Across Asia-Pacific and Japan (APJ), AI has moved quickly from innovation labs into boardrooms, national strategies, and enterprise roadmaps. Leaders are talking about sovereign AI, edge inference, and AI factories. They are also assuming, often a little too casually, that their existing infrastructure can absorb the next wave of cloud-native workloads. That assumption deserves a closer look.
Most enterprises are not building AI on a blank slate. They are building it on decades of software, long-running applications running mostly in virtual machines, fixed networks, mainframes, and operating models that were never designed for elastic, container-native AI.
This does not mean AI is failing because the models are weak. The issue is more practical. AI pilots can work beautifully in a sandbox, especially when they run in a clean Kubernetes environment with limited dependencies. Yet, production AI is a different story. It has to connect to enterprise data, security, networking, governance, and applications that may have been running reliably for 10, 20, or 30 years.
That is where the real work begins.
When AI leaves the sandbox, legacy systems push back
Containers are attractive for a simple reason: They help teams ship software faster. Developers can build in smaller increments, deploy more often, and move with less friction. That matters enormously as enterprises build new AI-enabled applications.
But virtual machines remain for a very practical reason: they work, they are everywhere, and many still run applications the business cannot simply switch off.
Nutanix’s latest Enterprise Cloud Index report shows this hybrid reality already taking shape: 71% of global organisations run AI-enabled apps across both virtual machines and containers, while only 14% run them directly on bare metal servers. At the same time, 83% are building new applications in containers, showing where development is heading without suggesting the existing estate disappears overnight.
The future is not containers instead of virtual machines. It is containers and virtual machines working together.
Kubernetes can unify AI operations, or deepen fragmentation
Kubernetes has become the default deployment model for new applications, including AI applications. That is not surprising. It brings portability, consistency, and a strong open-source ecosystem. The trouble starts when Kubernetes is treated as an infrastructure island. If you build a separate stack with dedicated hardware and a siloed team, you haven’t modernised your infrastructure, you’ve just rebuilt your organisational chart inside the data centre.
This is not just a minor nuisance. The same ECI study found that 82% of global organisations believe silos between business units and IT make technology initiatives harder to execute. The same problem often appears in infrastructure: Every team makes a rational decision in isolation, but the enterprise ends up with systems that struggle to work together.
When a containerised AI application needs to talk to a virtual machine-based application, those divides become very real. Different networking models, identity systems, and security policies can make that interaction feel almost external. In APJ, this becomes even more complicated with enterprises operating across multiple countries, each with different data localisation rules, compliance expectations, and operating realities.
Kubernetes should not become another high-cost silo. It needs to be part of a connective layer between modern AI workloads and the enterprise systems that still run the business.
Build for the hybrid reality
The practical path forward starts with accepting the hybrid reality. AI workloads will span cloud, on-premises, and edge environments. That means infrastructure decisions should focus less on purity and more on cohesion.
Enterprises need a common foundation where Kubernetes and virtual machines can run side by side, share consistent networking and security models, and avoid forcing teams into separate operational worlds. They also need Kubernetes environments that remain upstream and conformant, preserving portability and reducing the risk of future lock-in.
Just as importantly, Kubernetes must become an operational non-event. Hyperscalers succeeded because they abstracted complexity: click a button, get a conformant cluster. Enterprises need that same experience on-premises and at the edge, especially when data, latency, cost, or regulation make public cloud the wrong place for a workload.
The goal is not to move everything everywhere. It is to give leaders the flexibility to run the right workload in the right place, without being constrained by whichever tool happens to be easiest to access.
APJ’s AI race will be won through cohesive architecture
AI is accelerating faster than infrastructure modernisation across much of APJ. That gap is becoming a leadership issue, not just an IT issue.
Legacy gravity will not disappear. Nor should every existing system be treated as technical debt. Many applications are still reliable, valuable, and deeply embedded in how organisations operate. The task is to connect them intelligently to the new world of AI, not pretend they no longer matter.
The organisations that succeed in APJ’s AI race will not simply be those with the biggest models or the flashiest pilots. They will be the ones that build architectures capable of bringing old and new together without multiplying complexity, modernising where it matters, without breaking what already works.
