Speaking on RCR AITech Talk, Obinna Isiadinso, World Bank Group’s International Finance Corp. global sector lead for data center investments, says he expects 3 to 4 years of constraint or lag time for energized power, which means it’ll be 2029 or so before the U.S. sees 5-15 GW of new capacity. According to Isiadinso, this makes carrier-neutral, multi-tenant data centers projects in emerging economies an attractive option for data center investors and lenders like IFC, which increasingly looks to wholesale, retail, and hyperscale data center projects around the world (as with Scala in Brazil, Yondr in Malaysia, Kio in Mexico, Liquid in Africa). In the below video and interview summary, Isiadinso explores:
- Data center supply and demand in the U.S. and emerging markets;
- Future-proofing and industrialization of AI data centers;
- High enterprise value and EBITDA;
- Shift from equity to debt financing.
Uneven Supply and Demand
The majority of hyperscalers’ capital and resources reside in the U.S. today, where “energized power” is the single largest physical constraint for AI infrastructure expansion. While chip supply was previously the primary bottleneck, the focus has shifted to the inability of electrical grids to deliver reliable power at the scale required for massive AI training and inference. As that happens, Isiadinso predicts an expansion into emerging markets like Brazil, India, Malaysia, Mexico, Kenya, Morocco, Egypt, and Nigeria.
“Today, the U.S. market has about 30-50 GW of total capacity, but hyperscalers – the biggest customers and owners for data center capacity – want to add 30-50 GW over the next 5 to 7 years – a very compressed timeframe.” Isiadinso says that considering it’s taken the U.S. 30 years to get to that first 30-50 GW, the race to replicate that in a matter of a few years is going to be challenging. “We watch what the hyperscalers are doing, and where they are expanding, as they are the ones moving the market.”
He notes that just a few years ago, the biggest project was 50 MW, but larger projects of 100 MW to 1 GW and above will take about 5 to 7 years because of the need to get power to data center sites. “Depending on the state, there are regional utilities that have the infrastructure in place for projects of 20-25 MW and below. Anything larger is going to require more infrastructure, such as the interconnections, transformers, substations, and everything else that has to connect to the data center site.” Isiadinso reminds us that there’s “ already a long queue of data center investors and developers with applications already in for new projects – projects that are already online.”
This means it will take years to work through that pipeline of power infrastructure and deployment for new data centers. “Our sense is that it will take 3 to 4 years of constraint or lag time, which means it’ll be 2029 or so before we see 5-15 GW of new capacity in the U.S., and maybe 10 to 15 years to build out the type of hyperscaler capacity that will make us competitive on a global scale.”
The lure of emerging markets
Like the state-by-state dynamics in the U.S. that shape the energized-power landscape, there are unique dynamics within each developing region. For the Middle East, Asia, and Latin America, there is no blanket prediction, but, in Latin America and South America, Isiadinso notes Brazil does stands out, “with 30 GW of excess renewable energy, Brazil has a surplus position and the ability to deploy more data center infrastructure.” He notes, however, that transmission is a challenge. “They have renewable energy in the North, but the grid quality to the South is not great right now. They have to build out the transmission infrastructure to keep up with the significant renewable energy they produce.”
Africa, on the other hand, is in an energy-deficit position, with generation as an issue. “They are not producing enough energy, so there is an opportunity for investors like us to go there and generate more power, especially renewables. And, we could help improve the quality of the grid. Kenya, Egypt, and Morocco have access to renewables and natural gas, so they are in a better position than others in that region.”
Isiadinso also believes Malaysia, Thailand and other parts of Southeast Asia as prime for opportunity, with projections that data center capacity will possibly triple by 2030 thanks to the region’s strategic positioning to Australia, China, India, and Japan, as well as lower construction costs than the average in developed markets. According to Isiadinso, there is a strong push for solar, wind and hyrdropower that will make this region attractive for renewable energy in the upcoming years. “They use a lot of coal but are rapidly switching to renewable energy sources, especially in Malaysia and Thailand,” he points out.
