Power-constrained data architecture curbing AI ambitions

Time-to-power reshapes architecture and procurement

The timeline from site selection to an energized data center facility has pushed well beyond typical project assumptions, and the delays are coming from multiple directions at once. An interconnection agreement does not necessarily mean power will arrive soon, even if data platforms are ready to deploy.

Dell'Oro Group's February 2026 forecast projects the worldwide data center physical infrastructure market to grow by a mid-teens percentage each year through 2030, surpassing $80 billion by 2030. However, the market research firm said power scarcity is driving increased demand for on-site generation as data center operators work around grid constraints. On-site approaches are shifting from a contingency to a practical necessity for large facilities, it noted.

Interconnection delays for power plants are also affecting data center timelines. A December 2025 report from Lawrence Berkeley National Laboratory found that power generation and storage capacity projects built in the U.S. from 2018 to 2024 took a median of more than four years to go from the interconnection request to commercial operation. In Northern Virginia, the country's largest data center market, transmission expansion is now a multiyear planning issue, extending to eight years or more. Timelines are comparable in parts of Europe, including the U.K. and Germany.

The pressure is not limited to hyperscalers and large colocation providers that are securing hundreds of megawatts for AI deployments. Alex Cordovil, research director at Dell'Oro Group, said enterprise teams are already seeing downstream effects.

"When hyperscalers and colocators absorb the lion's share of available utility capacity in a given market, enterprise teams face a knock-on effect, whether that's longer lead times for their own on-premises expansions, reduced availability at their preferred colocation facilities, or less favorable contract terms," Cordovil said.

Energization dates -- when a facility can draw grid power -- now carry as much weight in site selection as network latency or real estate cost. Bloom Energy's 2025 data center power survey found that 84% of respondents ranked power access among their top three site-selection considerations. When timelines slip, organizations might need to move AI workloads to other regions, delay deployments or accept limited power access with potential interruptions during peak demand.

The transition to on-site power and grid-interactive operations

In addition to on-site power infrastructure they control directly, data center operators are turning to grid-interactive operations that adjust electricity use based on overall demand, reflecting a shift toward power-aware data platform design.

Gartner expects energy alternatives, such as green hydrogen and small modular nuclear reactors, to become more viable for data center microgrids by the end of the decade. Dell'Oro Group's quarterly tracking shows that worldwide data center physical infrastructure spending increased 20% year-over-year in the fourth quarter of 2025. Power distribution was one of the stronger growth areas, driven by demand for higher-density AI infrastructure.

The shift is playing out across five technology categories:

• Strategic siting. Colocating capacity near existing grid-connected infrastructure can help shorten deployment timelines.
• Microgrids. Combining on-site generation with battery storage, microgrids enable facilities to operate independently of the grid during stress events to reduce the exposure of critical data workloads to instability.
• Fuel cells. Solid-oxide fuel cells offer a faster path to primary power than traditional grid connections, deploying in weeks rather than years. They also run at significantly higher electrical efficiency than diesel generators, but buyers should assess cost, emissions reporting and runtime assumptions.
• Battery energy storage systems. BESS can come online during outages, reducing dependence on diesel generators and supporting grid independence when a utility's power supply is interrupted.
• Workload flexibility. AI and data workloads are increasingly designed to shed or shift load during peak grid stress without taking systems offline. Training workloads that tolerate latency are especially suited to this model.

Provider questions for power-constrained AI workloads

Understanding what data center operators are deploying is only part of the equation. Data leaders and other enterprise buyers also need to know whether the cloud and data platform providers they are evaluating are keeping pace.

Cordovil said most buyers are still evaluating providers on the wrong metrics. Power usage effectiveness, redundancy tier and price per kilowatt are baseline measures. The questions that matter now, he added, are about power pathway resilience, future-proofing and flexibility:

• What is the provider's contracted utility capacity versus what is currently energized?
• How much headroom exists for density growth, not just at the facility level but at the rack and row level?
• What is the provider's strategy for on-site generation, grid interactivity and demand response participation?
• How flexible is the infrastructure to accommodate shifting power and cooling profiles as AI architectures evolve? Can it scale incrementally, or does it require a full replacement?
• What does the power roadmap look like over the next five years, not just what is available today?

Tony Harvey, an analyst at Gartner, said teams should also ask whether promised power delivery is firm or can be reduced during grid stress events. If it's the latter, they should ask what conditions would trigger a reduction, how often it could happen and whether critical workloads can be protected from curtailment. If on-site generation is the backup plan, buyers should confirm whether it can carry the full power load and for how long.

Planning ahead for power grid risk

Even with better provider selection, power risk remains a resilience issue for enterprise data teams. Power remains the leading cause of impactful data center incidents, according to Uptime Institute's annual outage analysis report for 2025. On-site generation and workload flexibility reduce exposure but do not eliminate risk.

A complete resilience posture requires action across three areas:

• Physical separation across regions is the foundation. Logical separation within a single facility provides no protection against a shared power failure. The baseline is cross-region replication of applications, data and logs across sites on separate feeds from separate providers, with failover tested under load rather than on paper.
• Generator and UPS testing must reflect actual failure scenarios. A quarterly exercise that never reaches full load does not validate performance. Best practices include full load-bank testing quarterly; an annual black-start test where generators start cold and an uninterruptible power supply (UPS) bridges the gap; and verification that security systems and access controls hold during switchovers. Uptime Institute found failure to follow established procedures is the leading addressable cause of outages, making process review more impactful than new hardware.
• Workload classification is a prerequisite for everything else. Teams that have mapped AI workloads by delay tolerance can negotiate demand-response agreements with utilities. That classification also provides a plan for shifting workloads across regions and power providers when a primary site faces constraints.

Terms should reflect power availability risk

AI workloads are increasingly moving closer to enterprise infrastructure for model fine-tuning, inference at the edge and retrieval-augmented generation, and the data architectures underpinning them are evolving quickly. The power and cooling envelope an organization needs to support platforms and applications today might look very different in 18 months.

Power risk spans site selection, procurement and operations, but the contract with a provider defines an organization's exposure. Harvey said many buyers don't realize the terms are negotiable.

He recommended that they:

• Require proof of a signed interconnection agreement, completed required studies and a confirmed energization date.
• Define delivery requirements at the rack power distribution unit level, including redundant feeds and completed full-load testing.
• Make the colocation provider responsible for utility delays where possible.
• Review the force majeure clause and explicitly exclude utility delays from its scope.
• Attach penalties to non-delivery tied to documented business impact.
• Require zero curtailment for critical loads backed by a service-level agreement with financial penalties.

Getting the terms right is necessary, but it is only the starting point for data leaders. The broader risk of treating power as an afterthought is strategic, not just operational.

"Buyers who treat power as a static line item on a colocation contract are going to find themselves either capacity-constrained or locked into infrastructure that can't adapt as AI workloads scale and shift," Cordovil said.

Sean Michael Kerner is an IT consultant, technology enthusiast and tinkerer. He has pulled Token Ring, configured NetWare and been known to compile his own Linux kernel. He consults with industry and media organizations on technology issues.