The 'Whale Hunting' Era Of Data Center Development May Be Ending

The AI-driven data center boom of the past two years has been defined by hyperscale campuses in remote locations with hundreds of megawatts or even gigawatts of computing capacity — far larger than data centers have ever been built before.

But now, leaders from across the sector say the pendulum is swinging back in the other direction, with a growing share of development dollars flowing toward what are now considered small data centers. 

Five years ago, a data center with 100 megawatts of capacity would have been among the largest such facilities in the world. New data center projects for tech giants like Amazon, Google, Microsoft and Meta were measured in dozens of megawatts, not hundreds.

Today, all of the world’s largest technology companies are operating or in the process of developing multiple campuses that will have at least a gigawatt of capacity — an unprecedented jump in scale driven in large part by the computing needs of artificial intelligence.

Among the gigawatt-scale campuses in operation today are Amazon's $11B Indiana complex, xAI's Colossus data center cluster in Memphis, Tennessee, and Meta’s Prometheus campus in Ohio. Dozens more are under construction or being planned by the major hyperscalers and AI-model makers. 

With tech giants hungry for enormous blocks of power for AI training hubs — and increasingly agnostic about where those hubs are located — many third-party data center developers have shifted their businesses toward delivering these megacampuses, often far outside the industry’s traditional markets. 

A wave of new development entities focused on this segment — carrying varying degrees of credibility — has also emerged over the past 24 months. Many have announced plans for multigigawatt projects despite lacking commitments from major tech tenants.

This parade of ever-larger project announcements, combined with widespread skepticism over how many will ever come to fruition, has birthed a tongue-in-cheek piece of industry jargon: “bragawatts,” a reference to speculative capacity announcements that are more about publicity than any realistic development prospects.  

Yet the bragawatt gold rush may have already hit its peak, industry leaders said last week at Bisnow’s DICE: National event, held at the Bethesda North Marriott.

The smart money, they said, is starting to focus not on remote megacampuses but on smaller facilities of 100 MW or less closer to major population centers. 

“We've definitely seen a shift, with clients seeking out smaller deployments for various reasons,” said Josh Rudin, a data center attorney who co-chairs the digital infrastructure practice at Mintz. “We've definitely seen the spike in 20 MW to 90 MW, with some of the bigger players wanting smaller sites closer to primary downtown locations.”

The primary force driving demand for smaller data centers, several DICE panelists said, is the growing need for AI inference computing: the systems through which users interact with AI models in real time.

While the initial AI data center boom mainly saw the construction of facilities to train large language models, the rapid corporate adoption of AI products and services is driving a sharp rise in demand for inference computing.

Inference workloads typically need to be close to large population centers where the users are and in facilities where data can be transferred easily between different cloud providers and networks. That means a growing share of AI computing will need to be hosted in smaller facilities close to cities, not on rural megacampuses.

This shift is well underway. Data center REIT Equinix reported that inference workloads accounted for 60% of its new bookings last quarter, a stark increase over the year prior. Chipmaker Nvidia reported a similar shift toward inference in its sales numbers. 

Some of this demand is from corporate data center tenants building their own AI infrastructure, but it is also from the hyperscalers themselves, which are looking to deploy cloud computing edge nodes to serve customers in major population centers. Panelists said they expect Big Tech demand for capacity in the smaller “edge” data centers will accelerate as tech giants come under pressure from Wall Street to grow their user bases and monetize their AI infrastructure investments. 

“It’s the next wave of AI that we’re excited about. It’s going from training to inference, and ultimately, those latency-sensitive parts of the application will need to sit proximate to populations,” said Michael Borchetta, senior managing director and head of North American transactions for Harrison Street. 

“You can actually practically deliver 30, 40, 50 megawatts in proximity to population centers and do that on grid power in a reasonable time. To us, that’s where AI goes from here.”

There was broad consensus among DICE panelists that there is greater opportunity for developers in these smaller facilities than in gigawatt-scale campuses. 

While securing hundreds or thousands of megawatts on the timelines demanded by hyperscalers has become increasingly difficult — with yearslong wait times for grid connections driving developers toward complex solutions like behind-the-meter natural gas generation — sites with 100 MW or less remain far more attainable in many high-demand markets.

These “infill” opportunities are often former manufacturing plants or other industrial sites with existing grid connections that can be redeveloped as data centers. Because they already have power available on-site, they can often be brought online quickly. As demand shifts toward inference computing, adaptive reuse is likely to become an increasingly practical and common path for delivering new capacity. 

“There's still a lot of developers that are what we call ‘whale hunting’ within the industry — trying to land those large, multihundred-megawatt or gigawatt deployments — but those are getting harder and harder to come by,” said Alex Kang, vice president of acquisitions at Legacy Investing.

Legacy Investing's development strategy is increasingly focused on projects between 20 MW and 90 MW, Kang said. 

“Largely, you're able to get power from former advanced manufacturing facilities, large heavy industrial facilities like steel manufacturing and even refining facilities,” he said.  

And even for greenfield projects that require a new grid interconnection from a utility, investors said the process is typically easier and faster at smaller scale.

Utilities across the U.S. have been inundated with power requests from data center developers asking for hundreds or thousands of megawatts, forcing power providers to make the interconnection process more complex and expensive.

Amid this flood of massive power requests, many utilities are relieved to get applications from smaller projects that are likely to move forward, and they will often bump those to the top of the queue, said Adam Krupp, managing director at Wharton Equity Partners. 

“It's crazy because 80 MW is an enormous amount of power, but when you say to the utility that you want 50, 70, 80 megawatts, their response is: ‘Oh, thank God, you don't want a gigawatt,’ and they welcome you in,” Krupp said. 

Projects of this size are not only more feasible from a power acquisition perspective but also make it easier for developers to access capital.

Industry leaders said there is a far larger pool of lenders and equity investors able to participate in deals at this scale compared to gigawatt-scale projects. 

This is due, in part, to the real estate fundamentals of projects close to major economic hubs compared to more remote hyperscale sites, Harrison Street’s Borchetta said.

He and other panelists framed smaller edge data centers as a lower-risk bet than the massive remote campuses built primarily for AI training. If the data center demand landscape were to take a sharp downturn, the projects most vulnerable to that demand pullback would be the isolated sites built for a narrow set of use cases. 

By contrast, smaller data centers near population centers are considered more resilient because they can support traditional cloud computing workloads or be repurposed into other industrial uses. This carries weight with investors making long-term bets on a new and uncertain AI infrastructure landscape. 

The relative ease with which developers can access capital for these smaller projects, panelists said, is largely because their lower costs open the door to a much broader pool of lenders and investors, many eager to gain exposure to data centers but effectively priced out of hyperscale megaprojects with multibillion-dollar price tags. 

“A lot of the value in industry right now is pursuing those sub-100 MW deals because the check size is just smaller,” Kang said. “Calling a half-billion-dollar deal a small deal feels absolutely insane to me, but there's significantly larger equity pockets that are able to do those deals.”