Investment from hyperscalers is increasing, vacancies have all but disappeared and AI is pushing demand for data centers well beyond traditional hubs – but questions remain about power, people and long-term returns as development ripples through secondary and tertiary markets.
In summary, what you need to know:
Mega construction – Hyperscalers’ spending on AI infrastructure is changing dramatically, with the top five expected to deploy more than $600 billion in 2026 alone, largely driven by AI training workloads.
Critical Supply – vacancy rates are below 2% in major US markets and zero in hotspot areas like Northern Virginia, forcing expansion into new geographies.
Regional focus – Secondary and tertiary markets are the focus, but energy availability, labor attraction and future combination of AI training and inference investments.
There was a good round table last Sunday (January 18) in Honolulu, the first day of PTC’26on the “generational build cycle” in the US data center market, nicely presented by Joe Valenti, Managing Director and Co-Head of Media and Telecommunications at Bank of America. “The demand profile is huge,” he said, opening the discussion in front of a well-attended panel of data center analysts, consulting firms and operators.
He revealed some high-profile statistics: The five largest hyperscalers (Amazon, Microsoft, Google, Meta, Apple) will invest $602 billion of capital in data center infrastructure in 2026; four of them will each spend $100 billion – representing “more than 50%” of their revenues in some cases. “This is a milestone in the historic deployment of capital,” Valenti said. They will increase from 4% of total U.S. energy consumption in 2024 to 10% by 2030, he said.
Notably, there is virtually no excess capacity; The vacancy rate for U.S. real estate, whether powered or fully serviced, is just 1.6% in primary markets, he said – so only 1.6% of total data center inventory is currently unoccupied in historic AI infrastructure cores. “It’s well under one percent in some markets, Northern Virginia being the most obvious. So we have this huge imbalance between demand and supply.”
New regional markets
Demand is therefore exploding and supply is almost exhausted. And AI is the reason, related primarily to the task of training massive boundary language models (LLM). Let’s quote the discussion itself and a question from Valenti on propagation in secondary and tertiary markets, to compensate for the growing demand in energy-constrained primary markets. David Liggitt, head of data center analytics company Datacenter Hawk (stylized datacenterHawk), explained the demand profile in more detail.
He said: “[The] The main markets in the United States have been Chicago, Phoenix, Dallas, Atlanta – markets like this. We’ve seen secondary markets – like Minneapolis, Portland, Houston, Nashville – grow over time. And we have [also] has seen growth in tertiary markets over the past 24 months. [We] This was followed by a demand absorption of 15.6 GW in 2025 – which is the fancy real estate industry term for demand. To put this in context: it was 6.8 GW in 2024 and 3 GW in 2023.
“We therefore went from 3 GW to 6.8 GW then to 15.6 GW… The transactional value of these 15.6 GW [of capacity] represents over $2 trillion – for these leases, when combined. The value is therefore significant. Interestingly, Liggitt suggested that new deployments in so-called tertiary markets are outstripping their capacity equivalents in more familiar territories. “If Portland’s total market is 500/600 MW of capacity, a data center… for North Dakota could be double that. »
The meaning is a bit fuzzy, but the feeling is that the secondary market, in some areas, could be usurped – as long as electricity is available in more rural locations. “The question [is whether] these tertiary markets… can actually do more than [the] initial deployments – taking into account [they will] emphasize the public service system. There are many [to do] for some of these secondary markets to… carry out these projects. It will be fascinating to see what happens…over time.
Power and personnel
Jason Nance, head of client solutions at CBRE Data Center Solutions, a services group within commercial real estate services and investment firm CBRE, responded: “Tertiary markets are increasingly involved… [and] could very well become primary markets. But part of the challenge, beyond electrical power, lies in human power – and how to lure workers behind the scenes from nowhere to staff a large rural AI factory.
“Everyone knows [about] Abilene (in Texas, where the Stargate project is based)… [but] people forget that the infrastructure needed to support these projects [has] to be there… The biggest problem is getting people to move to these tertiary markets. People look at power and the network, but… at the end of the day, you still need people. And that could hinder the growth of these tertiary markets… Sometimes it’s a bridge too far… to move to North Dakota.
