Country Spotlight: Malaysia

MAIDC 2026: the conversation in Kuala Lumpur

That divergence was the central focus of the first Malaysia AI Data Centre Technology Awareness Conference (MAIDC 2026), held on 22 May at the Kuala Lumpur Golf & Country Club. Hosted by the Uptime Institute and the Malaysia Digital Economy Corporation (MDEC) and sponsored by STULZ, among other players in the industry, MAIDC 2026 brought operators, vendors, and policy bodies together for what was effectively a tour through every layer of an AI-ready data center. The framing question of the day was direct — what does it actually take to build a future-proof, AI-ready Malaysia? Moderated by Patrick Chan, Managing Director of the Uptime Institute, the opening panel pushed back on the easy answers of “more data centers” or “more AI startups” and argued that AI readiness is a stack problem rather than a real-estate one. Power, chips, storage, cooling, and day-to-day operations all have to work together while educating and retaining the right talent. The rest of the day’s speakers populated those layers one by one. The simplest way to follow the conversation is to walk the stack from the inside out.

Source: MAIDC'26, Panel Discussion, "How to build future-proof AI-ready Malaysia"
 

Silicon

It starts with the silicon. Malaysia's position in the global semiconductor supply chain is stronger than it might first appear. Built over decades of manufacturing and now accelerating into AI hardware. Supermicro, one of the world's leading AI infrastructure manufacturers, lists Malaysia as part of that supply chain among its four global production locations alongside the US, Europe, and Taiwan. NVIDIA's own hardware roadmap tells the same story from the demand side: each successive generation from the H100 to the GB200 NVL72 and the upcoming Vera Rubin brings higher rack densities, greater thermal loads, and more complex infrastructure requirements, raising the bar for every country that wants to host them. Malaysia's move from server hosting toward a broader semiconductor ecosystem is already underway, and its 2035 ambitions as a regional AI infrastructure hub are grounded in the broader technical base and production it is building today.

Storage

Silicon is necessary, but not sufficient alone. The next question tackled was the speed at which data reaches the chips efficiently. As the Senior Sales Engineer at WEKA noted, typical GPU utilisation sits around 30% in AI deployments, with the remaining 70% spent waiting for data rather than computing. Storage accounts for roughly 5% of infrastructure cost, but if neglected, it easily becomes 100% of the problem as operating expenditure follows capital expenditure. An idle GPU still draws power, consumes rack space, and needs cooling. The time cost, however, is the hardest to recover as slow storage directly translates into longer training runs.  Malaysia already has a working reference: YTL AI Cloud’s GB200 NVL72 deployment runs at around 1 GB/s sustained reads per GPU and sub-half-millisecond latency, meeting NVIDIA’s NCP reference architecture for GB200.

Cooling

To operate the GB200, traditional data hall layout and cooling design are no longer sufficient. Every new GPU generation pushes rack densities further: the H100 draws 70 kW per rack, the GB200 NVL72 around 130 kW, and Vera Rubin further still — up to 220–227 kW per rack. At densities around 20 kW, liquid cooling becomes the practical choice; above 50 kW, it is non-negotiable, as the laws of thermodynamics do not allow air alone to extract the heat generated. Regions like Johor, which are currently seeing heavy hyperscale expansion and were built for 10–20 kW racks, may become obsolete faster than expected as AI workloads shift toward 50–150+ kW deployments. To ensure a facility is future-proof and capable of accommodating AI and ML workloads, operators must plan from day one for liquid cooling infrastructure readiness, alongside higher power redundancy models, water availability, grid scalability, and modular infrastructure. Following this shift, cooling sits at the intersection of efficiency, density, and water. This places Water Usage Efficiency (WUE) on equal footing with Power Usage Effectiveness (PUE) within Malaysia’s National Sustainable Data Centre Framework. Malaysia’s tropical climate further compounds the challenge, adding specific risks like heat, humidity, flooding, drought variability, and grid stress, which make water and cooling resilience inseparable from broader climate resilience for AI infrastructure.

Reliability: Commissioning and Monitoring

AI deployments are creating a new specialised class of high-density data halls, distinct from traditional enterprise IT in cooling, power delivery, and even building layout. These distinctions have direct practical implications for the commissioning practices applied. AI facilities must be tested against extreme power and cooling transients, with prolonged capacity burn-in testing recommended during CxL4 to verify capacity and system functionality a step beyond what traditional commissioning required. Primary and secondary liquid cooling loops add an additional layer of testing: pre-flushing of cooling load banks before commissioning, re-flushing and fluid replacement after testing, and general water-quality cross-verification to ensure system health. The risk profile also changes, since higher-density racks are more susceptible to overheating when cooling is disrupted, demanding more vigilant on-site monitoring. Once the facility is live, the operational load scales non-linearly with AI capacity. A 1 GW AI campus generates more than two million monitored data points. This number is well past the 250,000-point ceiling a conventional BMS typically handles before it must fragment into parallel systems. Liquid cooling sharpens the requirement further, calling for one-second telemetry rather than the slower polling intervals that air-cooled halls tolerated. Finally, more liquid flowing in the facility directly translates into increasing the number of liquid contact points requiring close supervision to prevent any major system event.

Connectivity

Sitting on top of the infrastructure, the connective layer is a frequently overlooked element — yet it is the invisible fabric that makes silicon chips and hardware function as one system. That includes the high-speed network that lets GPUs communicate with each other, the controls that keep different users’ workloads securely separated, the connections between compute and storage, the billing systems, and the security wrapping all of it. Without that layer running smoothly, the most advanced hardware in the building sits idle. Pulled together into one platform, the same racks become a cloud that engineers can access at much higher speed.

Looking ahead

As the conference made clear, the global bottleneck for AI growth has shifted from algorithms to physical infrastructure, and MAIDC 2026 highlighted precisely what it takes for each layer to operate in the new high-density realities. Power, cooling, chips, and connectivity — each is a multi-year build, and each depends on the choices the country is making now. Malaysia built its initial momentum on lower land costs, competitive power pricing, and strategic location, which attracted international hyperscale players to expand in the country. Future-proofing, however, means moving past server hosting toward AI cloud clusters, dedicated GPU capacity, liquid cooling readiness, and a broader semiconductor ecosystem. The 2035 vision the panel sketched is Malaysia as an ASEAN AI infrastructure hub and renewable-powered digital economy on the condition that infrastructure growth is matched by talent, energy, and policy execution. The next decade will show whether Malaysia makes the transition from hosting foreign AI infrastructure to becoming a regional AI services leader, and STULZ is proud to support the country’s ambitions with reliable and efficient cooling solutions.