Power and Fiber Now Lead Data Center Siting Decisions
A primer on utility interconnection, power procurement, fiber engineering, and the infrastructure constraints reshaping AI data center development.
Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.
The sequence of data center development has inverted.
For most of the industry’s history, a developer secured land first, completed the supporting work, obtained entitlements, signed a lease, and arranged power near the end.
That order no longer holds for projects at institutional scale.
Power access and fiber connectivity now sit at the front of the site selection process, and the power purchase agreement is often the document that decides whether a project proceeds at all.
This primer explains how that process works.
It covers the utility interconnection sequence, the main power sourcing strategies, the fiber engineering that governs performance, and the way these constraints shape where AI infrastructure is being built.
Why the order changed
The cause is a timing mismatch. A data center shell can be built in 18 to 24 months.
The high-voltage transmission lines and substations that feed it take seven to ten years to permit and construct.
United States high-voltage transmission construction fell from roughly 1,700 miles per year in the early 2010s to about 180 miles per year by 2022.
The compute deployment cycle and the grid buildout cycle now run on entirely different clocks.
The AI load profile widened the gap.
A training cluster draws near-continuous full-load power.
That flat, predictable consumption curve looks nothing like the variable residential and commercial demand utilities were designed to serve.
A single campus requesting 100 MW or more triggers a multi-stage study sequence before any new infrastructure is built.
Goldman Sachs Research estimates roughly $720 billion in grid spending may be needed through 2030 to support data center expansion.
The utility interconnection sequence
Connecting a large load, generally 20 MW or more, moves through four stages.
The first is the feasibility study.
The developer submits an interconnection request, and the utility conducts a preliminary, non-binding assessment of the upgrades required for the connection.
The second is the system impact study, the substantive engineering phase.
The utility models load flows, short-circuit behavior, and voltage and dynamic stability to quantify the new load’s effect on the transmission system.
For large campuses, this stage frequently surfaces a need for transmission infrastructure the developer did not anticipate.
The third is the facilities study, which scopes, designs, and prices the required upgrades: new substations, transformer additions, line extensions, or reactive power compensation.
The developer is generally responsible for the costs it directly causes.
The fourth is the interconnection agreement and construction. This is where the longest delays occur.
PJM data shows projects spending more than three years to reach an agreement and roughly four more waiting to come online, with total time from application to operation averaging over seven years for AI projects entering service in 2025.
The backlog is immense. Estimates place 2,060–2,600 GW of capacity in U.S. interconnection queues roughly equal to or exceeding the installed capacity of the entire grid.
Conversion rates remain low: only about 40% of ERCOT projects and 24% of comparable PJM projects have reached agreement or become operational.
FERC Order 2023 replaced the old serial queue with an annual cluster study process, grouping requests that arrive within a 45-day window, and raised deposits and withdrawal penalties to clear speculative projects.
A subsequent FERC proceeding aims to build the first dedicated large-load framework for AI data centers, including an expedited track for loads that agree to curtail under grid stress.
Power sourcing strategies
Four approaches now define how projects secure electricity.
Developers now rely on four primary power procurement strategies, each with distinct tradeoffs.
Direct utility interconnection remains the most common, but agreements increasingly include curtailability, network upgrade costs, and minimum-billing commitments.
Northern Virginia illustrates the challenge: transmission constraints not generation have driven connection delays, with Dominion Energy reporting pending data center requests far exceeding its historical peak load.
Power purchase agreements (PPAs) have shifted from an operational decision to a financing priority.
Physical PPAs secure direct supply, while virtual PPAs hedge price risk but introduce basis risk.
Nuclear PPAs have expanded rapidly, highlighted by Microsoft’s Three Mile Island agreement and Amazon’s restructured long-term Susquehanna deal.
Behind-the-meter generation bypasses grid queues through on-site power, typically gas turbines or fuel cells.
While it reduces transmission risk, it creates fuel supply and pipeline dependency, as seen in xAI’s 422 MW off-grid deployment in Memphis.
Hybrid configurations combining grid power, dedicated generation, and storage are emerging as the preferred model for large AI campuses, with gas providing near-term baseload and nuclear paired with storage representing the longer-term direction.
Fiber route engineering
Power governs whether a facility can operate. Fiber governs how well it performs.
Inside the data center, AI workloads require an ultra-low-latency fabric connecting GPUs during training, with bandwidth demands rising each generation from roughly 800 Gbps for the H200 to over 1,200 Gbps for the B200.
As a result, building design, including ceiling heights and conduit routes, is increasingly driven by network requirements.
Outside the facility, resilience depends on true carrier diversity: independent providers using physically separate routes, entry points, and equipment.
Shared conduits do not provide redundancy, making dark fiber ownership and long-term route agreements strategic assets.
Latency also shapes deployment.
Training clusters can locate where power is cheapest, while inference workloads require sub-10-millisecond proximity to users, anchoring them to metro markets.
This rural training–urban inference split is becoming the defining architecture of AI infrastructure.
How the constraints shape geography
The same forces are reshaping every major market differently.
Northern Virginia remains the largest hub, supported by decades of internet exchange and subsea cable infrastructure despite power constraints.
Dallas-Fort Worth continues to attract investors, but Oncor’s 255 GW interconnection queue against a 31 GW system peak signals mounting congestion.
Johor has emerged as Southeast Asia’s fastest-growing hub, driven by low-latency connectivity to Singapore and strong subsea cable access.
Abu Dhabi has accelerated development through a sovereign speed-to-power model that aligns utility control, grid investment, and major cable landings.
The underlying pattern is the same: generation is seldom the constraint power delivery is.
Markets that secure power and fiber connectivity early command higher valuations and faster lease-up.
Developers that lock in grid access and fiber routes capture the greatest upside, while those that delay face rising costs and stronger competition in a market where tenants have global options.