For much of the past two years, the data center industry’s central question was whether artificial intelligence demand could sustain the unprecedented scale of infrastructure being announced in its name.

That is no longer the most urgent question.

The more immediate concern is whether developers can assemble the power, land, cooling systems, interconnections, capital, political support and community acceptance required to convert that demand into operating capacity.

“AI infrastructure has moved from being a fast-growing demand segment to becoming the organizing principle of data center development,” said Colby Cox, Managing Director for the Americas at DC Byte, during a recent appearance on the Data Center Frontier Show podcast.

The phrase captures a fundamental shift. AI is no longer simply one workload category competing for space inside the conventional data center market. It is beginning to determine where facilities are built, how campuses are financed, how electrical and mechanical systems are designed, and which regions can realistically participate in the next phase of digital infrastructure growth.

The market has also moved beyond incremental expansion. Campuses planned around hundreds of megawatts—and increasingly multiple gigawatts—are no longer treated strictly as outliers. They are being conceived from the beginning around GPU density, liquid cooling, accelerated deployment and enormous concentrations of electrical load.

Behind that transformation lies a second, increasingly decisive reality: Demand may be abundant, but deployable power is not.

Executable Power Separates Announcements From Infrastructure

From DC Byte’s market-intelligence vantage point, the dividing line between announced capacity and capacity likely to reach operation is increasingly straightforward.

Has the power been secured? When can it be energized? Can the grid—or an onsite alternative—support the project’s intended phasing?

“The market is no longer constrained primarily by demand or capital,” Cox said. “It is constrained by executable power.”

That distinction matters because a headline capacity figure says little about whether a project can be delivered. A developer may control land, possess financing and identify a willing customer, yet remain years away from an electrical service date capable of supporting the proposed load.

DC Byte classifies projects across four broad stages: live, under construction, committed and early stage. Cox said the firm has recorded a roughly 20% increase in the volume of projects remaining in the committed or early-stage categories over the past several years.

Power availability is a major reason.

The resulting backlog is pushing development along two general pathways. The first leads toward markets where existing generation, transmission or industrial power infrastructure can be expanded or repurposed. The second leads toward projects where developers are prepared to create significant portions of the power stack themselves.

That helps explain growing interest in parts of Texas, Indiana, Ohio and the broader Midwest. In some locations, utilities and state governments are attempting to position large data center loads as catalysts for broader grid investment. In others, the proposed campuses are so large that new energy systems must effectively be developed around them.

Power availability has therefore become more than a site-selection criterion. Cox described it as one of the strongest predictors of development velocity.

The implication is that the AI pipeline should no longer be evaluated strictly in megawatts or gigawatts. It should be evaluated in credible energization pathways.

From Data Center to AI Factory

The distinction between a traditional data center and an AI factory becomes clearest, Cox suggested, at the intersection of power density and capital intensity.

Traditional facilities have generally been organized around availability, occupancy and the flexible delivery of space and power to a range of customers. AI factories are organized more explicitly around compute output.

That difference produces larger campuses, denser computing halls, liquid-cooled systems, compressed construction schedules and more direct relationships among operators, utilities, hardware vendors and capital providers.

It is also changing how capacity is underwritten.

Leading hyperscalers and AI platforms are securing infrastructure years ahead of expected consumption, supported by strategic partnerships, vendor financing, debt and lease commitments that can extend 15, 20 or even 35 years.

The result may be a more stratified data center market.

At one level, hyperscalers, neocloud providers and well-capitalized developers are pursuing enormous AI opportunities measured in hundreds of megawatts or gigawatts. At another, colocation providers and enterprise facilities continue to support conventional cloud, network, storage and business workloads that remain essential to the digital economy.

Both markets are substantial. But they increasingly operate according to different assumptions about scale, capital, density and risk.

DC Byte’s project data illustrates how quickly the upper end has expanded. At the beginning of 2023, the firm tracked three committed projects of at least 900 MW and seven more in the early-stage category.

Today, Cox said, DC Byte is tracking 17 committed projects above 900 MW and another 49 in early-stage development. Fifteen of those early-stage projects exceed 2 GW, with several approaching 10 GW.

