Over the past year, we've spoken with customers across manufacturing, financial services, telecommunications, and the public sector who are all asking a similar question: what comes after the AI pilot?
Getting a model to work in a controlled environment is one challenge. Running it dependably across real business operations is another entirely.
This shift is changing the types of questions our customers ask. Discussions used to start with how much computing capacity was available. Now they often move quickly to topics like where the data sits, how much heat the system generates, whether the network can keep up, and who will manage the environment once the pilot is no longer a pilot.
Simply adding more GPUs doesn't address the deeper problems. Organisations need to build a deployment foundation that enables teams to use AI repeatedly without creating a new support burden every time they scale a use case.
Why AI workloads require reimagining infrastructure design
Pilots can often get by with borrowed capacity, isolated data, and a small technical team. That changes quickly when AI is expected to support complex operations like a factory line, a hospital operation, or a customer service process. Weak points that were manageable in testing become much harder to ignore.
Organisations are often surprised that the root problem isn’t what they expect. It’s not uncommon for a team to start by worrying about compute capacity, only to discover that storage throughput was the limiting factor.
Others find that cooling, networking, or operating processes create the bigger constraint. This complexity is compounded by the fact that the bottleneck tends to shift as projects mature.
Through our work, we’ve found that priorities vary across the Asia-Pacific region. For example, conversations in India generally emphasise growth and capacity expansion. In Singapore, reliability and responsiveness tend to come up earlier. These organisations are not all solving the same problem, but they are asking more practical questions than before.
When projects do fall behind, people are often quick to blame the model. However, it’s often the case that problems only become visible after real-world implementation commences. Handling this effectively can influence success as much as the technology itself.
Where AI projects begin to slow down
Most enterprises no longer need convincing that AI has value. The harder question is what happens after the first successful test. Can the system connect with existing tools? Is the data good enough? Are security teams comfortable with the way it operates? Can business teams depend on it when conditions change?
- Paul Ju, Senior Vice President, ASUS and Co-head of Infrastructure Solutions Business Group.
A few years ago, power and cooling rarely featured in early conversations about AI. Today, they often come up much sooner, especially as organisations plan higher-density systems. Integration is becoming more complex.
AI rarely runs in isolation; it must work with existing applications, data platforms, security controls, and operational processes. Data is still scattered across many organisations, which may slow projects down and make it harder for teams to maintain quality as use cases expand. And, as AI extends across data centers, cloud platforms, and edge locations, it becomes harder for teams to maintain a clear view of what is running where.
Taiwan offers a useful example. Technical discussions can look very different in areas like smart manufacturing and healthcare. A factory may care most about response time on the line. Meanwhile, a hospital may focus on data protection and reliability. High-performance systems are now built for dense AI workloads that require better cooling and more efficient energy use. This shows how scientific computing and industrial AI are coming together, and consistent performance is now just as important as reaching top speeds.
Likewise, in Vietnam, growing digital infrastructure and increased use of cloud services are driving more investment in AI-ready systems. These are built to run nonstop in areas like manufacturing, finance, and digital services. The focus is moving from adding more capacity to ensuring systems are reliable and always running.
This is where the gap between a good demonstration and a dependable service becomes clear: real value is created only when the system continues to operate reliably as demand rises, data changes, or something in the underlying platform needs attention. This is where many organisations underestimate the change. Scaling AI affects how teams work, how systems are supported, and how decisions are made after the first project goes live.
Building a future-ready AI infrastructure
Hybrid and distributed models are often discussed as architecture choices. For customers, they are usually more practical than that. Some data cannot move freely. Some systems need to sit close to users, machines, or devices. Other tasks belong in the cloud because flexibility matters more than proximity.
Five years ago, very few of our AI discussions included cooling strategies. That’s beginning to change, but the wider point is planning. A faster computing platform is useful only if you consider energy use, heat, serviceability, and long-term operating requirements from the start.
Another change we’ve noticed across the region is that infrastructure conversations are happening much earlier than before. A few years ago, many organisations focused first on
use cases and addressed the computing environment later. Today, more customers want to understand both at the same time.
The focus is no longer on how quickly a pilot can be launched. Customers increasingly ask how AI fits into their existing operations, such as long-term support, integration, internal controls, and the people responsible for keeping systems running.
We also see growing interest in approaches that let work move between cloud, data center, and edge environments as business requirements change. That flexibility matters most in industries where response time, data location, or continuity are central to the service, such as healthcare, financial services, manufacturing, retail, and public services.
No single company can solve all this alone. In many cases, customers need partners who understand the local operating context and can help connect hardware, software, services, and support into something that works for their market and their use case.
At the same time, trust has become a much larger part of the conversation. Questions about data sources, monitoring, accountability, and security are often raised much earlier than they were even a few years ago.
The most useful AI projects are usually tied to a specific process, customer need, or operational pain point. Projects that remain separate from how the organisation actually works will struggle to move beyond pilots. Those that make daily work faster, safer, or more consistent are easier to justify and expand.
For Asia’s enterprises, the opportunity is clear, but it will reward discipline as much as ambition. AI's long-term impact will not be determined by how many pilots an organisation launches. Its success will be determined by whether those projects become part of day-to-day operations in a practical and sustainable way.
For many organisations across the region, that work is only beginning.
Paul Ju is ASUS Senior Vice President and Co-head of Infrastructure Solutions Business Group.
