Data center backlash could put CIO AI plans on hold

US data center restrictions are breaking old AI cost assumptions

Protests against new data centers were organized across 42 US states in mid-July. At that point, 10 states, including Florida, Georgia, and Virginia, had active construction moratoriums, while eight others had pending legislation, according to datacenterbans.com. The issue has moved beyond local opposition. It now affects the economics of enterprise AI.

Power is the immediate constraint. As of May, 23 states had approved large-load tariffs, according to Arif Gasilov, partner in the natural resources and built environment division at sustainability advisory firm Gasilov Group. These tariffs can require data centers to pay the full infrastructure cost associated with serving their facilities. Those costs can ultimately affect the price enterprises pay for cloud and colocation capacity.

This makes AI business cases based on older electricity assumptions less reliable. “What this means for CIOs is that power cost assumptions built in 2023 are wrong in close to half the country,” Gasilov said. He advises CIOs planning AI deployments in affected states to ask cloud and colocation providers how new tariffs change their rate structures, then recalculate project economics.

CIOs should make that reassessment at the workload level. A higher infrastructure bill changes the economics of training, inference, data storage and accelerator use differently. It can also affect decisions about which models to run, where to run them and whether dedicated capacity still makes financial sense.

Building smaller data centers is not a universal workaround. Gasilov notes that some state rules can target groups of facilities located close together rather than only the size of an individual site. Splitting one large project into several smaller projects may therefore leave the underlying regulatory problem unchanged.

For executives, the key change is clear: power availability and local infrastructure policy now belong inside AI financial planning. A technically sound deployment can still fail its business case if the electricity and data center assumptions underneath it have changed.

Fewer data centers will restrict where and how companies deploy AI

The impact is not limited to price. Fewer approved data centers mean fewer places to obtain the dense computing capacity needed for large AI workloads. That reduces deployment flexibility and can increase the time required to secure capacity.

Chuck Girt, CTO at fiber-optic network provider FiberLight, expects this constraint to change AI deployment rather than stop AI adoption. “I don’t think the rate of data center construction changes the direction AI is headed, but it could influence how organizations deploy and access AI at scale,” he said.

The reason is structural. Most enterprises do not own the physical infrastructure behind their AI services. “Most enterprises aren’t going to build this infrastructure themselves; they’re going to rely on cloud and data center environments to provide the compute AI requires,” Girt said. Restrictions on new facilities therefore flow through cloud and colocation providers to their enterprise customers.

Kevin Surace, CEO of biometric security vendor TokenCore and an AI and green energy expert, identifies the resulting risks: less available capacity, fewer options for geographic redundancy, longer provisioning times and greater dependence on a limited number of cloud providers and locations. He describes compute capacity as strategically important alongside electricity, semiconductors and network connectivity.

Geographic concentration deserves particular attention from executives. AI workloads concentrated in a small number of regions create operational and procurement dependencies. Organizations may have fewer alternatives when a provider runs short of accelerator capacity, faces a local power constraint or cannot meet a required deployment schedule. Geographic redundancy can also matter for resilience and business continuity.

Capacity planning therefore needs to happen before an AI application is ready for production. Surace warns that companies that have not secured capacity may discover that their AI strategy works technically but cannot be executed on schedule. For large AI programs, access to physical compute is becoming a prerequisite rather than a procurement task that can be left until deployment.

The practical response is to reduce concentration risk early. CIOs should identify the regions and providers on which important AI workloads depend, determine how much capacity is actually committed, and test whether credible alternatives exist. The objective is not to build redundant infrastructure everywhere. It is to prevent a shortage in one provider or region from determining the pace of the company’s AI strategy.

Constrained data center capacity will raise AI infrastructure costs

AI compute demand is rising while new data center projects face tighter limits. That combination creates a clear pricing risk. If the supply of facilities, electricity and high-density computing capacity grows more slowly than demand, enterprises should expect to pay more for AI infrastructure.

Kevin Surace, CEO of biometric security vendor TokenCore and an AI and green energy expert, expects the pressure to affect several parts of the market. “Demand for AI compute is accelerating, so constraining the supply of facilities, electricity and high-density capacity will place upward pressure on cloud pricing, colocation, accelerator access, and long-term capacity contracts,” he said.

Accelerator availability is especially important. GPUs and other AI accelerators require significant power and cooling, so buying more chips does not solve the problem if suitable data center capacity is unavailable. For CIOs, the constraint is therefore the complete system: compute hardware, electrical capacity, cooling and physical space must all be available in the required location.

Large enterprises have more options to manage this risk. Surace says organizations with access to capacity can protect themselves through multiyear agreements and dedicated infrastructure. Locking in capacity can improve planning certainty, although executives must balance that benefit against the risk of committing too much capital to infrastructure as AI models and hardware become more efficient.

The impact could be more severe for smaller buyers. Surace expects smaller organizations, startups and universities to face the largest percentage increases. Some could be priced out of leading-edge AI capabilities. Large cloud customers generally have greater purchasing power and more options for long-term commitments, while smaller users have less ability to absorb sharp changes in infrastructure prices.

This makes AI cost control a capacity-planning issue. CFOs and CIOs should model how higher cloud rates, accelerator costs and capacity commitments affect the full economics of AI projects. They should also separate workloads that genuinely require expensive, leading-edge infrastructure from those that can run on smaller models or less costly hardware.

The objective should be predictable access at an acceptable cost. Enterprises do not need maximum compute for every AI use case. They need enough secured capacity for workloads that create measurable business value, with flexibility to shift less demanding work to cheaper infrastructure.

Compute and energy now require strategic supply planning

CIOs should treat access to compute and electricity as strategic dependencies for major AI programs. The central issue is availability. An organization cannot deploy a workload on schedule if its provider lacks sufficient power, accelerator capacity or approved data center space, regardless of how well the application itself performs.

Surace recommends securing capacity as early as practical and avoiding dependence on one cloud provider or geographic region. He also advises companies to use smaller, more efficient AI models where they can meet the required business outcome. These measures address both cost and availability without requiring enterprises to build their own large-scale infrastructure.

Diversification needs discipline. Adding multiple providers does not automatically reduce risk if they depend on the same region, electricity infrastructure or underlying data centers. CIOs should understand where critical workloads physically run, which regions provide alternatives, and how quickly workloads can move between environments. Contracted capacity should also be distinguished from capacity that a provider merely expects to make available later.

Environmental and community issues belong in this assessment because they can directly affect project approval and continuity. Surace recommends asking data center providers where their water comes from, whether cooling systems use a closed loop, who pays for new electrical grid infrastructure, what share of power is generated onsite, and what environmental monitoring results are publicly reported.

These questions address risks that extend beyond sustainability reporting. Water availability can restrict cooling options. Grid upgrades can add costs and extend construction schedules. Local opposition can delay or cancel projects. Surace argues that “Transparency and early community engagement are far less expensive than lawsuits, project cancellations, and moratoriums.”

CIOs should therefore connect AI infrastructure planning with finance, procurement, sustainability and business continuity. Capacity requirements need to be forecast alongside application demand, while contracts should address price, availability, location and expansion rights. Smaller or more efficient models should be considered when they can reduce infrastructure requirements without compromising the intended result.

The executive priority is to secure the resources required for important AI workloads before scarcity dictates the company’s options. AI strategy increasingly depends on models and software, and on confirmed access to power and compute at a viable price and in the right locations.

Data center constraints will push AI strategies toward efficiency

The strongest response to limited data center capacity is not simply to secure more infrastructure. Enterprises can also reduce how much infrastructure each AI workload requires. That means getting more useful output from each GPU, watt of power and dollar of spending.

Anurag Gurtu, cofounder and CEO of agentic AI platform provider Airrived, argues that this is where competitive advantage will shift. “Enterprises don’t actually want more data centers; they want more intelligence per watt, per GPU, and per dollar,” he said. “The winners won’t be those with the biggest infrastructure footprint, but those extracting the most value from every unit of compute.”

This changes an important assumption behind enterprise AI planning. Strategies that depend on almost unlimited computing capacity will be more exposed to higher electricity prices, data center restrictions and accelerator shortages. Adding hardware can support growth, but it becomes less attractive when the marginal business value grows more slowly than infrastructure spending.

Gurtu expects the next phase of AI development to operate under tighter compute, power and economic limits. “Organizations that optimize models, deploy domain-specific AI, and leverage hybrid architectures will continue to innovate, while those relying solely on scaling hardware will face diminishing returns,” he said.

Model choice is therefore becoming a financial and infrastructure decision. A large general-purpose model may be justified for complex tasks that require its capabilities. It is not automatically the best choice for narrow and repeatable business processes. Smaller or domain-specific models can require less memory and compute, reduce inference costs and improve response times when they deliver sufficient accuracy for the task.

Hybrid architectures provide another option. Enterprises can assign different workloads to different models and computing environments rather than forcing every application onto the most expensive infrastructure. Sensitive or predictable workloads may run on dedicated resources, while other tasks use cloud services. The right design depends on performance, security, data governance, availability and cost requirements.

Higher infrastructure prices may also create stronger incentives to improve the software layer. Gurtu points to model optimization, inference efficiency and intelligent orchestration as areas likely to benefit. “Rising infrastructure costs also accelerate innovation in model optimization, inference efficiency, and intelligent orchestration,” he said. These improvements can reduce the amount of compute needed to deliver the same business outcome.

For C-suite leaders, the relevant measure is no longer how much AI capacity the company can acquire. It is how effectively that capacity turns into business results. CIOs should track compute cost by workload, utilization of expensive accelerators, inference costs and the business value produced by each AI service. Those measures make it easier to identify applications that deserve additional capacity and those that need redesign.

Data center constraints do not make enterprise AI unworkable. They make inefficient AI more expensive. Companies that improve model selection, infrastructure utilization and workload placement can continue to expand AI while reducing their exposure to scarce power and compute.

Key takeaways for decision-makers

• Reprice AI infrastructure plans: Data center opposition, construction moratoriums and new power tariffs are changing AI economics across the US. CIOs should update older power and capacity assumptions before approving major deployments.
• Secure capacity before it becomes a constraint: Fewer data center options can mean longer provisioning times, less geographic redundancy and greater cloud concentration. Leaders should confirm committed capacity across multiple providers and regions before critical AI workloads reach production.
• Prepare for higher compute costs: Limited data center, power and accelerator capacity could increase cloud, colocation and long-term contract prices. Prioritize scarce compute for high-value workloads and assess multiyear capacity agreements where demand is predictable.
• Manage compute and energy as strategic supply risks: AI programs increasingly depend on confirmed access to power, cooling and physical capacity. CIOs should diversify infrastructure sources and assess providers for grid, water, environmental and expansion risks.
• Make AI efficiency a competitive priority: Infrastructure constraints increase the value of smaller models, domain-specific AI, efficient inference and hybrid architectures. Measure business value per unit of compute rather than relying on continual hardware expansion.