Why power availability, not land cost, has become the binding constraint on data center siting decisions — especially for AI training and inference workloads.
A decade ago, data center siting was mostly a land-and-connectivity exercise. Today, for any facility built to support AI training or high-density inference, the first question is no longer where the land is cheapest — it's where enough power, at the right redundancy tier, can actually be interconnected within a financeable timeline.
This report sets out the siting criteria INA applies to datacenter projects, why power availability now dominates every other variable, and the redundancy, cooling and regulatory considerations that determine whether a candidate site is actually financeable.
“Land is a one-time decision. The power interconnection queue is the decision that actually controls your delivery date.”INA Project Structuring Framework™ — Field Notes, 2026
AI training clusters can require rack densities several multiples higher than a traditional enterprise data hall — which changes the siting calculus for power, cooling and even structural load in ways that a land-cost comparison alone will never capture.
A site with cheaper land but a multi-year power interconnection queue is not a cheaper site — it's a delayed site, and delay is what actually erodes a data center business case.
The recurring siting mistakes INA sees across public and private datacenter projects.
Utility interconnection queues can run years long. Sponsors that check queue position after site selection frequently discover the timeline no longer fits the financing plan.
High-density AI racks push cooling demand well beyond traditional air-cooled designs, and water-stressed regions can turn a promising site into a permitting fight.
Tier III versus Tier IV changes both capex and the site's power and cooling footprint substantially — it needs to be a Phase II input, not a Phase IV surprise.
Inference workloads are far more latency-sensitive than training workloads — a site optimized purely for power can still be the wrong site for a latency-sensitive service.
Water use, noise and grid-load concerns increasingly trigger community opposition — engaging regulators and communities early avoids late-stage permitting reversals.
INA evaluates candidate sites against six criteria, weighted according to workload type, as part of the Project Structuring Framework™ (F1) Phase II business case for datacenter projects.
| Criterion | What It Measures | Why It Matters for AI Workloads |
|---|---|---|
| Power Availability & Cost | Interconnection queue position, substation capacity, tariff structure | Often the single largest operating cost and the primary delivery-timeline constraint |
| Redundancy Architecture | Required uptime tier (N+1, 2N) versus site's grid and backup capacity | Training clusters can tolerate more scheduled downtime than production inference services |
| Cooling & Water Access | Local water stress, availability of liquid cooling infrastructure, climate | High-density AI racks frequently require liquid cooling, not traditional air cooling |
| Network Latency | Round-trip time to the primary demand centers the facility will serve | Critical for inference; largely irrelevant for batch training workloads |
| Land & Climate Risk | Flood zones, seismic exposure, long-term climate projections | A 15–20 year asset life means today's climate risk models matter more than today's weather |
| Regulatory & Tax Environment | Permitting timeline, incentive programs, data residency requirements | Data residency rules increasingly shape where AI infrastructure can legally be sited |
Redundancy tier is a Phase II decision, not a construction detail — it drives power, cooling and capex simultaneously.
| Tier | Description | Target Uptime | Typical Use Case |
|---|---|---|---|
| Tier I | Single path for power and cooling, no redundancy | ~99.67% | Non-critical batch training, development environments |
| Tier II | Redundant capacity components, single distribution path | ~99.75% | Small colocation, edge sites |
| Tier III | Concurrently maintainable, multiple distribution paths, one active | ~99.98% | Most hyperscale and AI training facilities |
| Tier IV | Fault tolerant, multiple active distribution paths | ~99.995% | Latency-sensitive production inference, mission-critical services |
| Risk | Probability | Impact | Primary Mitigation |
|---|---|---|---|
| Power interconnection queue delay | High | High | Queue position confirmed with the utility before site selection is finalized |
| Water stress limiting cooling capacity | Medium | High | Water availability studies and liquid-cooling alternatives assessed at siting stage |
| Climate exposure over asset lifetime | Medium | High | Long-horizon climate risk modeling included in site due diligence, not just current flood maps |
| Community or regulatory opposition | Medium | Medium | Community and regulator engagement started during site evaluation, not after permitting begins |
| Grid interconnection cost overrun | Medium | Medium | Utility interconnection cost estimate independently validated before financial close |
Siting is where a datacenter project's real economics get decided, long before ground is broken. Power queue position, redundancy tier and cooling strategy need to be evaluated together, at the same stage as the financial model — not sequenced after a site is already chosen for its land price.
Published by International Network Advisors (INA), September 2026. Part of the INA Knowledge library, drawing on the INA Project Structuring Framework™ (F1).
INA's advisory team can run a site feasibility review against your candidate locations before you commit capital. Request Advisory →