Gartner predicts that two out of five of Agentic AI projects will be abandoned by 2027. This is not due to any flaws in AI itself, but because enterprises are forcing modern autonomy onto outdated systems.
In Asia Pacific specifically, IDC found that a similar number of organisations say their current architecture cannot support new AI application development without significant modernisation.
In other words, the real bottleneck to scaling AI is no longer intelligence itself, it’s the infrastructure it runs on. Imagine installing a high-performance engine into a decades-old sedan: it will start, it may even roar, but the moment you push it, the rest of the chassis cannot keep up.
This is where many APAC enterprises find themselves in 2026. After a decade of rapid cloud adoption, many organisations are now operating environments that are fragmented, inefficient, or constrained by legacy architectures.
In trying to accelerate AI, past infrastructure decisions are now having the opposite effect: slowing deployment, increasing cost, and limiting scale.
- Keith Lee, Cloud Business Director, Sangfor Technologies.
So what are the current challenges and what actually needs to change?
The reality of today’s infrastructure: what is holding AI scale back
The problem is not a single failure point. It is a system that has gradually drifted out of alignment with what AI now demands.
It starts with how infrastructure evolved. Over time, many enterprises adopted a multi-cloud strategy. What started as a promise of ultimate flexibility became a nightmare of complexity. Data is distributed across environments. Tooling is duplicated. Visibility is inconsistent.
For AI workloads, this creates immediate friction because data has to move constantly between systems, introducing latency, increasing cost, and slowing iteration cycles. What used to be manageable complexity becomes a barrier to scale.
At the same time, the legacy infrastructure that many companies once adopted was designed for stable, predictable applications. AI is the opposite: it is data-intensive, distributed, and highly dynamic.
Legacy systems struggle because they lack coordination between compute, storage, and networking, so data pipelines cannot keep up. Resources sit idle while waiting for other parts of the stack to respond.
Then there is the cost pressure, which is rising and becoming harder to control. Multi-cloud architectures increase costs primarily due to hidden data transfer fees, management complexity, and resource fragmentation.
At the same time, licensing shifts in virtualisation and infrastructure platforms are locking organisations into long-term, bundled subscription models. This creates a double burden: rising infrastructure spend and limited flexibility to optimise it. AI workloads, which are already resource-intensive, amplify this cost challenge even further.
Caught between these pressures, many organisations are now trying to modernise. According to Avasant’s benchmark data, over 70 percent of enterprises are now actively exploring mitigation strategies to offset cost escalations and maintain operational continuity.
But migration itself has become one of the biggest challenges, because it affects storage, networking, automation, and operational processes built over years. Compatibility becomes a critical issue. So does risk. As a result, many enterprises are stuck in transition, unable to fully move forward, but unable to stay where they are.
Taken together, these challenges explain the disconnect between AI ambition and execution. The infrastructure exists, but it is not built to support what enterprises are trying to do with it.
What an AI-ready cloud infrastructure chassis actually requires
Fixing this does not mean starting from scratch. It means addressing the fragmentation, rigidity, cost inefficiencies and migration risks that are holding AI back today.
Here are some tips which enterprises can take heed from:
- First, infrastructure must be integrated. AI workloads depend on coordination. If compute, storage, and networking are not aligned, performance breaks down. This is why organisations are moving toward more integrated models, where infrastructure layers are managed in coordination rather than stitched across separate systems. Reducing interdependencies improves performance, simplifies operations, and removes friction from AI workloads.
- Second, modernisation must be incremental and compatibility-driven. Most enterprises cannot rip and replace their existing systems, risking downtime or losing years of technical investment. The priority is to modernise without disruption, allowing new AI workloads to run alongside existing applications while gradually reducing dependency on rigid systems. Compatibility is the foundation of any successful migration strategy.
- Third, hybrid environments must be intentional. Hybrid cloud is now the norm, unmanaged hybrid is the problem. AI forces organisations to be deliberate about where workloads run, where data resides, and how resources are allocated. Public cloud provides elasticity. Private infrastructure provides control and predictable cost. The challenge is not choosing one over the other, it is ensuring they work together coherently without creating additional complexity.
- Finally, using existing resources more effectively will define how far AI can scale. In many enterprise environments, resources like memory are over-provisioned but underutilised while costs continue to rise. This is driving a shift toward smarter resource management, where infrastructure dynamically prioritises active workloads and optimises how capacity is used. Approaches such as memory tiering are gaining attention as organisations look to expand capability without proportionally increasing cost.
The cloud conversation in APAC is entering a new phase. The question is no longer how much cloud an organisation uses. It is whether its infrastructure is engineered to support AI at scale, and whether it can evolve as those demands continue to grow.
Keith Lee is Cloud Business Director at Sangfor Technologies.
