Key takeaways

  • Only about 1 % of data‑center space is vacant, but most of that cannot host 140‑160 kW AI racks.
  • Power density for AI has jumped from under 10 kW per rack a few years ago to 40‑130 kW today, with some GPU platforms exceeding 200 kW.
  • Building a new AI‑focused campus now costs roughly $488 /ft², while AI‑optimized builds push past $1,100 /ft².
  • Labor shortages and transformer lead‑times of up to six years are the top cost and schedule drivers.
  • Liquid‑to‑chip cooling delivers PUE around 1.10‑1.20, compared with 1.55‑1.67 for legacy air‑cooled plants, effectively freeing megawatts for compute.
  • Power availability, not land, is the primary siting factor; megawatt‑scale campuses are being sited at existing industrial sites with grid capacity already in place.

1. Space is not just scarce – it’s unusable for AI

Even though overall vacancy sits near 1 %, the subset of that space capable of supporting the 140‑160 kW per‑rack loads needed for modern generative‑AI models is dramatically smaller. Operators must also pass credit‑screening and liquid‑cooling readiness checks before a rack can be considered usable. The result is a de‑facto shortage of AI‑ready real estate, despite the appearance of open capacity.

2. Power density has become the new design driver

AI workloads are reshaping the rack‑level power envelope. A few years back a typical rack consumed 5‑10 kW; today the baseline for AI‑focused racks sits between 40 and 130 kW, and the most aggressive GPU platforms push beyond 200 kW per rack. This shift forces a redesign of electrical distribution, UPS sizing, and cooling capacity.

3. Liquid cooling is moving from niche to default

Traditional air‑cooling can no longer keep pace with the heat generated by dense AI racks. In many installations, the heat load already outstrips what airflow can move, making direct‑to‑chip liquid cooling a practical requirement. Industry observers note a growing consensus that liquid cooling will dominate high‑performance and hyperscale facilities.

4. Construction economics are exploding

The average cost to erect a data‑center footprint in 2026 is $488 per square foot. Facilities optimized for AI‑scale power and cooling routinely exceed $1,100 per square foot. The cost surge is driven primarily by the need for robust power infrastructure, high‑density cooling loops, and extensive site‑specific engineering.

5. Labor and component lead‑times dominate budgets

Skilled‑labor shortages now top the list of cost drivers, with the largest campus builds requiring 4,000‑5,000 workers at peak. In addition, critical electrical components such as high‑voltage transformers can have lead‑times of up to 72 months, extending overall project schedules.

6. Grid access is the biggest siting constraint

Power availability has overtaken traditional site‑selection criteria like land price or proximity to fiber. In many U.S. markets, securing new megawatt capacity can take three to four years—longer than the physical construction of a data‑center. As a result, developers are gravitating toward sites with existing grid interconnections, even repurposing former industrial facilities.

7. Mega‑campus architecture: power, cooling, and modularity

Designing a multi‑hundred‑megawatt AI campus now follows a layered approach:

  • Utility‑first planning – lock in megawatt‑scale grid interconnection before breaking ground.
  • Integrated power‑to‑chip architecture – combine UPS, lithium‑ion battery buffers, and direct‑to‑chip liquid cooling to collapse the power‑to‑cooling path.
  • Modular construction – prefabricated power skids, rack‑level coolant distribution units, and containerized cooling plants enable phased commissioning.
  • Digital twins – simulate power distribution, coolant flow, and network topology to validate design choices and predict bottlenecks before hardware arrives.

A leading example is the Lake Mariner campus, which re‑uses an existing industrial grid connection and is planned to scale to 750 MW using a unified power‑and‑liquid‑cooling infrastructure.

8. Cooling efficiency translates directly to compute capacity

Legacy air‑cooled plants typically run with Power Usage Effectiveness (PUE) between 1.55 and 1.67, meaning roughly a third of incoming electricity powers the cooling plant rather than the servers. Direct‑to‑chip liquid cooling can lower PUE to the 1.10‑1.20 range and cut cooling‑energy consumption by 30‑60 %. Those efficiency gains effectively free megawatts for additional AI racks without waiting for new grid capacity.

9. How density, cooling, and efficiency intersect (comparison)

Rack power density Typical cooling approach Observed PUE range
≤15 kW (legacy) Air‑side CRAC units 1.55 – 1.67
30‑100 kW (current AI) Direct‑to‑chip liquid loops / rear‑door heat exchangers 1.10 – 1.20
>100 kW (next‑gen GPU platforms) Immersion or high‑flow liquid systems 1.10 – 1.20

The table illustrates that as rack density climbs, operators must adopt liquid‑based cooling to maintain reasonable PUE and preserve electrical headroom.

10. Speed versus scale: modular vs traditional builds

Conventional builds still require 18‑36 months for civil work, power installation, and commissioning. AI demand cycles, however, can shrink to a few months. To bridge that gap, many providers are standardizing on modular data‑center pods that arrive pre‑wired for power and cooling, allowing rapid stacking of compute capacity while the surrounding infrastructure is phased in.

Conclusion

The AI compute boom has turned high‑density space into a premium commodity. Operators must treat power, cooling, and construction as a single, interdependent system. Securing megawatt‑scale grid access, investing in direct‑to‑chip liquid cooling, and embracing modular, digitally‑validated designs are critical ways to keep pace with the accelerating demand for AI racks. The companies that can align these variables will be the ones that successfully launch the next generation of AI factories.

Sources

This article was researched and fact-checked against the following sources: