Hyperscaler capex is surging, emptiness has all however vanished, and AI is pushing information centre demand far past conventional hubs – however questions stay over energy, individuals, and long-term returns as growth spills into secondary and tertiary markets.
In sum – what to know:
Mega buildout – hyperscaler spending on AI infra is at a step change, with the highest 5 set to deploy greater than $600bn in 2026 alone, pushed largely by AI coaching workloads.
Important provide – emptiness charges are under two % in main US markets and successfully zero in hotspots like Northern Virginia, forcing growth into new geographies.
Regional focus – secondary and tertiary markets are in focus, however energy availability, workforce attraction, and the longer term mixture of AI coaching versus inference investments.
There was a great panel dialogue final Sunday (January 18) in Honolulu, the primary day of PTC’26, concerning the “generational build-cycle” within the US information centre market, teed-up properly by Joe Valenti, managing director and co-head for media and telecom at Financial institution of America. “The demand profile is big,” he mentioned, opening the dialogue for a well-attended panel comprising data-centre analyst, consultancy, and operator corporations.
He reeled off some top-line stats: the highest 5 hyperscaler corporations (Amazon, Microsoft, Google, Meta, Apple) will make investments $602 billion of capital in information centre infrastructure in 2026; 4 of them will spend $100 billion every – which represents “north of fifty %” of their income in some instances. “It’s a step operate in historic capital deployment,” mentioned Valenti. They are going to go from 4 % of complete US vitality consumption in 2024, to 10 % by 2030, he mentioned.
Notably, there’s just about zero over-capacity; the emptiness charge for US actual property, whether or not powered shell amenities or absolutely fitted areas, is simply 1.6 % in main markets, he mentioned – so just one.6 % of complete data-centre stock is presently unoccupied in historic AI-infrastructure heartlands. “It’s effectively lower than one % in sure markets – Northern Virginia being the obvious one. So we’ve got this enormous imbalance of demand relative to provide.”
New regional markets
So demand is exploding and provide is sort of exhausted. And AI is the explanation, linked principally to the duty to coach large frontier language (LLM) fashions. Cue the dialogue, correct, and a query from Valenti concerning the unfold into secondary and tertiary markets, to offset rising demand in energy-constrained main markets. David Liggitt, chief at information centre analytics agency Datacenter Hawk (stylised datacenterHawk), additional defined the demand profile.
He mentioned: “[The] main markets within the US have been Chicago, Phoenix, Dallas, Atlanta – markets like this. We’ve seen secondary markets – like Minneapolis, Portland, Houston, Nashville – develop over time. And we’ve [also] seen progress into tertiary markets within the final 24 months. [We] tracked 15.6 GW of demand absorption in 2025 – which is the flamboyant time period in the true property business for demand. To place that in context: it was 6.8 GW in 2024, and three GW in 2023.
“So we’ve got gone from 3 GW to six.8 GW to fifteen.6 GW… The transactional worth of that 15.6 GW [of capacity] is over $2 trillion – for these leases, when put collectively. So the worth is critical.” Apparently, Liggitt advised new deployments in so-called tertiary markets are outgunning their equivalents for capability in additional acquainted territories. “If the whole market in Portland is 500/600 MW of capability, an information centre… going into North Dakota could be twice that.”
The that means is slightly unclear, however the sense is the secondary market, in some areas, could possibly be usurped – as long as energy is offered in additional rural websites. “The query [is whether] these tertiary markets… can really do greater than [the] preliminary deployments – contemplating [they will] stress the utility system. There’s rather a lot [to do] for a few of these secondary markets to… full these tasks. It will likely be fascinating to see what occurs… over time.”
Energy and personnel
Jason Nance, answerable for consumer options at CBRE Knowledge Middle Options, a service group inside industrial real-estate providers and funding agency CBRE, responded: “The tertiary markets are increasingly within the combine… [and] may very effectively change into main markets.” However part of the problem, past electrical energy, is (hu)man energy – and learn how to entice staff to the back-end of nowhere to employees a giant rural AI manufacturing unit.
“All people is aware of [about] Abilene (in Texas, the place the Stargate undertaking is predicated)… [but] individuals neglect that the infrastructure for individuals to assist these tasks [has] to be there… The most important downside is getting individuals to maneuver to those tertiary markets. Individuals take a look at the ability and the community, however… you continue to want individuals on the finish of the day. And that will stymie progress in these tertiary markets… Typically it’s a bridge too far… to maneuver to North Dakota.”
