Data centers are needed to anchor a burgeoning digital revolution. In January, Mizuho hosted the Mizuho Insights Symposium: A Holistic View of Data Centers in our New York office. The event provided our clients and investors with a deep dive and comprehensive view of the data center landscape – from the energy that powers them, the artificial intelligence (AI) models driving the hyperscaler spend and an analysis of the global expansion the industry is now seeing.
Below are key takeaways:
- Data Growth Driving Current Supply / Demand Conundrum - The amount of data consumed is growing exponentially – on an hourly basis. To train large language models and create functioning generative AI, these centers, along with uninterruptible power generation, are vital. As recently as 2018 there was supply and demand equilibrium between the number of data centers and users, Tony Wanger, CEO & Founder of Regnaw Capital told event attendees.
- Building Data Centers Requires a Lift from Multiple Industries - We’ve since turned a corner – with demand for these data centers significantly exceeding supply and very little (if any) vacancy in the core data center markets. Construction lag time for both data centers and power projects to support them are into multi-year periods. “I think the data center industry has been on a journey for a long time to really capitalize this opportunity,” Wanger said, adding that a mix of industries including real estate, energy, telecom and technology are all needed to solve the data center capacity problem. For example, semiconductor firm Nvidia needs support from an “ecosystem” of several different industries.“They’re not an air conditioning company. They’re not a liquid cooling company. They are not a cabinets company. They’re not a real estate company. But they need this ecosystem to sell more servers and chips,” Wanger said.
- U.S. Has Led the Way – But Other Geographies Leaning into Development - While the U.S. has been the leader in terms of data center development and exposure (accounting for 75% of all data centers), other geographies are showing increased growth. Europe, LatAm and Asia Pacific remain areas of focus for significant amount of infrastructure capital. This capital is leaning toward both greenfield build opportunities as well as targeted M&A.
A more detailed discussion from the conference panels is highlighted below.
Models, Inference, AI & M&A
Machine learning and artificial intelligence have been around a long time but with advancements in applications such as ChatGPT, inference functions are expected to drive the bulk of the tech space advancements. Microsoft’s Copilot is another example. “Inferencing” refers to “the process of running live data through a trained AI model to make a prediction or solve a task,” according to IBM’s own definition. Model sizes continue to grow to train generative AI and, with that, an increase in spending, power demand and sophisticated chips.
One panelist likened inference modeling to cars on a roadway.
“It’s like building out the equivalent of the United States roadway and bridge and tunnel infrastructure,” said Jordan Klein, Managing Director, U.S. Equity Sales Trading, Mizuho Americas. “Then the cars on the road which take you to the places you want to go, that’s inferencing. Inferencing is really where the rubber meets the road. That’s where you monetize, that’s where you customize.”
For example, ChatGPT was a game changer in that it infers answers to questions from data that has been consumed and catalogued. To be sure, it takes enormous capex to run and train models (as well as energy consumption).
Jennifer Fritzsche, Managing Director, Technology, Media & Telecom, Mizuho | Greenhill who moderated the panel posed the question: How long are investors willing to wait before they see a return on investment?
Microsoft, for one, is building out inferencing models and investing in AI “because the demand is there,” Klein said. “I don’t think any investor believes that companies that went through that transition and built out all that infrastructure would make a mistake and overbuild for something that’s not going to materialize.”
Once the winners and losers are parsed out, which could take another few years, consolidation is likely.
“We’ve already seen a whole bunch of foundation models consolidate with Anthropic and Open AI and Microsoft making acquisitions, so we will continue to see that happen,” said Aravind Natarajan, Managing Director, Technology, Media & Telecom, Mizuho | Greenhill during the panel discussion.
An “All Hands-on Deck” Approach for Power Resources
Inference blueprints need to be solidified into working models, and, so far, require huge amounts of electric power demand in order to reach that goal. Recent reports suggest that Chinese-owned DeepSeek has created a model that uses less energy and is less expensive to run, but that still needs bearing out.
