Southeast Asia Data Centers: The AI Boom, Power Struggles & Future Hotspots

Introduction

In May 2025, I joined the research firm SemiAnalysis on a datacenter information field trip in Malaysia and Singapore. This video is not sponsored by them; I paid my own way. We visited actual data center sites, went through security, talked to operators, and got a first-hand look at the industry. The datacenter boom in Southeast Asia is hot, but there are challenges due to resource shortages and geopolitics.

Beginnings

• According to industry veterans, the Southeast Asian data center boom began in the early 2000s in Singapore.

• In 2000, Singtel and another telecom set up a small data center on the island, working with Singapore's larger real estate businesses.

• They also helped set up the first connection cables in the area.

• Other companies, such as Keppel and Equinix, followed suit and connected to each other for faster data exchange.

• This led to the emergence of Singapore's "Data Center Alley," one of the world's largest concentrated hubs. There are two Data Center Alleys in Singapore, one in the East and the other in the West.

• The data center industry's growth soon outpaced Singapore's ability to provide sufficient water and power. In 2019, the country placed a moratorium on new data center builds, although investments in pre-existing developments can continue.

• Developers then moved to Malaysia or Indonesia, where power and water are more plentiful. However, they often site their campuses close to Singapore to connect to the island nation's fiber cable connections to the rest of the world.

Data Centers

• A data center is a facility that hosts and powers various types of IT equipment 24/7.

• It is built to be a big, standalone computer with facilities and equipment for networking, storage, compute, power supply, and cooling.

• The data center operator builds the facility, secures the water and power supply, patrols the perimeter for intruders, and ensures that the facility operates without incident per the Service Level Agreement.

• The customer brings in the chips and IT equipment that goes into the racks and uses the facility. They often lease the facility for an extended period of time, such as 10 years.

Data Center Variety

• Data centers can come in a variety of shapes and sizes.

• We observed a small, legacy data center located inside the basement of an office building. This data center once served a multinational corporation and is now largely empty.

• There are also medium-sized, multi-floor data centers for multiple tenants, such as government, telecoms, and regional IT companies. These data centers tend to use about a dozen megawatts per floor.

• The truly massive facilities are built and owned by the tech giants in the US and China. Most of these facilities are black boxes, with little information about how they operate or what they use.

• The equipment inside data centers can also vary. Almost all of the data centers we visited in Southeast Asia were cloud servers, which had CPUs, hard disk drives, and networking to run programs on the cloud.

• The COVID-19 pandemic increased the demand for cloud IT services, which kicked off the most recent data center boom. AI then picked up the baton, with chatbots like ChatGPT and Claude being more compute-thirsty than almost any other software service product.

• The data centers that train or run inference for these AI services are often called "AI Factories" by Jensen Huang and others in the industry. They take in power and electrons and generate tokens, with the goal of producing as many tokens as possible efficiently.

Driving in Malaysia

• As soon as you drive over the bridge leaving Singapore to enter Malaysia, you realize you are in a different place.

• We took a scenic route to the first data centers, which gave us a chance to see the villages and the countryside.

• Malaysia has a lot of land, most of which is covered by oil palms. Oil palms are a huge business in Malaysia, as they can be harvested and crushed to create palm oil, which is useful as a cooking oil, food product ingredient, cosmetic ingredient, or even biodiesel.

• However, the oil palm industry has been criticized for incentivizing widespread rainforest destruction. Driving through Malaysia, it is easy to see that whatever was once there has been long obliterated, and all that is left are rows of oil palms that stretch as far as the eye can see.

Sedenak

• After about 45 minutes of driving, we arrived at a massive data center cluster region in an area called Sedenak.

• The experience reminded me of first visiting the semiconductor fabs in Tainan. All of a sudden, the green grass receded and we could see these huge facilities spanning multiple blocks.

• This land hosts several data center campuses, owned and operated by several companies I had never heard of.

• Land in Malaysia is relatively easy to obtain, but the key question is whether it has access to water and power. A major reason why this particular site exists where it does is because the national electricity utility, Tenaga Nasional, was building a substation there.

• This attracted the attention of one of the US tech giants, who agreed to invest some ten years ago. Then, in 2021 and 2022, several smaller data center operators began moving into the area on behalf of their clients. Today, those smaller operators take up the majority of the current region, and the tech giant has hardly a presence there now.

• The land is being developed by a subsidiary of a state-owned company called the Johor Corporation, which also runs a chain of fast food restaurants. This is why I found the lack of convenient restaurants intolerable.

Inside the Data Centers

• The data center campus is entirely fenced off. We arrived at the gate, and a security detail met us outside.