Not all will be built as announced. Some will be phased, resized, delayed or abandoned. But the magnitude of the pipeline demonstrates how thoroughly AI has altered the industry’s scale horizon.

The gigawatt has moved from abstraction to planning unit.

Rack Density Is Becoming the Architecture

The campus-level numbers are extraordinary. The transformation inside the white space may be even more consequential.

“Rack density assumptions are changing faster than most of the market’s design language has even caught up,” Cox said.

Average densities of approximately 6 kW per rack characterized much of the industry when Cox began working in the sector. Thirty-kilowatt racks once represented an ambitious high-density deployment. The industry is now seriously planning around 100-kW racks, while some AI configurations are pushing toward 300 kW and beyond.

Those are not simply larger versions of familiar data halls. They represent a different infrastructure problem.

At those densities, total facility power tells only part of the story. Operators must be able to distribute, manage and cool the load at the rack and pod level. That changes electrical topology, busway selection, breaker coordination, floor loading, mechanical plant design, piping, controls, commissioning and heat rejection.

Cooling, in particular, can no longer be treated as an enhancement applied after the primary facility design has been established.

Direct-to-chip systems, coolant distribution units, rear-door heat exchangers, facility water loops and heat-rejection equipment must increasingly be considered elements of the base architecture.

The density transition also complicates space planning. A facility may possess sufficient aggregate megawatts but still be unable to deploy them efficiently if its rooms, distribution systems or cooling infrastructure cannot support the concentration required by current AI hardware.

That raises a difficult question for the installed base.

The Legacy Data Center Utilization Trap

Older facilities may not face a conventional power shortage. They may instead encounter a utilization problem.

“The issue is not whether an older site has enough total power,” Cox said. “It’s really whether it can hit the new rack densities without leaving parts of the hall underutilized or operationally compromised.”

A data hall designed for broadly distributed air-cooled loads may technically contain several megawatts of capacity. But concentrating that power into a much smaller number of AI racks can overwhelm local electrical paths, cooling systems or structural assumptions long before the building reaches its nameplate limit.

Operators may therefore confront an uncomfortable choice: undertake extensive retrofits, dedicate only portions of a facility to high-density deployments, or accept that some buildings will remain better suited to traditional workloads.

For multi-tenant colocation providers, the answer is unlikely to be an immediate conversion to all-liquid infrastructure.

Cox suggested that an initial design assumption anticipating liquid cooling across roughly 30% of capacity may provide a more practical starting point, particularly when paired with floor plans that allow liquid-cooled and air-cooled customers to coexist.

The underlying challenge is preserving optionality.

Colocation operators must accommodate today’s hardware without locking facilities into a single cooling architecture or density profile that may prove limiting during the next generation of systems. They must also maintain connectivity among different customer environments while avoiding stranded space, stranded power or unnecessary mechanical complexity.

Flexibility, once primarily a leasing and capacity-management advantage, is becoming a physical-design requirement.

The AI Map Expands Beyond Primary Markets

Power constraints and campus scale are also redrawing the geography of North American data center development.

Established markets will remain important because of their fiber density, customer ecosystems, labor pools and existing infrastructure. But the AI buildout is directing increasing attention toward secondary and tertiary locations capable of offering credible power pathways.

Texas remains a leading example, although the opportunity now extends well beyond Dallas-Fort Worth. Cox pointed to activity around Pecos, Abilene, San Antonio and Austin as developers search for combinations of land, energy and permitting support.

Louisiana provides an even more dramatic illustration. According to Cox, the state had approximately 9 MW of total data center capacity only two-and-a-half to three years ago. It now has several gigawatts in various stages of development.

That growth has been supported by recruitment incentives and coordination around utility generation and transmission.

Indiana has similarly benefited from proactive state and utility positioning. West Virginia has emerged through legislation and incentives intended to attract data center investment, drawing interest from companies including Google, Microsoft and Penzance. Pennsylvania is receiving more attention, while development around Atlanta continues to spread farther into surrounding Georgia markets.

What these locations share is not simply inexpensive land.