1623 Farnam, based in Nebraska, operates an interconnection facility in Omaha, considered a secondary market; it is “one of the most connected in the entire Midwest,” hosting around sixty “carriers” and more than 35 broadband providers. “There are eight hyperscale campuses within a 40-mile radius of our facilities,” said Bill Severn, the company’s president and CEO. “All of them are asking for double the power they asked for three years ago. »
1623 Farnam is close to the local power company and sees the challenge for both parties. “How do they solve this problem? Because these are not new constructions; these are existing, pre-existing constructions… [And] I have to work diligently with the power company to find the nooks and crannies where some electricity could be extracted. Beyond the vagaries of AI planning, there is the existential question of how such large projects will be profitable once the major AI work is completed.
Training and inference
Valenti explained: “Most deployments are used for generative AI, right? We know we’re in an arms race [where] multiple parties are trying to create the best model… But we will get to a point where the models are sufficiently trained – quote unquote – quote unquote. And then we’ll have massive capacity in places like North Dakota, Mississippi and Louisiana… [which will need to absorb] things other than generative AI. What do you guys think? »
How do data center operators, private equity firms and sovereign wealth funds assess long-term returns from new data center builds, particularly in remote tertiary markets – far from the edge, where all the real AI action is happening; where data is generated and exploited, and where AI inference is increasingly taking hold? Nance responded: “There’s a lot of exuberance around AI, and if we’re not careful, it can become irrational exuberance, right?
“Everyone with a dollar and a piece of land… wants to build a data center. But anyone who invests money in the market needs to understand the fundamentals… Even the most famous [AI] companies, which are constantly in the news, must monetize [their infrastructure investments] at the end of the day… [They have to ask] if the investment… is the right one – because they are not all the same. They must be aware of irrational things.
Liggitt countered: “Another term we hear quite often… is fungibility – relating to sites being acquired, and whether they will be used in other use cases in the future? That’s probably the best one. [question to ask]: if there are others [AI] use cases that… [will] emerge when these large patterns are formed [that will] take this type of power…. [Because] the amount of electricity in some more rural markets is…difficult to understand. I mean, these are such large amounts of money.
He added: “The other thing [to consider] is the wave and maturity of inference, and how that affects our space geographically – where it affects things; how close inference needs to be to major cities, as it matures over the next five years. How far can we [data centres] really go there? Because it will also impact where the energy will be deployed. » Indeed, there is still debate about where the computing power for AI inference will be located.
Something else: networks
Sunday afternoon, and PTC has just started; Valenti has already had half a dozen different conversations with data center operators, and there is little real consensus on where inference workloads should be placed, he says. “I’ve received feedback that there are many non-latency sensitive inference workloads that don’t need to be moved elsewhere; that they can be deployed in some of the regions we’re talking about. Things like video rendering, for example.”
The question is posed to Severn at 1623 Farnam. “We read all the headlines about these billion-dollar deals, and everyone’s talking about AI inference or AI training. But what gets lost is [in the network] between the two. It’s just not as sexy. But where we win [with AI] is in the network – because this component must be in place. And the networks are arriving in the tertiary markets, where we are an interconnection tool.
Valenti left the discussion on the network, prompting Severn to ask a question about interest and feedback on fiber optic interconnectivity – beyond the big discussions about AI data centers and training models. Severn responded with insight into 1623 Farnam’s strategy, and who to let into his transport hotel/meeting room in Omaha, and who to keep out. “If a client doesn’t add anything to the building ecosystem, we say no,” Severn said.
“Everything has to be additive. Our customers have on average 13 cross-connections with other people in the building. Which shows that it’s a really tricky environment. And if you think about all the AI startups, our building could be full right now if we wanted it to be. We just turned down someone who wanted the rest and could have generated several million dollars in cash flow a year. But that would take up 40% of the space and wouldn’t give us the same return.”