Nebraska-based 1623 Farnam runs an interconnection facility in Omaha, thought-about to be a secondary market; it’s “one of the crucial linked in all the Midwest”, internet hosting 60-odd ‘carriers’ and greater than 35 broadband suppliers. “There are eight hyperscale campuses inside 40 miles of our facility,” defined Invoice Severn, president and chief government on the agency. “All of them are asking for double the ability that they requested for 3 years in the past.”
1623 Farnam is shut with the native energy firm, and sees the problem for either side. “How do they resolve that? As a result of these will not be new builds; they’re present and preexisting… [And] should work diligently with the ability firm to seek out these nooks and crannies the place slightly energy could be squeezed out.” Past the vagaries of AI planning, there’s an existential query about how such huge tasks pays out after the large AI work is accomplished.
Coaching and inference
Valenti defined: “Most deployments are getting used for generative AI, proper? We all know we’re in an arms race [where] a number of events are attempting to create the perfect mannequin … However we’ll get to a degree the place the fashions are skilled – quote/unquote – sufficient. After which we can have large capability in locations like North Dakota, Mississippi, and Louisiana… [which will need to absorb] issues aside from generative AI. How do you guys take into consideration that?”
How do information centre operators, personal fairness companies, and sovereign wealth funds assess long-term returns on new data-centre builds, particularly in distant tertiary markets – a great distance from the sting, the place all the true AI motion is; the place the information is generated and acted upon, and AI inference more and more takes up residency? Nance responded: “There’s a whole lot of exuberance round AI, and if we’re not cautious it may well change into irrational exuberance, proper?
“All people with a greenback and a chunk of land… desires to construct an information centre. However anyone placing cash into the market wants to grasp the basics… Even the storied [AI] corporations on the market, that are within the information on a regular basis, should monetize [their infrastructure investments] on the finish of the day… [They have to ask] if the funding… is the correct one – as a result of they’re not all the identical. They’ve to pay attention to the irrational issues.”
Liggitt rejoined: “One other time period we’re listening to fairly a bit… is fungibility – associated to websites being acquired, and whether or not they into different use instances down the street? That’s most likely the perfect [question to ask]: whether or not there are different [AI] use instances that… [will] emerge when these huge fashions are skilled [that will] take that sort of energy…. [Because] the quantity of energy in some extra rural markets is… onerous to fathom. I imply, these are such giant quantities.”
He added: “The opposite factor [to consider] is the wave and the maturity of inference, and the way that impacts our area geographically – the place it impacts issues; how near the larger cities does inference must be, because it matures over the following 5 years. How far out can [data centres] actually go? As a result of that can even have an effect on the place energy is deployed.” Certainly, there’s nonetheless some debate concerning the localisation of compute energy for AI inference.
One other factor: networks
Sunday afternoon, and PTC has solely simply began; Valenti has had half a dozen completely different conversations with information centre operators already, and there’s little actual consensus about the place to place inference workloads, he says. “I’ve had suggestions that there are many non-latency-sensitive inference workloads that don’t have to go anyplace else; that they are often deployed in among the areas we’re speaking about. Issues like video rendering, for instance.”
The query is put to Severn at 1623 Farnam. “All of us learn the headlines about these billion-dollar offers, and everyone talks about AI inference or coaching. However the bit that will get misplaced is [in the network] in-between. It’s simply not practically as horny. However the place we’re successful [with AI] is within the community – as a result of that element must be in place. And the networks are coming to tertiary markets, the place we’re an interconnection facility.”
Valenti let the community dialogue go, prompting Severn with a query concerning the curiosity in and returns on fibre interconnectivity – past simply the large discuss AI information centres and coaching fashions. Severn responded with an perception concerning the technique at 1623 Farnam, and who to let into its service lodge / meet-me room set-up in Omaha, and who to maintain out. “If a consumer doesn’t add to the ecosystem within the constructing, we are saying no,” mentioned Severn.
“Every thing must be additive. Our shoppers common 13 cross-connects to others within the constructing. Which reveals it’s a extremely sticky atmosphere. And if you consider all of the AI startups, our constructing could possibly be full proper now if we needed to. We simply turned down somebody that needed the remainder of it, and will have generated a number of million {dollars} price of money circulation a 12 months. However it might take up 40 % of the area, and wouldn’t give us the identical return.”
