Vacancy in North American data centers is sitting at roughly 1% — the third straight year it’s been pinned that low, according to JLL’s midyear 2026 report. On paper, that should terrify anyone shopping for capacity. In practice, the number that should terrify you more is this one: fewer than 4% of US data centers can currently support a 100 kW-per-rack deployment, according to Uptime Institute’s 2025 Global Data Center Survey. Vacancy is measuring the wrong thing. If you’re buying capacity for an AI workload in 2026 or 2027, “available” and “usable” have quietly become two different questions, and a lot of buyers are finding that out mid-negotiation rather than before they sign.
The Density Gap Is Bigger Than the Vacancy Numbers Suggest
Uptime’s survey of operators found that 82% of respondents reported their highest-density rack running below 30 kW, while only 9% had anything at 50 kW or above. Uptime’s Daniel Bizo puts the share of US facilities that can currently support 100 kW+ racks at under 4%, and he’s careful to flag that as indicative rather than precise given sample size — but even generously interpreted, it describes a market where the overwhelming majority of built capacity was designed for a different era of compute. Most facilities built over the last decade were engineered around roughly 20 kW average rack power, with many able to stretch to 30-40 kW through relatively simple upgrades. That’s a different animal entirely from what a modern GPU cluster needs.
The mismatch shows up immediately once you look at actual hardware draw. A single Nvidia DGX H100 system can pull over 10 kW on its own — more than some legacy facilities budget for an entire cabinet. Nvidia’s current GB300 NVL72 platform, the one most hyperscalers and neoclouds are deploying right now, requires roughly 140-160 kW per rack once you account for networking and storage alongside the compute trays, per site-selection advisor James Mercer’s analysis. Air cooling becomes impractical above roughly 50 kW, Bizo notes, which means direct liquid cooling isn’t optional at these densities — it’s the entry ticket.
Why “They Have Space” Doesn’t Mean “They Have Your Space”
Mercer’s own client search is the clearest illustration of the gap between headline vacancy and usable capacity. Looking for 5 MW at 140-160 kW per rack, he approached more than ten North American operators. Nearly all of them said yes, they had space. Almost none of them could actually support the required density with direct liquid cooling once he got past the first conversation. Only two purpose-built, liquid-cooled operators engaged substantively with the requirement at all.
This is the trap buyers walk into: a provider’s marketing deck says “AI-ready,” a tour shows open floor space, and a term sheet gets drafted before anyone has verified that the facility’s power distribution, cooling plant, and structural floor loading can actually carry a 140 kW rack — let alone the 2,000+ pounds a fully loaded GPU rack with its cooling distribution unit can weigh. JLL’s own research backs this up: more than 66 GW of capacity is under construction across North America, with 77% of it in frontier markets built specifically for this load profile — which is a tacit admission that the existing, already-built stock mostly can’t do the job.
Nautilus Data Technologies, which builds liquid-cooled facilities, frames the underlying problem bluntly: most colocation contracts were written around 5-10 kW per rack assumptions, with facility design often sitting at 8-12 kW per cabinet under air cooling. A tenant that signs a standard colo agreement and then shows up with a GPU cluster is creating what Nautilus calls a “fundamental mismatch” — one where the operator becomes an unwilling co-participant in a cooling retrofit that the lease never contemplated. We’ve covered what those retrofits actually cost in CRAC/CRAH-to-liquid conversions, and the gap between what a legacy facility can deliver and what a modern rack demands is exactly why those costs climb so fast once you’re past a simple upgrade.
Need Expert Guidance?
Talk to a Data Center Expert
21+ years of hands-on experience in data center design, operations & infrastructure. Book a quick discovery call to discuss your project.
📞 Book a Discovery CallThe Credit Check Nobody Warned You About
There’s a second filter buyers aren’t expecting: creditworthiness. Mercer reports that the two liquid-cooled operators willing to engage on his client’s 140-160 kW requirement both asked, early in the conversation, whether the client was investment grade — a screening question he says he’s only seen attached to this specific density tier. It’s a rational response from operators who are financing increasingly expensive, purpose-built infrastructure against a single tenant’s commitment, but it means mid-market buyers without an investment-grade balance sheet can find themselves locked out of the facilities that can actually run their hardware, regardless of budget.
The financing math explains why. JLL reports that construction loans for non-credit-tenant deals carry spreads 200-300 basis points wider than investment-grade-backed loans, with lenders capping leverage at 70-80% versus up to 85% for top-tier hyperscaler-anchored projects. JLL’s Carl Beardsley recommends developers — and by extension, buyers — bring a financing advisor into site-selection conversations early, specifically so the lease terms being negotiated are ones a lender will actually underwrite. If you’re a buyer without a hyperscaler-grade balance sheet, that advisor conversation needs to happen before you fall in love with a floor plan, not after.
What’s Coming Makes Today’s Gap Look Small
None of this is a temporary bottleneck that resolves once the current generation of retrofits catches up. Nvidia laid out the roadmap at GTC 2025: standard Vera Rubin NVL144 racks, arriving in the second half of 2026, run at roughly 120-130 kW — comparable to today’s high end. But Rubin Ultra’s NVL576 “Kyber” rack, slated for the second half of 2027, is designed to draw 600 kW per rack, about five times current systems, and NVIDIA’s own figures suggest the architecture after that — Feynman, in 2028 — pushes toward 1 MW racks. CoolIT Systems and Accelsius have already demonstrated 250 kW rack cooling using 40°C inlet water, so the engineering path exists, but it requires facilities most buyers are touring today weren’t built for and many won’t be retrofitted for cheaply.
The cooling bill alone is instructive. Morgan Stanley’s teardown of Nvidia’s bill of materials puts the liquid-cooling hardware for a single GB300 NVL72 rack at roughly $49,860 — about the price of a Tesla Model Y — rising to an estimated $55,710 for the Vera Rubin NVL144 generation. That’s cooling hardware cost alone, before facility-level chilled water plant upgrades, power distribution, or structural reinforcement. Multiply that across a few hundred racks and it becomes obvious why operators are asking about your balance sheet before they commit capital to retrofit for you.
What This Means If You’re Actually Buying Capacity
Three things are worth doing before you sign anything. First, treat “vacancy” as a starting filter, not an answer — ask explicitly what density the facility can deliver today with direct liquid cooling in place, not what it could theoretically support with future upgrades, and get that in writing with per-cabinet (not just per-cage) power caps. Second, if you’re not an investment-grade tenant, bring your financing position into the conversation early; operators building for 140 kW+ tenants are increasingly underwriting the tenant, not just the lease. Third, build in contract language for what happens when your density requirements change — advance notice windows, a defined approval process for cooling modifications, and clarity on who bears structural and liability risk for heavier, liquid-cooled racks — because the gap between a 30 kW legacy cabinet and a 160 kW GPU rack isn’t closing on its own, and the next jump to 600 kW is already on Nvidia’s roadmap. For a closer look at how to structure these negotiations, see our data center buyer’s toolkit.
Sources
- JLL — North America Data Center Report, Midyear 2026
- Uptime Institute — Global Data Center Survey 2025
Written by
Raajeev Ratra
Data Center Infrastructure Expert | 21+ Years in DC Design, Operations & Project Management
Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.