Most estimates from banks, consultants and government agencies call for sharp increases in electric power consumption for AI-focused data centers – as much as a 30% increase by 2030, according to McKinsey. This could challenge ESG goals since uninterrupted power supply is what a data center needs to effectively run, and renewables can’t guarantee that.
More traditional fuels that can provide reliable power are likely to prevail.
“There’s been a 180 in natural gas over the last 12-14 months,” said Paul McNutt, Head of Power, Energy & Infrastructure, Mizuho | Greenhill. “And it’s not driven by the gas industry, it’s not driven by LNG, it’s not even driven by the utility sector. It’s driven by the socialized need or acceptance that the U.S. has to be a leader in artificial intelligence. And to do that, you need 24/7 power.”
Hyperscalers (the large tech companies such as Amazon and Google) have an advantage because they are already in the process of building out data centers near power centers or have applied to build their own power sources such as small nuclear reactors.
Capitalizing the Build
All of this requires, in a word, funding.
There are two roles the banks play in financing a data center: lending via BBS – (borrowing base style) for the land and construction for high-quality credits or an “originate to distribute” project finance model where a bank finances the construction with a takeout by investors on the other side via asset back securities (ABS) or commercial mortgage-backed securities (CMBS), said Ryan Atkinson, Chief Credit Officer, Mizuho Americas.
The amount of capital cascading into the sector is growing in tandem with data volumes. There is an estimated “$25 billion a quarter in data center deals” being project financed right now, notwithstanding borrowing based facilities, Sarah Kanes, Managing Director, Real Estate, Mizuho | Greenhill, told attendees.
“We as a bank, we can’t absorb all of this,” Atkinson said. “So the capital markets have to remain open for all of this to work.”
Fortunately, the capital markets are open for data center funding, said Xavier Auge, Executive Director, Structured Credit Group Mizuho | Greenhill. “Data centers have these unique benefits that they can tap different type of capital markets. It’s pretty rare. They can tap the ABS market, they can tap the 144 Reg S market, they can tap the CMBS market, and they can tap the USPP market.”
Assuming smooth construction and functioning once up and running, data centers can eventually generate bond-like returns for long-term investors, said Noel Purcell, Head of Real Estate, Mizuho | Greenhill, who moderated the panel.
Private markets (equity and credit) have already stepped up to fund data center builds where public markets may not take on the risk. Data centers offer infrastructure fund investors stable cash flows until they are ready to tap the equity markets in an IPO, spin it out or sell it. Private credit is another avenue for development projects looking for pre-lease, speculative lenders.
U.S. Market and Beyond
The U.S. has been ahead of the curve in data center development. Of the (total) 8 gigawatts of data center capacity deployed, 6 gigawatts were in the U.S.
Asia Pacific is not far behind in development. Blackstone recently acquired AirTrunk, a data center in Asia Pacific and the firm’s largest ever transaction in the region. Mizuho acted as Lender and Arranger for AirTrunk’s $393 million five-year loan facility to expand its data center in Johor, Malaysia, after the Blackstone acquisition.
Latin America is very dependent on where electricity is located for data center capacity with each nation having a different power scheme and source, noted Fernando Soriano, Head of Latin America for Mizuho | Greenhill. For example, Chile has a lot of geothermal resources, he said.
Europe is constrained by geography, energy capacity, cost and regulatory schemes. The cities known as the “FLAP-D” (Frankfurt, London, Amsterdam, Paris and Dublin) are more mature markets, but new construction is expected in Italy, Portugal and Spain, notwithstanding the varying environmental regulations which may hinder development.
Conclusion
The world will continue to see industries converge to meet the demand for data centers to support the technology revolution. While data center expansion is still in early innings, the expectation is that hundreds of billions of dollars in capital will continue to be needed to support and facilitate the continued rollout of AI infrastructure.
To view more content from the Mizuho Symposium, please click here.