Malaysia has done something few markets in the region have managed. It has moved electricity to data centres at remarkable speed.
A July 2026 parliamentary answer from the Malaysian ministry Petra records that Tenaga Nasional Berhad (TNB) supplied 36 operating data centres with around 4.5GW of planned supply capacity through the first quarter of 2026 and met demand for 23 more projects under construction carrying 3.8GW. That builds on TNB’s Green Lane Pathway, which has sought to compress grid connection timelines from the usual 36 to 48 months to a mere 12 months, and an RM43 billion (US$10.75 billion) capital allocation for grid investment across 2025 to 2027.
That is a tremendous breakthrough. But it also raises the question that Malaysia needs to ask next. For the first phase of the data centre race, the challenge was getting enough power. As AI workloads become denser and more unpredictable, the next challenge is what happens to that power once it arrives.
Data centres are projected to consume up to about a third of Malaysia’s total electricity demand by 2035, up from 7 percent this year. At that scale, how efficiently electricity is converted into useful computing power is no longer simply a facility-level engineering question. It becomes a question about national infrastructure. Because a megawatt delivered to a data centre is not the same as a megawatt delivered to a GPU.
Where the power actually goes
Between the substation and the chip sits a chain of infrastructure: transformers, switchgear, UPS systems, power distribution and cooling. Every stage determines how much of the electricity entering the facility ultimately becomes useful compute, and how much is lost as heat or consumed by the infrastructure supporting it. That distinction matters because AI is changing the density of the load itself.
- Jason Yuan is President of Southeast Asia and ANZ at Delta Electronics.
At single-facility scale, a difference of 0.1 or 0.2 in Power Usage Effectiveness (PUE) is an engineering detail; across the load Malaysia is now planning for, it is hundreds of megawatts built, paid for, and spent on heat.
The same applies to stranded capacity. An approved megawatt that a building cannot cool does not become useful AI capacity. The grid connection is consumed either way, but the computing capacity is not guaranteed. Efficiency and resilience both share a root cause. Small amounts of airflow inefficiency add millions in annual energy costs and contribute materially to outages. A facility that wastes power is also one that is less resilient.
Malaysia is calling for efficiency through policy. The Ministry of Investment, Trade and Industry (MITI)’s Guidelines for Sustainable Development of Data Center set a design PUE limit of 1.4 or lower for hyperscale facilities above 21MW, alongside a water usage effectiveness ceiling of 2.2m/MWh.
The standard exists, so the next question is whether the infrastructure being deployed today can continue to meet it as AI density rises. That question matters even more alongside a national target of 70 percent clean energy by 2050. Efficiency is the most immediate form of energy supply because a watt that is not wasted never needs to be generated, transmitted or replaced with additional clean capacity.
AI is changing the infrastructure equation
AFCOM’s 2026 State of the Data Centre report found average rack density jumped from 16kW to 27kW in a single year; the largest year-on-year increase in the report’s decade-long history. In the same survey, 39 percent of operators said their cooling does not meet all operational requirements, while 36 percent have deployed liquid cooling and another 28 percent plan to within 12 to 24 months.
At the frontier, the divergence is sharper. NVIDIA’s Vera Rubin platform is pushing AI infrastructure into a new class of rack-scale computing, with the industry already planning for power architectures capable of supporting 1MW IT racks and beyond. As rack power requirements move beyond 500kW, however, the challenge is no longer simply delivering more power. The electrical architecture itself needs to evolve, with 800VDC emerging as an important approach for supporting these ultra-high-density AI systems.
When rack densities move this quickly, power and cooling can no longer be treated as separate procurement decisions. The power architecture determines how electricity is delivered to increasingly dense loads. The cooling architecture determines how effectively that resulting heat can be removed. And together, they determine how much computing capacity can ultimately be delivered from the grid connection Malaysia has worked so quickly to secure.
In other words, speed to power only creates an advantage if that power can be converted into usable, resilient compute.
Why the investment numbers still matter
None of this diminishes the investment story. It is precisely why the next phase matters.
In July, Malaysia’s Digital Minister Gobind Singh Deo was asked whether operators were choosing Malaysia mainly for cheap utilities. He answered by citing Bank Negara’s report, sharing that senior roles pay RM10,000 to RM30,000 (US$2,500 to US$7,500) a month against a 2024 median wage of RM2,793 (about US$700), with the return coming through high-skilled jobs, technology transfer and AI ecosystem development. The value of the data center sector therefore extends beyond the buildings themselves.
Wood Mackenzie estimates Johor alone has drawn US$42 billion in cumulative investment from hyperscalers and technology firms. Capital at that scale funds the grid that makes the next tranche viable, signals credibility to the next investor, and, critically, increases the importance of how efficiently that infrastructure is used.
The next bottleneck sits behind the grid
Malaysia’s next infrastructure challenge sits behind the grid. As AI rack densities move rapidly from today’s industry averages toward the 100kW-plus range and ultimately toward megawatt-scale systems, power and cooling can no longer be planned as separate packages or added later. They need to be designed together from the outset, particularly as critical electrical equipment such as high-capacity transformers faces multi-year lead times.
This also creates an opportunity for Malaysia to move further up the value chain. The National Semiconductor Strategy’s ambition to move from “Made in Malaysia” toward “Made by Malaysia”, backed by RM25 billion (US$6.25 billion) in fiscal support and a 60,000-engineer training target, also provides a foundation for developing local capabilities in the power and cooling infrastructure that AI data centres increasingly require.
The next measure of readiness for Malaysia
Malaysia has already demonstrated that it can move quickly when it comes to getting power to data centres. The next test is whether it can turn that power into useful AI capacity just as efficiently.
That means looking beyond megawatts approved and connections delivered. It means asking how much compute can be supported by each megawatt, how efficiently that power can be converted and distributed, how effectively the resulting heat can be removed, and how resilient the infrastructure remains as rack densities move toward the next generation of AI systems.
For Malaysia, the next phase of the data centre race is therefore not simply about securing more power. It is about making every watt count.
Jason Yuan is President of Southeast Asia and ANZ at Delta Electronics.
