The wave of new data centers under construction in the Malaysian state of Johor is beginning to take a toll on the local power grid. The region, historically Singapore's industrial hinterland, has turned into one of Southeast Asia's most attractive hubs for digital infrastructure. But the speed at which projects are materializing is generating a silent short-circuit: energy demand, also driven by artificial intelligence workloads, is growing faster than generation and distribution capacity.

The data center race in Southeast Asia

Over the past three years, Johor has seen a multiplication of investment announcements. Major cloud operators and infrastructure funds have identified the area as a natural extension of Singapore's financial hub, where limited physical space and moratoriums on new construction have pushed expansion plans outward. The result is a data center pipeline that, according to industry analysts, could double installed capacity in a short time.

Proximity to submarine cables, relatively low land costs, and a favorable regulatory framework have done the rest. However, electricity availability is becoming the limiting factor. Local utilities are struggling to keep pace and have already signaled the need for substantial grid upgrades to avoid bottlenecks that would ripple across the entire region.

Hardware and consumption: the energy puzzle for on-premise AI

The sharpest acceleration comes from AI workloads. Training and Inference on LLMs require cutting-edge GPUs, often in clusters of tens or hundreds of units, with consumption easily exceeding one kilowatt per single node. In an on-premise context, where the organization retains direct control over hardware, the energy bill significantly impacts the TCO. Added to this are cooling costs: the shift to liquid cooling, now almost mandatory for the latest GPUs, introduces further infrastructural complexity.

Those designing a local infrastructure for LLMs must therefore include energy availability and cost among the top variables, alongside quantization choices, the amount of VRAM needed, and memory bandwidth. Johor's experience shows that even with political will and private investment, the physical electricity grid can become an obstacle that's hard to bypass.

What changes for those evaluating local infrastructure

The strain on Johor's grid is not an isolated case but a signal of a tension emerging in many digital hubs. For companies considering on-premise deployments – driven by data sovereignty needs, GDPR compliance, or model control – this dynamic adds a layer of uncertainty. It's no longer enough to compare the cost of servers or software licenses: an integrated assessment is required, covering the stability of electricity supply over the medium term and the possibility of accessing renewable sources through direct contracts (PPAs).

In some scenarios, a hybrid approach or colocation in edge data centers – which offer low-latency connectivity to offices but share energy infrastructure with other tenants – can represent a compromise. The balance among CapEx, OpEx, and operational risk becomes the decisive factor.

Beyond the cloud: sovereignty and real costs

The Malaysian case is instructive for anyone designing an AI strategy away from the 'all-in-cloud' approach. Deployment choices are never just technological: they touch on energy geography, industrial policies, and grid resilience. In Europe, the digital sovereignty debate has often focused on data localization, but the availability of stable, predictably priced electric power is equally strategic.

For those evaluating a self-hosted environment, the analysis published on AI-RADAR's /llm-onpremise provides tools to weigh these trade-offs: from the impact of quantization on required VRAM to cost models that integrate energy as a primary factor. The goal is not to suggest a one-size-fits-all solution but to offer a map for navigating a landscape where electricity has become the true raw material of artificial intelligence.