Power delivery, rack density and cooling architecture now determine deployable AI capacity. Floor space is no longer a useful proxy.
Most data centers advertise available capacity. Far fewer can actually support a modern AI deployment.
The gap shows up in due diligence constantly. A provider lists tens of thousands of square feet of vacant space, the site tour goes well, and then the engineering review finds the facility can power perhaps a tenth of what the buyer needs. Nothing about the space was misrepresented. Space was simply the wrong question.
AI data center capacity is constrained by power delivery, rack density and cooling architecture rather than physical space. A conventional enterprise rack drew 5 to 15 kW. A current-generation Blackwell-class rack draws roughly 120 to 140 kW, and denser platforms are close behind. Floor space and deployable AI capacity have effectively decoupled.
What Does “AI-Ready” Data Center Capacity Actually Mean?
AI-ready capacity refers to a facility’s ability to deliver contracted power, rack-level density and liquid cooling sufficient for modern GPU clusters. Physical space is a secondary variable. A facility can be largely vacant and still unable to accommodate a high-density deployment.
Most colocation contracts and RFP templates continue to express capacity in square feet. The convention dates from an era when a fully loaded rack drew about 10 kW, and it no longer describes anything useful: a single AI rack now draws more power than an entire row of legacy racks. Organizations evaluating AI-ready colocation need to underwrite power, not space.
Why Available Floor Space No Longer Predicts Deployable AI Capacity
Power availability has become the primary constraint on AI deployments. At Data Center World 2026, Google’s data center technology team emphasized that as rack densities rise, power, not compute, is becoming the limiting factor.
A facility can have an entire vacant data hall and still power only a fraction of the racks that hall could physically support at AI density. The relevant question is contracted power per rack and per hall, not total facility square footage.
The one area where space still matters is structural, not dimensional. A fully populated GB200 NVL72 rack weighs about 1.36 metric tons, several times the load of a traditional server rack, so these systems often require reinforced flooring or a slab with verified load capacity.
How Much Power Do AI Racks Actually Require?
Modern AI racks draw between roughly 80 kW and more than 130 kW. The trajectory from here is steep and already public. NVIDIA’s Vera Rubin (VR200 NVL72), in volume production for the second half of 2026, is rated at roughly 190 to 230 kW per rack, and the Rubin Ultra Kyber rack is specified at about 600 kW for the second half of 2027.
The broader market is moving in the same direction. AFCOM’s 2026 survey put average rack density at 27 kW, up from 16 kW a year earlier and about 7 kW in 2021, the largest year-over-year jump in the report’s decade of data. AI-specific deployments sit well above even that rising average.
| Era | Typical Rack Density | Deployment Type |
| 2021 | ~7 kW per rack | Traditional enterprise compute (global average) |
| 2025 | ~16 kW per rack | Mixed enterprise and cloud (global average) |
| 2026 | ~27 kW per rack | Global average, per AFCOM 2026 |
| 2026 (AI-optimized) | 80 to 140 kW per rack | GB200/GB300 NVL72-class GPU clusters |
| H2 2026 (shipping) | ~190 to 230 kW per rack | NVIDIA Vera Rubin (VR200 NVL72) |
| H2 2027 (announced) | ~600 kW per rack | NVIDIA Rubin Ultra Kyber (NVL576) |
The trajectory matters more than the current figures. A hall engineered for 15 kW racks cannot be upgraded to 130 kW by adding circuits. The change reaches from the utility feed through the busway to the rack, and in most facilities it amounts to a rebuild rather than an upgrade.
Why Cooling Readiness Is a Hard Constraint, Not an Upgrade Path
Air cooling cannot reject the heat produced above roughly 30 to 40 kW per rack. Beyond that threshold, liquid cooling is a requirement. Direct-to-chip liquid cooling is standard for systems in the 120 to 140 kW range; rear-door heat exchangers, designed for 30 to 40 kW racks, cannot carry the load. Immersion cooling supports densities above 200 kW.
| Cooling Method | Supported Density | Status for AI Workloads |
| Air cooling | Up to ~30 to 40 kW per rack | Insufficient for current AI racks |
| Direct-to-chip liquid | ~140 kW today; designs shown to ~250 kW | Standard for current-generation GPU systems |
| Immersion cooling | 200+ kW per rack | Emerging option for extreme density |
Retrofit timelines are the practical constraint. Coolant distribution units, leak detection and compatible rack designs cannot be added quickly to a facility that lacks them. Cooling architecture is a design decision made at construction, which is why it belongs early in any capacity evaluation rather than as a follow-up item.
One detail that surprises teams new to these deployments: liquid-cooled racks still reject part of their load to air. On a deployed GB200 NVL72, roughly 115 kW is handled by the liquid loop and about 17 kW, mostly networking and storage, is air-cooled. A facility’s air handling still matters at high density. It just stops being the primary system.
What Role Does Deployment Timeline Play in AI-Ready Capacity?
Securing power is now the longest lead-time item in most AI infrastructure projects, frequently exceeding GPU procurement timelines. Grid interconnection queues in constrained markets such as Northern Virginia have extended to five years or more for new connections, and some markets, including Dublin, have paused new data center connections altogether.
Operators have responded with prefabricated electrical and cooling modules, standardized rack designs and planning horizons that span multiple GPU generations. For enterprises, the test is narrower: an availability date is meaningful only if the power behind it is already contracted and the cooling is already installed.
A Practical Framework for Evaluating AI-Ready Data Center Capacity
In practice, AI infrastructure planning for a GPU deployment comes down to five factors. They should be assessed together; a deficiency in any one negates the others.
- Contracted power per rack and per hall, not total facility square footage
- Electrical architecture that delivers AI-scale density without a full retrofit
- Installed liquid cooling infrastructure, including coolant distribution units and leak detection
- A contracted deployment timeline backed by secured power and cooling capacity
- Headroom for the next generation of GPU density, not only the current one
A facility that performs well on floor space and poorly on these factors is legacy capacity, whatever the marketing materials indicate.
What Questions Should Enterprises Ask Colocation Providers?
Provider due diligence should quantify power per rack, distinguish installed cooling from planned cooling and verify that stated timelines are backed by secured utility or on-site power commitments. Five questions cover most of it:
- What is the maximum contracted power per rack available today, not on a future roadmap?
- Is direct-to-chip or immersion cooling already installed, or does it require a build-out?
- What is the facility’s current PUE (Power Usage Effectiveness, the ratio of total facility power to IT power), and how does it compare to the 1.1 to 1.2 achieved by leading operators?
- How does grid interconnection status in this market affect the deployment timeline?
- What headroom exists for the next generation of GPU rack density?
What AI-Ready Capacity Really Measures
AI has fundamentally changed what capacity means inside a data center. Available floor space may still matter, but it no longer answers the question enterprises actually need to ask.
The real measure of AI-ready capacity is the ability to deliver power, cooling and deployment readiness at the density modern GPU infrastructure requires. Organizations that assess AI infrastructure capacity on those terms first make better decisions and avoid costly deployment delays.
Vertical Data helps enterprises plan, finance and deploy AI infrastructure based on real-world operational requirements, not legacy capacity metrics.
