Artificial intelligence may seem intangible at first glance, but there is a huge physical supply chain behind it, and at least for now, that’s where the money is to be made.
The chips, developed by Nvidia, AMD and others and mostly manufactured by Taiwan Semiconductor Manufacturing Company, have countless parts and suppliers. The chips must be organized and connected by another series of components. All of this technology then has to be built into a purpose-built building. An AI-driven boom is happening at every level of the chain, including the big concrete box that everything is built into.
Although data centers often look like simple boxes, they are anything but simple. Cooling and energy management for these structures, measured in megawatts rather than square meters, are key and represent a whole other category of technological innovation.
Although Amazon, Google and Microsoft own much of the cloud capacity, new entrants have been entering the market regularly since the advent of AI cloud computing.
DataBank is one of the few players able to spot trends in the big tech industry, as the 19-year-old company leases space in more than 65 data centers in 27 markets. The company has faced new competition as Blackstone and industrial real estate giant Prologis have followed the prospect of revenue into the space.
But DataBank CEO Raul Martynek has experienced boom phases before.
In the early 2000s, during the dot-com boom, Martynek ran a company that was a leader in dedicated Internet services. Today, he is once again building the rails on which new technologies run – and is an eyewitness to how the biggest players, with contracts that run years into the future, are vying for dominance in the cloud.
Martynek spoke to BI about the need for computing power for AI, hardware issues, and the still unanswered question of return on AI investments.
These questions and answers have been edited for clarity and length.
BI: How have you experienced the demand for AI data centers over the last 18 months?
We are data center developers, so we build the buildings that house all the equipment. This industry has been around for about 25 years, since the dot-com bubble, and it’s growing steadily.
Over the past decade, the biggest influence has been the growth of public clouds. These didn’t exist in 2010, but today they account for hundreds of billions in revenue. Hyperscalers are building many of their own data centers, but are also outsourcing much of this work to companies like DataBank.
ChatGPT was in November 2022, then 2023 came and actually the first quarter and a half were relatively slow. We wondered what was going on. Then in mid-2023 there was a tsunami of demand.
The hyperscalers and other major technology companies have started to accelerate their AI initiatives. Every time Nvidia sells a GPU, it has to go into a data center, so there is a direct correlation between the number of units shipped and the need for data center capacity.
In fact, we’ve seen sustained demand outstripping supply for about 12 months. A year ago, data center inventory was tight — it’s not that people are overbuilding in this space because it’s expensive — but there was capacity. In the last 12 months, prices have doubled and supply has dropped to the point where vacancy rates are in the low single digits.
DataBank CEO Raul Martynek database
BI: Data centers are planned years in advance, so you’re constantly thinking about the future. What year do you have in mind most of the time?
I think it’s a question of split thinking. I’m thinking about this year because we have existing inventory and we’re focused on staying within our budget for this year, but at the same time we have to think 24 to 36 months ahead.
For example, we are developing data center capacities of over 800 megawatts. They will come online in 2025, 2026, 2027 and partly in 2028. So we know roughly what we will be doing in the next three and a half years.
It all starts with the land.
We acquired the land 12 or 18 months ago. We spent a year planning the property, getting it permitted, zoning it, negotiating the power contract with the utility. To give you an example, one of our new developments is 95 acres in Atlanta. It’s a 120-megawatt data center campus – we’re just starting to break ground on it, and next year the first building will be coming out of the ground, and then by the second quarter of 2026 we’ll have customers.
In our business, we call this a pre-lease, and pre-leases were pretty common in the last few years before ChatGPT. The earliest pre-lease was 12 months. Due to AI-driven demand and the scarcity of data center capacity, that window will extend to 24 months.
In response to demand for AI, Databank is building its fourth data center in the Atlanta area. database
Databank’s next data center in Atlanta will consume 120 megawatts of power and occupy 95 acres of land. database
BI: Recently it was announced that some Nvidia chip models delayed or cancelledTo what extent do fluctuations in the AI hardware market affect your planning and how disruptive are they to your customers?
It doesn’t matter. We are obviously very conscious of that because with these developments we can step on the gas in terms of the speed at which we bring capacity online and we can also extend it a little bit. So we want to moderate that depending on what we hear from customers.
I think customers are upset in some ways, but also relieved in others, because there have been a lot of complaints about Nvidia wanting to release a new architecture every year. People who buy these chips can’t engage with them and use them to their full lifespan.
But if we build, say, a 40-megawatt data center, there are ten 4-megawatt data halls. There are ten small data centers in the building itself. If we were to build a data center without a pre-lease, we would only equip the first four megawatts of it. So you’re staggering your capital.
BI: When it comes to data center deals, do you think about the ROI associated with AI?
This is a fantastic topic, and this is especially true for me, having been in the industry since the mid-90s and living through the dot-com crash. I think everyone realized that the internet was real, it’s just that the business models at the time weren’t profitable, and so things went downhill. You could argue that the same dynamic could happen with GenAI.
We certainly track this in terms of whether we believe our clients are getting a return on their investment.
It’s early days, but this technology is truly revolutionary. Frankly, it’s probably one of the most disruptive technologies since the internet or maybe since cell phones. We’re going to look back on this in seven years and say, “Oh my God, that was so primitive, wasn’t it?”
Some companies can demonstrate an ROI, but it is certainly not consistent at this point. However, that is to be expected when adopting such a transformative technology.
We obviously check the customer’s creditworthiness. Like any landlord, we want to make sure our tenants are viable for the duration of the lease, right? The good news is that this technology is so sophisticated and so expensive that you have to be relatively financially stable and have the technical skills to use it.
We’re seeing more and more larger companies – hyperscalers and large technology companies – adopting these GPU deployments, and I think we’re going to see a lot more of that in the next year.
If a company was founded just four months ago, got an allocation of a few thousand GPUs, and wants to use five megawatts of data center space, we don’t do it. We let someone else support it. We stick with tenants with very good credit, and if they are early stage companies, they are well financed.
Do you have a tip or insight you would like to share? Contact senior reporter Emma Cosgrove at [email protected] or use the secure messaging app Signal: 443-333-9088