Across all of these regions, 100 MW and above is not yet the norm: “Right now, only in certain countries will you will see 100 MW and above, as so few have access to power, and renewables in particular. For 20 MW and below, there is less constraint,” he says.
Preventing obsolescence
In emerging markets, there is the benefit of leapfrogging air cooling for advanced liquid cooling to meet the thermal demands of AI. “You’ll see heavier load-bearing floors, higher ceilings, wider hallways to future-proof what will be billion-dollar investments,” says Isiadinso, who adds that “GPU and cooling technology is changing every 18 months or so.”
Because server and rack technology changes “every 5 to 7 years,” it’s important that data centers be built in such a way that new technology can be moved in and out rapidly. “They are designing data centers that make it easy to replace the IT infrastructure within…load bearing floors to handle bigger, larger racks; modularity so that each module handles the rack requirements, the cooling; and the larger hallways to allow room for replacing equipment.” Isiadinso estimates that a data center shell built in this manner can last 30 years or so. “The data hall is where the evolution takes place, changing every 2 years. Every 5 years, you change out the equipment.”
Growing valuations, shift from equity to debt financing
Investing in AI data centers, versus traditional Cloud data centers, requires more patience and longer-term outlooks, according to Isiadinso. “We are still early days in terms of valuations for these investments. For us, most of our investments are in cloud, hyperscale, colocation types of data center projects.” He says investors are looking at traditional data centers with a long-term average 17x EV/EBITDA multiple (Enterprise Value / EBITDA). “When looking at a 10-year average for bigger data center players like Equinix or Digital Realty, there’s typically a 17x EBITDA, but right now, the 10-year average for a U.S. data center players is 25x EBITDA. When looking at M&A transactions though, they go north of 25x maybe even up to 30x.” He explains that when you factor in future growth, the multiples do come down, which is a big driver of high valuations, “because so much growth is expected.”
Because the business models are attractive, with long-term cash flows and contracts with AAA rated clients, recurring revenues, investors feel confident and that feeds valuations. “It’s something that can be highly profitable, once you scale up and stabilize the assets,” added Isiadinso.
He notes that AI data centers like CoreWeave, IREN, and other companies with high valuations are seeing EBITDA of 20-25x, “but multiples have been coming down over the past couple of months. Fundamentals are very strong and demand for the infrastructure is strong.” He notes that because hyperscalers “can’t do it all on their own,” there is a need for third-party firms to help them develop the infrastructure. That will keep multiples elevated for some time.
As CoreWeave, Nvidia, OpenAI, and others increase their investment in “ecosystems that support their own ecosystems,” Isiadinso says there will be a continued shift from equity toward debt financing. “When you look at the numbers with 30 to 50 GW and something like $3 trillion to $5 trillion in capex, about 50% to 60% of that will be equity, and 40% to 50% will be debt financing. No one player can do that all on their own balance sheet, so they have to raise that capital from other investors.”
Will there be a ‘killer-app’?
“There’s a big difference now than what we saw in the early 2000s when fiber companies were investing in some of their customers. This is a bit different,” said Isiadinso. “For example, Nvidia and the demand for that infrastructure isn’t going away any time soon, so I think year after year there will be a progression, with no ‘silver bullet’ year,” every year –much like the evolution of the Internet.” He notes that similar to the 30-year evolution, and the amount of capital that went into its global buildout, the AI buildout will be similar in that continuous march, as opposed to it suddenly materializing in a certain number of years. “Again, it’s the longer-term view into which we will be investing. In AI infrastructure, for 10 to 20 years in the U.S. and globally, we expect every year to see great applications emerge.” He points to OpenAI, which was the first with 800 million weekly average users and 40 million paying subscribers. “If there’s a 5% conversion now that grows to 10%, that turns into 100 million users, and they will find ways to monetize the infrastructure.”
