AI data center space bottleneck exposes power and cooling gaps

Solmar Insights

A North American search for 5 MW of rack-scale, liquid-cooled GPU colocation, commissioning in early 2027, revealed a severe gap between vacant data center space and infrastructure actually capable of supporting high-density AI hardware. More than ten operators confirmed available square footage, but almost none could guarantee the technical requirements or bring supporting systems online within the needed timeline.

Vacancy versus usable power

According to CBRE, North America’s primary data center markets posted a record-low vacancy rate of 1.4 percent in the first half of 2026. However, for AI and GPU deployments, technical infrastructure, not just floor space, is the main gating factor. The reported 5 MW search targeted liquid-cooled, rack-scale GPU systems at densities up to 160 kW per rack, such as NVIDIA’s GB300 NVL72 cluster.

The overwhelming majority of operators queried had unoccupied space, but critical infrastructure for ultra-high-density racks was usually absent or insufficient. Most facilities are still designed for average rack densities between 10 and 30 kW, with only a fraction reporting any capability above 30 kW and racks over 100 kW per position remain rare nationwide.

Three technical barriers

Operators identified three non-negotiable technical constraints: power delivery per rack, cooling infrastructure at the rack row, and floor loading. Data halls originally designed for 15 to 30 kW racks cannot rapidly or inexpensively deliver 10x that density, even if total site MW is available. Retrofitting electrical distribution and busways to each rack is a complex, time-consuming process.

Cooling represents a discrete hurdle; “liquid-cooling ready” commonly means that the building is engineered to eventually support liquid cooling, but does not actually have facility water loop hookups extended to the white space. Facility cooling lead times, equipment, installation, commissioning, were quoted at up to 12 months after building completion, not concurrent. In addition, a fully loaded AI rack can weigh around 3,000 pounds, and many raised floors cannot accommodate that live load.

Liquid cooling timelines and limits

A recurring theme in provider responses was the distinction between being liquid-cooling ready and actually delivering liquid cooling to the rack row. While the server hardware often ships with in-rack coolant distribution units, the larger facility-side loop and associated infrastructure are the operator’s responsibility and may lag building delivery by up to a year. For buyers, the meaningful readiness date is when chilled water or equivalent is plumbed and active at the rack position, not at central plant level, and not at building handover.

Operators consistently cited bottlenecks not just in equipment procurement but also in permitting and labor to extend new liquid cooling loops into operational halls. Buyers seeking rapid deployment, especially in AI, are thus left with very limited viable options. Selection must go beyond traditional metrics of gross power and area and focus on rack-level, room-by-room certification.

Credit as a hidden filter

Interestingly, creditworthiness emerged as an unspoken but material constraint. Two operators with serious interest in the 5 MW opportunity independently raised questions about whether the tenant was investment grade, with one operator making future capacity negotiations conditional on positive credit assessment. In the current environment, megawatt-scale AI deals represent large, multi-year receivables, and with very tight supply, landlords can be selective about tenant risk.

This dynamic means some providers may publicly cite lack of capacity as the reason for declining bids, when in fact buyer credit or lack of deposit or guarantee is a factor. AI tenants without investment-grade credit should anticipate the need to provide additional surety or creative structuring to secure AI-ready capacity.

Practical questions for AI deployments

The findings suggest AI compute buyers need to change how they assess and shortlist data center candidates. Rather than relying on headline MW and rack numbers, the critical questions focus on sustained density in the specific offered space, concrete timelines to have cooling live at the rack row, and the true floor load allowance in the target halls. In addition, evaluating the expansion path, and whether it is fully approved or just a roadmap concept, prevents mid-cycle surprises. Finally, getting credit requirements on the table from RFP stage can save weeks of lost time and bid churn.

Flexibility also proved decisive; buyers able to phase their deployments, split loads across sites, or use air-cooled platforms at initial stages will increase their options in the current constrained market. These practical shifts, rather than waiting for new mega-campus stock, align with the real bottlenecks in North American AI data center capacity in 2026.

Reporting via the original publisher