For High-Density AI, Available Data Center Space May Not Be Usable

North American data center vacancy remains at 1%, but JLL reports that only a very small share of that capacity can support high-density AI deployments, with most vacancies in legacy products. A 5 MW buyer seeking 140-160 kW per rack found nearly every operator had space, but almost none could support the required density with direct liquid cooling, according to James Mercer of Metro Colo Advisory. Uptime Institute's 2025 survey found 82% of respondents had a highest-density rack below 30 kW, and only 9% reported 50 kW or more, with fewer than 4% of US data centers able to support 100 kW+ racks, per research director Daniel Bizo.

For High-Density AI, Available Data Center Space May Not Be Usable North American vacancy remains at 1%, but JLL says only a very small share of that capacity can support high-density AI deployments. A 5 MW data center buyer recently approached more than 10 North American operators seeking capacity for 140-160 kilowatts per compute rack. Nearly every operator had space. Almost none could support the required density with direct liquid cooling /cooling/cooling-becomes-the-system-constraint-as-ai-reshapes-data-center-design , according to James Mercer, principal of Metro Colo Advisory, who conducted the search for a client. Mercer agreed to be quoted on the record but declined to identify the operators, citing ongoing channel relationships with them. The search highlights a distinction between available data center capacity and capacity capable of supporting the highest-density AI systems. JLL’s local brokerage teams track available capacity by rack density, according to Andrew Batson, who leads the firm’s data center research. Only a “very small percent” of the 1% vacancy can support high-density deployments such as the requirement Mercer was shopping for, Batson told Data Center Knowledge. He said JLL views liquid-cooling infrastructure as a requirement for AI deployments at these densities. 1% Vacancy, Little AI-Ready Capacity JLL’s Midyear 2026 North America Data Center Report https://www.jll.com/en-us/insights/market-dynamics/north-america-data-centers puts North American vacancy at 1% for the third consecutive year. Available capacity is limited to small, fragmented blocks, while most tenants securing space are contracting for 2028 deliveries. More than 66 GW of capacity is under construction across North America, JLL said, with 77% of that capacity in frontier markets. For an AI buyer seeking 140-160 kW per rack, however, the market is much narrower than the headline vacancy rate suggests. “Most vacancies are in legacy products,” Batson said. Over a multiyear horizon, JLL expects more AI-enabled capacity to come online. The gap between nominal vacancy and AI-ready capacity is unchanged, he noted. Few Facilities Can Support 100 kW+ Racks Uptime Institute’s 2025 Global Data Center Survey https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf?version=0 found that 82% of respondents had a highest-density rack below 30 kW, while only 9% reported a highest-density rack of 50 kW or more. Uptime identified some cabinets exceeding 100 kW but said those remained rare. “Probably a low single digit under 4% of data centers in the US can currently support these racks,” said Daniel Bizo, research director at Uptime Intelligence, Uptime Institute’s market research unit. Bizo told Data Center Knowledge that the figure is an inference from Uptime’s survey data and should be treated as indicative due to sample size limitations. Uptime’s surveys https://intelligence.uptimeinstitute.com/index.php/resource/2025-global-data-center-survey-results-and-crosstabs show fewer than 1% of operators running 100 kW+ racks as a standard configuration, Bizo said. Fewer than one in 10 US operators may be able to support a limited number of 100 kW racks, such as a single row. “With AI computing for training still proliferating, 100 kW+ racks are becoming more common,” Bizo said. That puts Mercer’s 140-160 kW requirement at the extreme end of the market. The 30-kW threshold reflects both the statistical distribution of rack densities and changes in hardware and rack configurations needed to support denser deployments, Bizo said. Most data centers designed over roughly the past decade were built around an average rack power of about 20 kW, even if that capacity was not fully fitted on day one. Those facilities can often accommodate 30-40-kW racks in limited numbers with relatively straightforward upgrades, he said. The picture changes above 50 kW. “Above 50 kW, it becomes increasingly impractical to cool with air due to space limitations in the IT chassis/rack, mandating direct liquid cooling,” Bizo said. At scale, higher-density racks also create power-distribution challenges involving larger and heavier busways and PDUs, power cables, breakers, and potentially additional circuits, he said. Electrical room size and layout can become constraints. Some components have 6- to 12-month lead times due to high demand, while additional power-distribution equipment adds weight to already heavy racks. That can create structural issues in multistory buildings. “100 kW+ racks pose a bigger electrical challenge than thermal,” Bizo said. Bizo added that a 140-160 kW requirement does not mean every rack in a larger deployment will operate continuously at that level. A deployment