regional hubs such as Singapore and Jakarta through Digital Realty’s wider platform. Similarly, our Jakarta campus combines local infrastructure and interconnection with access to our broader regional and global ecosystem.
For regional businesses, this creates greater flexibility to use different markets according to their regulatory, capacity and workload requirements without treating each deployment as a standalone environment.
South Korean respondents place unusual emphasis on interconnection. As AI architectures become more distributed, where does network design become as important as compute capacity?
When the components supporting an AI workload are distributed, the network becomes part of the workload’s performance.
Compute may sit in one environment, enterprise data in another and cloud services somewhere else. Having sufficient computing capacity is therefore only part of the equation. The workload also needs to reach the data and services it depends on quickly, reliably and securely.
That becomes even more important for inference, where latency can directly affect the user experience, and for AI applications that need to draw on enterprise data in real time.
As AI architectures become more distributed, enterprises will increasingly look at compute and connectivity as one infrastructure decision rather than two separate ones.
How should enterprises think about avoiding infrastructure or cloud lock-in while still getting the performance benefits of tightly integrated AI stacks?
The priority should be to balance optimization with flexibility. Enterprises will naturally optimize individual workloads around the infrastructure, cloud or technology stack that gives them the performance they need. But AI is evolving quickly, and the requirements of those workloads will change.
That makes connectivity and portability important architectural considerations. Enterprises should ask whether their data and workloads can connect across different infrastructure environments and locations without having to redesign the architecture each time requirements change.
This is particularly relevant in APAC, where capacity, regulation and data-location requirements vary between markets. The goal should be enough integration to optimize performance without designing an environment that limits future choices.
Which metrics should technology leaders track to know whether infrastructure investment is actually improving AI economics rather than simply adding capacity?
At the infrastructure level, enterprises can look at metrics such as utilization, workload latency, deployment time, availability and the cost of running or moving workloads and data. These indicate whether infrastructure is being used efficiently and improving the performance of AI workloads.
The next level is connecting those measures to business outcomes. Are more AI use cases reaching production? Is the time required to deploy them falling? Can the organization support greater AI usage without costs increasing at the same rate?
Taken together, these measures give technology leaders a better indication of whether additional infrastructure is making AI deployment more efficient and reliable, rather than simply increasing the amount of available capacity.
Over the next 12 to 24 months, what infrastructure constraints do you expect to become the most difficult bottleneck for APAC enterprises moving AI into production?
One of the biggest challenges will be bringing the different infrastructure requirements for AI together in the right places.
Power, compute and cooling capacity will remain important, but those resources also need to be available where workloads can meet their data-location, connectivity, latency and resilience requirements. In APAC, that becomes more complex because these conditions differ considerably between markets.
This is why infrastructure planning needs to consider these requirements together rather than treating power, compute, connectivity and data location as separate decisions. Optimizing for capacity alone can create another constraint if a workload ultimately cannot run where its data needs to reside or where its performance requirements can be met.
For enterprises, building flexibility into workload placement will become increasingly important. Different markets can support different requirements, but those environments need to remain connected as AI demand evolves.
Serene Nah is Managing Director and Head of APAC at Digital Realty.
Editor’s note: This Q&A has been lightly edited for clarity and TNGlobal house style. The substance of the interviewee’s responses has been preserved.
Share your perspective: TNGlobal welcomes contributed insights and expert commentary from across Asia’s technology and innovation ecosystem. Submit a contribution for editorial consideration, or explore more conversations in our TNGlobal INSIDER and TNGlobal Q&A and Interviews archive.
Automation review notes
Source: https://www.linkedin.com/in/serene-nah-7642466/
Prepared from Serene Nah’s completed emailed Q&A received October 5, 2026, with bio, LinkedIn profile and headshot subsequently completed. Copy-edit only. Survey percentages and market comparisons are Digital Realty research claims and should remain attributed. Featured image intentionally left null for legacy importer safety; headshot is available in the Gmail thread.