The strongest emerging markets are those where power, policy, development and community interests can be brought into alignment. The winning jurisdictions will be those capable of treating data center growth as a coordinated infrastructure strategy rather than a series of isolated real estate projects.

Geography may shift again as AI workloads evolve.

Large training clusters favor enormous concentrations of computing and power. More widespread inference and agentic AI applications could eventually place greater emphasis on distributed capacity, network latency, fiber availability and middle-mile connectivity.

The next AI infrastructure map may therefore be more geographically diverse than the first.

Power First, Politics Second

The greatest obstacle to that expansion may not be technical.

Local resistance has become a material factor in determining where large-scale AI infrastructure can be financed, permitted and delivered.

Cox said investors and operators are no longer evaluating markets solely through vacancy, pricing, competitive capacity and utility rates. They are examining surrounding counties to determine how many projects have been announced, how many remain active and why others encountered opposition.

Was the concern air quality? Water use? Noise? Grid cost? Land consumption? A lack of credible local economic benefits?

Those questions are increasingly part of project diligence.

“The AI campus map is being redrawn by power first and politics second,” Cox said.

Community resistance does not eliminate data center demand. It narrows the geography in which investors believe that demand can be served with acceptable risk.

The jurisdictions most likely to succeed will be those where utilities, governments and developers engage early enough to explain—and demonstrate—how a project functions as an infrastructure asset rather than a large private load imposed on local residents.

That requires more than communications strategy. It requires credible plans for energy procurement, grid investment, environmental performance, construction, tax revenue and durable employment.

“Local acceptance is now part of power underwriting,” Cox said.

A project may look viable in megawatts while remaining weak on permitting, political support or community trust. The market is beginning to price that weakness as delivery risk.

Behind the Meter: Necessity, Not Ambition

The collision between AI demand and utility timelines has made behind-the-meter generation one of the sector’s defining strategic questions.

Cox does not believe most data center companies aspire to become power companies. They would generally prefer to purchase reliable utility power and concentrate on developing and operating digital infrastructure.

But preference does not eliminate necessity.

In some markets, onsite generation may support a long-term operating model. In others, it will serve as a bridge while transmission, substations and utility generation catch up with the requested load.

Natural gas currently dominates many near-term behind-the-meter proposals because it can provide firm power at scales and on schedules that intermittent resources alone cannot easily match. Battery storage and emerging energy technologies may increasingly complement those systems, but the optimum configuration will depend on local economics, policy, fuel availability and project scale.

Behind-the-meter power will not work everywhere. It introduces additional capital requirements, operating responsibilities, permitting challenges and community concerns. It can also complicate the environmental narrative surrounding AI infrastructure.

Yet until utilities can deliver capacity at the speed demanded by AI customers, onsite power will remain a meaningful part of the development equation.

Cox believes many developers would still prefer to fund grid improvements that accelerate access to utility power rather than permanently operate their own generation.

That may be the most important point in the behind-the-meter debate. The trend is not necessarily evidence that data center companies want to abandon the grid. It is evidence that the schedules of AI infrastructure and the schedules of the power system have diverged.

Execution Becomes the Market

The AI infrastructure boom is often described through demand forecasts, capital expenditures and pipeline announcements. DC Byte’s data suggests the more revealing story lies in execution:

• Can power be delivered on the required schedule?
• Can the facility support the required density at the rack?
• Can cooling be integrated as a core system rather than a retrofit?
• Can developers secure local support before opposition hardens?
• Can enormous capacity commitments survive the distance between early-stage ambition and an energized campus?

The industry has largely moved beyond debating whether AI will reshape the data center market. It is now discovering what that transformation requires in physical, financial and political terms.

Demand may determine the size of the opportunity. But power, density and community alignment will determine where—and whether—that opportunity gets built.

At Data Center Frontier, we talk the industry talk and walk the industry walk. In that spirit, DCF Staff members may occasionally use AI tools to assist with content. Elements of this article were created with help from OpenAI's GPT5.

Keep pace with the fast-moving world of data centers and cloud computing by connecting with Data Center Frontier on LinkedIn, following us on X/Twitter and Facebook, as well as on BlueSky, and signing up for our weekly newsletters using the form below.