• Before being allowed inside, the security guards took our phones and locked them inside clear plastic bags, like we were taking them on a SCUBA dive. We could still use the phones, but only through the bags, and the plastic strategically blocked the cameras.

• This particular site is one of several in the campus, but it is the largest of the bunch. It is anticipated to be one of the largest in the country, sized at several hundred megawatts in total. It is being built in phases across three massive plots.

• It is interesting to see how the customer has changed their approach over time. The first phase began in 2023, and those buildings were built extremely hastily. A lot of the power and cooling equipment had been prefabricated on site at the factory in China, shipped in standard shipping containers, and installed onsite. This allowed them to get something up in less than a year.

• For later phases, however, the operator and customer decided that this approach was not worth the compromises it forced upon them. So, they decided to go for a more integrated construction approach, despite the longer ramp-up time of 18 months to 2 years.

• Inside the data center, we walked down long, empty hallways, like an industrial warehouse or Costco, to tour the various support rooms. The facility's heart is its Data Hall, which houses the server computers and other IT equipment, usually in horizontal racks. We were generally not allowed into the data halls, as getting inside means possibly getting direct access to customer data, which is a big deal.

• Fences and subdivisions inside the data hall are common. Some customers might be so strict and secretive that they erect privacy screens to keep people from even seeing inside.

• The rooms themselves are clean but spartan, with concrete and high ceilings, and might contain equipment from companies like Schneider Electric for routing power, stepping down high voltage power from the grid, and circuit-breaking.

• One memorable room contained chillers, basically large fridges. The equipment in the data halls runs very hot, and to cool this equipment, almost every data center operator here outside of the hyperscalers blows air through the data halls. The air heats up, and is then collected and cooled with massive chillers. For this one building, the operator said that there are four chiller rooms, three of which are constantly operating, with the fourth held idle as contingency.

• It is important that the data centers can switch to backup sources of power in case something happens with the grid. This includes battery power and backup diesel generators. I recall visiting one warehouse filled with rows upon rows of diesel fuel drums.

• I asked about hiring and training talents, and they told me it was a massive challenge. Junior people are hired right out of Malaysian universities and trained internally, mostly electrical engineering or HVAC majors, or similar. That being said, I did not see many people onsite. During the day, the site might have one or two dozen people on the premises, and during the night shifts, half a dozen, maybe.

Other Data Centers

• After that visit, we visited about five or six more data centers. Once you have seen one, you sorta seen them all, but there are a few interesting differences.

• One that stood out to me was the batteries. Battery rooms are essential, and some of the more recent battery rooms feature lithium-ion batteries made by CATL. However, there are plenty others with simple lead-acid batteries. Operators explained that a recent spate of battery fires in data centers had caused some consternation from the local governments about lithium ion.

• One hazard of Johor is flooding. It apparently floods more than a few times a year during the monsoon season, and more than a few companies in the area emphasize their anti-flooding measures.

• Another thing that caught my attention was the lack of liquid cooling amongst the non-hyperscalers. To me, liquid cooling is a sign that the data center is running power-hungry AI hardware inside its data hall. Operators will never tell you what is running inside their data halls, and I don't think their customers tell them either. But you can make a good guess based on the customer's requests for power, and what cooling system they want to use. If it's a single floor building that uses a lot of power, or uses liquid cooling, it's most likely running AI workloads.

• Operators asked us about it too. More than a few times we were asked the question, "Have you spoken to anyone who is using liquid cooling?" I am somewhat sure that the hyperscalers are running it, but everyone else seems to be using air cooling.

• We did visit a data center that does do it, but let me cover that trip in a future video.

Next Malaysia

• The hyperscalers are still interested in Malaysia, but there are indications that the data center boom there is starting to grow long in the tooth. In 2025, the sentiment within the industry seems to be that getting the water and power, mostly the power, allocated for a brand new, green-field data center project would be very difficult.

• The limitations are not about actual power generation, as Malaysia has plenty of hydrocarbons to burn, but distribution. How can we get the power over there?

• One recent case of a medium-sized, non-AI data center saw the land procured in early February 2024. Securing the water took another month, but securing the power took until November. And then the actual construction of the power connections will take until May 2026. The center is estimated to finish the next month, with the customer getting the keys in October 2026. We were told that this timeline would be even worse if you started in 2025 rather than even just a year prior in 2024.

• So, in light of these restrictions, companies are looking for other areas of expansion within Southeast Asia. But where else in the region can they go?

• There seems to be some growing interest in Australia by hyperscalers. This led to some data center growth in Sydney, which then caused the state government to consider a ban on new data centers. Since then, interest has moved over to the city of Melbourne, and the chatter seems to be that AWS Amazon might want to make that city their major AI hub for the APAC area. Assuming that they don't ban data center builds there either.