built around racks at that peak density could average about 80 kW per 19-inch rack position, since network, storage, and other racks are part of the installation. Compute racks also may not operate at maximum power simultaneously, although training workloads can produce recurring peaks that electrical systems must be sized to accommodate. Uptime does not have a direct cross-tab that compares liquid-cooling capability with ultra-high-density rack support. Bizo said support for liquid cooling is broader than support for 100 kW+ racks, meaning liquid-cooling capability alone is not a reliable proxy for extreme rack-density capability. More than half of US operators surveyed by Uptime brought some high-density capacity online during the 12 months through April 2026, Bizo noted. That share rises to about two-thirds for the current 12-month period. However, Uptime did not define high density specifically as racks of 100 kW+, so the figures do not directly measure the expansion of ultra-high-density capacity. Megawatts Aren’t Enough A buyer seeking 5 MW at 140-160 kW per rack is not looking for generic vacant capacity. The deployment requires electrical infrastructure capable of delivering the required power at the rack, cooling systems capable of removing the resulting heat, direct liquid-cooling capability and suitable contiguous space. Mercer’s published analysis says rack-scale GPU systems at these densities use direct-to-chip liquid cooling /cooling/data-center-cooling-methods-costs-vs-efficiency-vs-sustainability , with cold plates cooling processors and a coolant loop reaching the white space through a coolant distribution unit. His analysis also distinguishes between a provider saying it supports liquid cooling and a specific data hall having a liquid-cooling loop serving the white space today. That distinction matters for a buyer trying to deploy immediately. Mercer’s search found operators with space that could not meet the required density and cooling configuration. His analysis identifies Nvidia’s GB300 NVL72 platform as requiring roughly 140-160 kW per rack. JLL’s 1% vacancy rate captures overall market availability, while Uptime’s data show how uncommon extreme rack densities remain. Neither measures the amount of vacant capacity capable of supporting a 140-160 kW rack. Credit Screens Further Narrow Options The physical constraint was not the only obstacle Mercer encountered. The two purpose-built, liquid-cooled operators he approached were the only ones in his search that engaged substantively rather than declining outright. Both asked, without prompting, whether his client was investment grade. Mercer declined to identify those operators, citing his ongoing relationships with them. One confirmed it had no capacity for the requested timeline, mentioned a future project that might fit, and then asked whether the client was investment grade. Further discussion was contingent on the answer, Mercer said. Mercer has seen that screening on this requirement only. He does not claim it is a broader industry practice. The observation follows an April 2026 report /cloud/neocloud-storm-gathers-as-data-center-deals-stall-over-credit-risk by Data Center Knowledge on stalled neocloud deals, which found that operators were scrutinizing customers for credit quality, long-term viability, and balance-sheet durability. The reporting found that operators were requiring stronger guarantees, including letters of credit or parent backing, and were declining some neocloud deals despite aggressive commercial terms. The new observation raises a narrower question: When high-density capacity is already scarce, does customer credit become another filter on facilities capable of serving those deployments? JLL’s financing analysis concerns project lending, not operators’ customer-selection policies. It does not establish that colocation providers broadly require investment-grade customers. But tenant credit can affect how developers structure AI and high-density projects, according to Carl Beardsley, senior managing director and data centers leader at JLL Capital Markets. Asked whether tenant credit affects how developers approach AI and neocloud customers, particularly for specialized high-density deployments, Beardsley said: “The simple answer is, yes, it does.” “It is important for the developers to bring in an advisor or lender early in the process as they are structuring the lease to ensure it is financeable,” Beardsley said. JLL’s latest report provides context for the financing side of that question. Construction lending remains liquid across credit tiers, including AI companies and neoclouds, JLL said. But non-credit-tenant deals are evaluated on a case-by-case basis, with debt metrics depending on the tenant profile, credit support from the off-taker, and project location. JLL said credit spreads on those deals generally run 200 to 300 basis points wider than investment-grade construction loans, with leverage of 70% to 80% of loan cost compared with as much as 85% for top-tier credit hyperscalers. Two Screens for Scarce Capacity For AI developers, that leaves two potentially separate screens for scarce high-density capacity: whether a facility can physically support the workload and whether the customer can satisfy the operator’s credit requirements. Mercer’s search provides one example of those constraints appearing in the same procurement process.