1% Vacancy, 81% Preleased: Where Midmarket Compute Actually Deploys in 2026

JLL's H2 2025 Data Center Report recorded North American primary market vacancy below 1% (1.2% headline, functionally sub-1% once LOI and final-negotiation holds are factored in). CBRE's corresponding analysis showed 81.5% of all capacity under construction already preleased (CBRE North America Data Center Figures H2 2025). Cushman and Wakefield's 2025 year-end report confirmed the same pattern across all major metros.

For hyperscalers, this is a logistics problem they solve with 500MW campus pre-commitments two years in advance. For midmarket operators (the companies deploying 5MW to 50MW of AI inference, private cloud, or GPU compute) it's an existential constraint. You can't deploy infrastructure in space that doesn't exist.

The Vacancy Numbers: What 1% Actually Means

A 1% vacancy rate in data center markets isn't comparable to 1% vacancy in office or industrial real estate. Data center space has much longer procurement and build cycles. When vacancy drops below 3%, available inventory is typically low-quality: wrong power density, wrong location, wrong redundancy tier, or wrong lease terms. At 1%, there's essentially nothing usable on the market.

The 81.5% prelease figure from CBRE is arguably the more important number. It means that of the approximately 5.1 GW of capacity under construction across the U.S. at the end of 2025, roughly 4.16 GW was already committed to tenants before a single rack was installed. The remaining 940MW of uncommitted capacity under construction will be absorbed within months of delivery, most of it by hyperscalers expanding existing commitments.

Who's Taking All the Space

The answer is straightforward. Microsoft, Google, Amazon, Meta, and Oracle collectively committed to over $200B in data center capital expenditure through 2026-2028 (reported in respective 2025 earnings calls and capital allocation disclosures). These five companies account for an estimated 70-80% of new data center leasing activity in primary markets.

This isn't new. Hyperscalers have dominated leasing for years. What's changed is the scale. Individual campus commitments now regularly exceed 200MW, with Microsoft's recent deals in Wisconsin and Georgia each surpassing 500MW. At that scale, a single tenant can absorb an entire market's new construction pipeline.

That leaves midmarket operators competing for the scraps. And there aren't many scraps.

The Traditional Build Timeline Problem

Even if a midmarket operator secures land and power entitlements today, the traditional data center build timeline works against them. Turner and Townsend's 2024 Data Centre Cost Index documented average construction timelines of 18-24 months for a 10MW facility in the U.S., and that's after permitting, which adds another 6-12 months depending on jurisdiction.

The cost picture is equally challenging. Turner and Townsend reported traditional data center construction costs of $10.3-11.3M per MW of IT load in primary U.S. markets. A 10MW facility costs $103-113M before land, utility interconnection, and soft costs. For a midmarket company, that's a significant capital commitment for capacity that won't be available for 2-3 years.

Modular Infrastructure: The Market Response

The modular data center market exists because traditional construction can't keep pace with demand. Grand View Research valued the global modular data center market at $29.03B in 2023, projecting growth to $75.77B by 2030 at a 17.2% CAGR (Grand View Research, "Modular Data Center Market Size Report," 2024). That growth rate isn't driven by preference. It's driven by necessity.

What "Modular" Means in Practice

The term gets overloaded. Some vendors use it to describe prefabricated power and cooling modules assembled on-site. Others mean fully self-contained data center units shipped in standard containers. The distinction matters because the deployment timelines and cost structures differ significantly.

Compare those numbers to the traditional build: $10.3-11.3M per MW and 18-24 months of construction. Modular approaches deliver at 45-65% of the capital cost and 15-25% of the timeline. That trade-off isn't perfect. Traditional builds offer more customization and typically higher density per square foot of land. But when you need capacity within 2026 and traditional construction puts you in Q1 2028, the comparison is academic.

Secondary Markets: Where Power and Land Still Exist

The vacancy crisis is concentrated in primary markets. Secondary and tertiary markets tell a different story, not because demand is lower, but because supply hasn't been as thoroughly consumed.

Case Studies: Who's Already Moving

Crusoe Energy: Factory-Built GPU Infrastructure

Crusoe Energy's approach is worth studying. Their Spark Factory in Tyler, Texas produces modular data center units designed specifically for high-density GPU workloads. Rather than building traditional data centers, they manufacture standardized compute modules in a factory environment and deploy them to sites with available power, often in secondary markets where stranded or underused power capacity exists.

The factory model lets Crusoe control quality and timeline in ways that traditional construction can't. Factory environments have consistent labor, consistent tooling, and the ability to run 24/7 production schedules without weather delays. Crusoe hasn't disclosed exact unit costs, but the modular approach inherently benefits from manufacturing economics: learning curves, bulk procurement, and parallel production.

LG CNS: Modular at Enterprise Scale

LG CNS entered the U.S. modular data center market in 2024 with a platform designed for 5-20MW deployments. Their approach uses prefabricated power and cooling modules that ship in standard containers and assemble on-site in 90-120 days. The platform supports power densities up to 50 kW per rack with direct liquid cooling - competitive for enterprise AI workloads but below the ceiling of purpose-built GPU-first platforms.

What makes LG CNS relevant to midmarket operators specifically is the deployment scale. They're not targeting 200MW hyperscale campuses. They're targeting the 5-20MW deployments that don't have the capital or timeline for traditional construction but need more than a few racks in a colo facility.

The Upper End: Cold-Plate DLC at 200 kW/Rack

The modular vertical isn't one density band - it spans roughly 40 kW/rack on the air-and-water edge up to 200 kW/rack on purpose-built cold-plate direct liquid cooling platforms. Platforms engineered for GB200 NVL72, GB300, and Vera Rubin clusters routinely spec 150-200 kW/rack with factory-fitted rack-level supply and return manifolds and dripless quick-disconnect fittings. That's the density tier that matters for frontier-model training and dense inference; it's also the tier that most traditional retrofit facilities cannot reach.

The Cost Comparison: Traditional vs. Modular

The capital efficiency gap is significant. A midmarket operator deploying 10MW via modular infrastructure typically saves $25-75M in upfront capital compared to traditional construction, with the exact savings depending on which modular platform is selected and how much of the site work is pre-existing. That freed capital can fund the IT equipment itself, or it can support a phased deployment strategy where you start with 2MW and scale to 10MW as demand materializes.

But capital cost isn't the whole picture. The time-value of compute matters too. If a traditional build puts you live in Q1 2028 and a modular deployment puts you live in Q3 2026, that's 18 months of revenue you'd otherwise miss. For a GPU inference provider charging $2-4 per GPU-hour, 18 months of delay on a 10MW deployment represents tens of millions in lost revenue.

What This Means for 2026 Deployment Decisions

The data center market isn't going to loosen up in 2026. Every analyst report points the same direction: demand continues to outpace supply, hyperscalers continue to absorb the majority of new capacity, and primary market vacancy stays near zero.

For midmarket operators, the practical playbook looks like this:

The operators deploying successfully in 2026 aren't the ones with the biggest budgets. They're the ones who accepted the vacancy reality six months ago, identified secondary market sites with available power, and committed to modular deployment timelines that match when they actually need compute capacity.

Waiting for primary market vacancy to improve is a losing strategy. The math says it won't.