Japan, South Korea primed for data center growth in age of AI

There is strong and increasing demand for data centers in South Korea and Japan, and advances in AI will mean that Seoul, Osaka and Tokyo won’t be the only locations for largescale data center development going forward. Those are some of the main takeaways from a recent panel discussion hosted by Mingtiandi.

Japan topped the Asia Pacific region in 2022 by raising $4.2 billion for new data center projects, while South Korea tied with India for the second spot with $1.9 billion in commitments for new developments last year.

Patrick Boocock, CEO of private equity alternative assets for CapitaLand Investment’s real assets team, said “With increased 5G cloud computing, digitalization, AI, all of the tailwinds that are supporting the data center industry, coupled with … supply challenges in terms of finding adequate, suitable land. … There is generally pent-up demand in all major markets, in particular in established markets like Korea and Japan.”

Diarmid Massey, CEO of data centers at real asset manager ESR, said, “There’s no doubt that the demand and the potential growth for AI-driven applications is seeing an increased demand.” Massey attributed much of the demand to hyperscale and AI providers who are seeking to buy effective data center space wherever they can get it.

“We [ESR] are investing in building new state-of-the-art data centers in Japan and Korea because we believe in the demand curve,” he said. “We believe that the AI boom is here to stay, as well as the ongoing evolution to the cloud and digital transformation across industries in Asia Pac.”

Not only is AI increasing data center demand, but it is also altering the landscape of where those data centers can be built. The training for machine learning can be done far away from any urban center, as it “does not require proximity to end users or interconnection to other facilities,” wrote CBRE in a recent article on how AI will impact the data center market. Discussing the ability for machine-learning AI systems to train in more remote data centers, Boocock observed that U.S. hyperscalers “are developing very large data centers in the Nordics for exactly that reason.”

CapitaLand’s main targets continue to be Tokyo and Osaka, Boocock said, but “as we start to roll out AI, I think the proximity to the major availability zones becomes less of an issue.” The panel discussed Hokkaido, Japan, as one potential intriguing location for data center development. AI inference, when AI analyzes new information from outside its training knowledge base (for example, someone “conversing” with ChatGPT), requires high computing performance and low latency, which means a focus on edge data centers near large markets.

Jing Zhou, senior director of alternatives and strategic transactions at Nuveen, noted that we are seeing mostly business-to-business AI computing requirements. When the AI market progressively gets more focused on business-to-consumer and consumer-to-consumer transactions, Zhou said, “that will be the real edge data center decentralizing distributions across all markets,” with “much smaller distribution centers which have to be location sensitive and latency sensitive.”

This doesn’t mean the data centers providing computing for AI training will become obsolete, Zhou observed. “It’s the entire ecosystem. It’s not one replaces another,” he said. “It’s really they coexist with the four different users.”

Data centers from 20 years ago, however, which supply 2 kilowatts per rack, aren’t optimized for AI computing that often requires 50 kilowatts per rack. Some of the older centers “can be retrofitted with additional power and additional cooling,” said Massey, “but in many cases they’re going to continue to be a lower density sort of asset. Now, they may still continue to be productive, and they may just be used for networking. They may be used for lower-density applications.” But it is important for investors in today’s market to have flexibility in mind when developing data centers, he said, so that they can scale up to take on AI demands.

“It’s up to us as an industry to design that flexibility into the business,”  Massey said. “But equally as much, to be able to get power we need to have good relationships with the grid and with the power companies and, more importantly, renewable energy. So we won’t necessarily be able to get that power on site, but we need to have access to renewable energy to be able to provide the good renewable feeds to the AI demand.”

As for the actual fundraising and development of data center projects, the panelists seemed to agree that there is no one right way. Massey mentioned how ESR has been partnering in joint ventures with data center operators (for example, they have done two deals with Stack in South Korea and in Osaka), has more data center JVs in the pipeline, and that ESR would like to continue to do more JV data center projects in the future. “The aim is that effectively we leverage the ESR skill sets,” he said, “which is sourcing land, getting power, getting permitting in a country and getting financing obviously, and then tie up with a regional or global operator that wants to expand their footprint but doesn’t necessarily want to manage the construction process.”

Construction costs are stabilizing but remain high, labor costs are up, and MEP (Mechanical, Electrical and Plumbing) is more expensive now than it was two years ago, said Massey. “We’re starting to see that flowing into the construction of the sites now,” he added. “It’s prevalent in Japan and Korea, it’s prevalent across the world, and that will flow ultimately into leasing rates as new supply starts to hit the market built with a higher-cost model.”

Boocock agreed with Massey on the viability of working with data center operators and noted that “roughly two thirds of the cost of a data center development is in the MEP.” Focusing on staging out the rollout of the MEP in the racks can minimize the risk for investors as data center owners lease up over time, he said.

Bob Tan, executive director of capital markets for JLL’s data center practice group, said that because of the strong digital ecosystems of both South Korea and Japan, the countries are prime markets for AI to grow, and thus for data center demand to grow. Zhou affirmed the countries as prime markets, attributing it as well to their varied customer profile for data center leasing and because they both have well-established cloud providers “from both the West and also from China.”

In light of the high costs in general for building data centers, Boocock gave out two pieces of advice. “Number one, you need to be extremely diligent on your submarkets,” he said. “You need to be very confident that where you are planning to develop, number one, you can get the energy to the site. And number two, that you’ve got the right supply-demand analysis done, that you’re confident you’ll be able to lease it.”