Has Data Center Development Hit a Wall?

This article first appeared in the CPE Sneak Peek Newsletter. Subscribe now for early access to exclusive content, in-depth analysis and key trends!

As a new generation of tech infrastructure unfolds, new data center development risks are beginning to test the sector’s unprecedented growth. Construction activity continues to reach new highs each month, but an observable gap between announced and deliverable capacity is also taking shape. At the same time, talks of an AI bubble are getting increasingly louder, adding pressure to the sector’s long-term outlook.

Community backlash against data centers is playing a major role in this next chapter. New York recently approved a one-year moratorium on data center projects larger than 20 megawatts, a first for the nation. Other states are targeting tax exemptions, such as Arizona, where a three-year pause on such abatements was implemented last month.

Several major projects across the U.S. have also fallen through in recent months, some outright canceled, others scaled down from their initial scope. Oracle and OpenAI canceled expansion plans for their Stargate data center campus in Abilene, Texas; what was to be Arizona’s largest hyperscale campus, La Osa Data Center and Energy Campus in Pinal County, was scaled down by 80 percent; Microsoft scrapped its 244-acre data center in, Caledonia, Wis., after community pushback; and other projects have faced similar challenges.

Together, these circumstances create overarching headwinds that will have a deep impact on the sector going forward. Still, project-level failures and scale-downs are not unusual in a market expanding this quickly.

“Every project has a specific story that would explain any deviation from the original strategy,” Green Street Head of Global Data Center & Tower Research David Guarino told Commercial Property Executive. The industry, he added, is undergoing a massive transformation, with projects of this size and scale being attempted for the first time.

Demand is not the problem. Mammoth projects are still being leased by compute providers such as Amazon, Google and Microsoft. Instead, the unfolding story hinges on the type of projects that can actually get built, how quickly they come online and whether developers can adapt to a more demanding execution environment.

Project announcements get ahead of diligence

A growing number of newcomers are entering the data center sector with what Pond Robinson & Associates President Justin Lia describes as a “wildcatter mentality.” Drawn by the potential for quick returns, some developers underestimate the complexities and intricacies of the sector, overlooking the millions of dollars in soft costs. When that is paired with premature announcements about future capacity, projects can stall at various points in the development process.

“‘It’s just a warehouse, Justin.’ That’s what I heard. I had a set of construction documents for a million-square-foot, single-tenant warehouse in the same part of the world as a hyperscaler project,” Lia exemplified. “The warehouse had 200 sheets of drawings with the specifications incorporated into them. The hyperscaler data center project had 4,000 sheets of drawings with an 8,000-page spec book.”

The cost comparison is just as stark. For same-size structures, a warehouse can cost about $60 per square foot, while a powered shell can go as high as $600 a square foot. That translates to roughly $2.5 million per month in construction draw for one type of project, compared with $2 million a day, six days a week, for years, for the other, according to Lia.

In other words, execution itself becomes a real development risk for data center projects when mishandled. Within that equation, power that is not properly vetted often becomes the deal killer, believes JD Jones, senior vice president of data center services at Pond Robinson & Associates.

Power constraints are already top of mind for developers, but solving them can take far longer than inexperienced players expect. Jones noted that the market is starting to respond to these challenges. For example, Texas is in the early stages of upgrading its power transmission capabilities to 765-kilovolt lines, a project estimated to cost between $14 billion and $40 billion.

The chip cycle threat

Another factor adding complexity to data center execution is the chip cycle and hardware timing. NVIDIA currently releases a new AI GPU platform roughly every year. Int he long run, that means more efficient designs and greater computing capacity, but in the near term, the accelerated pace is introducing a new layer of challenges into the data center ecosystem. A medium-size data center can take up to three years to come online, so speed-to-market is gaining a new dimension for developers and compute providers.

Therefore, redesigns during the development process are becoming more common. A new chip architecture can affect floor loads, cooling requirements and other technical specifications, which can then spill over into power requirements, permitting and project timelines, according to Jones.

The semiconductor/chip supply chain is also under pressure from this pace. While TSMC and other manufacturers are rolling out more production capacity, those facilities take years to fully develop and meaningfully impact the market. Meanwhile, AI providers are racing to roll out better-performing models, which are becoming more power efficient—but that efficiency depends on new hardware, pressuring developers and data center providers to accommodate new architecture as quickly as possible.

Development risk mitigation as the new playbook

Under these circumstances, Lia and Jones both believe the need for overdesign is becoming more apparent. Some projects may not materialize in the way they were initially advertised, but elements that are difficult to retrofit—such as foundations, structural capacity, cooling pathways and redundancy—need to be addressed upfront because they are far more expensive to change later.

“The key assumption is delivery schedule,” Guarino added.

When evaluating a proposed hyperscale AI campus today, many financial assumptions are observable and relatively similar across projects, including construction costs, power delivery, tenant credit and exit cap rates. That makes speed-to-market one of the most pressured variables.

All these challenges are compounded by the legal landscape surrounding large-scale developments.

“Time will kill a project,” said Joe Restivo, managing partner at DarrowEverett LLP, when asked about the most common legal pressure points. “Apart from an outright prohibition through a moratorium, the zoning appeal process is the most effective means of derailing a project when there is local opposition.”

The significant uncertainty around high electricity demand and efforts to create special rate classes for data centers makes that landscape even harder to navigate because cost-shifting has a breaking point. While ratepayers should not subsidize data center development, Restivo said, if too many costs are pushed onto developers, some projects will become unfeasible.

As a result, risk mitigation is becoming essential for developers and investors at all stages of a given project. Chip obsolescence does not necessarily mean a building becomes worthless, but the accelerated hardware cycle can impact schedules, equipment orders, engineering contingencies and speed-to-market.

The rising need for evolving energy policies

Disputes over grid upgrade costs are already common and “transcend normal political or geographic lines,” as Restivo noted. Few places have sufficiently upgraded energy infrastructure to absorb the impact of large-scale data centers, and utilities are caught between pressure to achieve public policy goals and pressure to keep rates low.

Furthermore, a well-organized opposition movement combined with intricate regulatory approval processes can create circumstances where a project is stalled for years.

“For example, if a project needs zoning approval and the appeal process has the potential to tie a project up for years, then, for relatively low cost, a project could be effectively killed,” Restivo added.

Still, demand for more AI infrastructure remains solid, at least for now. If these new development risks can be overcome, data center projects can continue to generate strong returns for companies and providers with the experience and balance sheets to execute them.

Downscaling trend?

The past few years have produced a significant number of projects with questionable scalability and feasibility, but overall demand is not disappearing. Rather, the market is shifting toward greater discipline, more scrutiny and more attention to the needs and voices of the local communities affected by these massive developments.

However, all of the risks outlined here—from execution challenges and accelerated hardware cycles to rising policy concerns and community opposition—are tied to one larger factor: whether AI companies can become profitable enough to sustain the infrastructure buildout they are driving.

OpenAI estimated that it will spend hundreds of billions over the next few years training its models, FastCompany and The Wall Street Journal both reported. While its revenue numbers are not publicly available, the LLM provider is unlikely to break even earlier than 2030, according to its own projections.

Given the compounding pressure exerted by all these forces, the data center sector will undergo a huge shift in the coming months. Demand will still be there, but the shape of that demand and the projects capable of meeting it will need to adjust to a new reality.