• And as for Indonesia, Jakarta announced a great deal of new projects in 2021 and 2022, and those are still on the way. But there has not been as many new projects since then. Though Google is expanding a cloud region there. Demand seems to be an issue. Microsoft recently made business headlines for pausing work on a Jakarta site - part of what has been reported as a larger pulling back.

• Other options include the Philippines and Vietnam. But like with Indonesia, there seems to be insufficient demand in both countries even for cloud services, let alone AI. Someone in Vietnam told me that occupancy rates for national data center projects are under 20%.

• Another impediment in the case of Vietnam is the requirement to work with a local partner like FPT, Vietnam's so-called AI unicorn. At $6 billion, FPT is their most valuable listed tech firm. Working in Vietnam also means getting into bed with the Communist government. FPT's founder is the son-in-law of the legendary General Vo Nguyen Giap, former commander-in-chief of the army and former secretary of the Central Military Commission.

• So, the data center builders have targeted Thailand as the "next Malaysia". Thailand has plentiful water and energy, living costs are not insane, the government is somewhat business-friendly, and demand seems to be good. Probably the thing most in shortage are general contractors to build these projects.

• Google and Amazon are starting off in Thailand with small first phases that they are expected to scale up over the years. Google says they will invest a billion dollars there. ByteDance is also rumored to be hunting for space there.

• There is also a local player - Siam AI, an Nvidia partner. They are associated with the politically powerful Shinawatra clan, and have announced that they would build the country's largest AI data center, 100 megawatts.

The China Thing

• The chip geopolitics, particularly regarding GPUs, are ever-present here. By far, the majority of the hyperscalers building in Malaysia and Thailand are American - Microsoft, Google, and AWS. Oracle is a recent new name on the market.

• But it is also obvious on the ground that the Chinese are building here and very fast. A number of the data center operators have Chinese or Chinese via Singaporean origins. Many of the equipment suppliers and even construction companies are Chinese. And the ultimate end users are Chinese. Alibaba, a bit. Tencent too. But ByteDance is the unquestioned giant. TikTok is legitimately booming in Southeast Asia. They have a very large office in Singapore where they do an extensive amount of AI research.

• The Chinese data center operators have taken steps to appear less "Chinese". This might include a corporate separation and rebranding. For example, the Chinese data center firm GDS had an international arm that they separated out and renamed to DayOne.

• So far as I know, it is legal for the Chinese hyperscalers to rent AI GPU time, so long as the data center operators and actual owners are not Chinese. Don't quote me on that. But it is not clear whether Southeast Asian countries, Malaysia in particular, can secure said GPUs.

• One operator in Malaysia says that he was not sure whether export controls like the recently repealed AI Diffusion Laws were an issue. Or to be more precise, their customers have not said anything to them about it, and they are not sure how it is being dealt with.

• I reckon it is up to the individual trade deal that Malaysia ends up striking with the Americans. And Malaysia is working hard to balance both sides. One such kerfuffle recently happened when a government official said in a speech that Malaysian companies would buy and use Huawei's Ascend AI chips. Except the United States government has since said that it would consider anyone using Huawei chips anywhere in the world to be in violation of export controls. This triggered a very rapid turnaround on the whole matter by the Malaysian government. It is not clear whether or not the plan is still on the way.

• Anyway. As of this writing, the current mood is hopeful. It is hopeful that a deal will be struck and Nvidia AI chips will soon be allowed back into areas seen as more "okay", and data center progression there can continue.

Conclusion

• I want to thank SemiAnalysis, and especially Dan, for this opportunity. It was a truly amazing experience.

• I can get why some might not find data centers all that interesting. The companies running them are super discreet. Many people in the industry readily acknowledge that the whole thing investment-wise is a real estate and infrastructure play.

• But these are where the chips live. And it is where the current bottlenecks are as well. The whole AI supply chain is like an engine. It needs the exact right amounts of air, fuel and spark to go. If these ingredients are not present in the right form or amount, then the engine will sputter and roll to a stop.

• A year or so ago, the bottleneck was TSMC's advanced packaging capacity. But TSMC has ramped up to such an extent that that is no longer the case. I even caught some random chatter of CoWoS order reductions.

• Right now, the constraints are in power. Full stop. These AI factories require unprecedented amounts of power and it is taking literally years to connect them to the grid. Clearing out these bottlenecks, however, will not come quickly.

• Malaysia as a data center hub seems to be on the wane. Thailand is ascending. But with politics the way they are, things can shift in an instant.