For most of my career, "high density" meant something specific and relatively stable: a rack pulling eight to ten kilowatts was worth a conversation, and anything above 15kW got flagged for special mechanical treatment. Those assumptions held long enough that entire categories of due diligence – floor loading, fire protection, electrical service sizing and cooling – were built around them. Those were reasonable numbers to build a checklist on.

That checklist is no longer sufficient. Computing hardware is moving faster than the assumptions behind existing facilities and the loads that today's and tomorrow's AI systems may require.

An NVIDIA GB200 NVL72 rack draws roughly 120 to 132kW at nominal load, depending on configuration, and HPE recommends provisioning busway for peaks of about 192kW. The GB300 NVL72 runs higher at roughly 135 to 142kW nominal with peaks near 155kW. About 90 percent of that heat is captured by liquid cooling, but the remaining ten percent, roughly 14kW of air-side heat per rack, is on its own about as much as an entire rack my old checklist would have flagged as high density.

And the roadmap does not stop there. NVIDIA has publicly targeted roughly 600kW per rack for its Rubin Ultra NVL576 “Kyber” systems in the second half of 2027 and has published an 800 volt DC power architecture designed to support racks of 1MW and beyond. These are no longer laboratory concepts or isolated hyperscaler experiments – they are increasingly part of the conversations showing up in data-center RFPs.

The problem is not a lack of standards. Standards for high-density and liquid-cooled facilities exist. But standards and codes don't retrofit a facility for the next generation of equipment. We're seeing operators evaluate facilities that were considered AI-ready only a short time ago discover structural, cooling, water, electrical or resiliency assumptions that no longer match the tenant's deployment plans.

Where the gap actually lives

Start with cooling, because it's the most visible symptom. As rack densities rise, conventional air cooling becomes increasingly difficult to scale. At some point, the amount of airflow required to remove the heat becomes impractical for a conventional air-cooled architecture, whether raised floor or slab. A 132kW GPU rack is not simply a higher-density version of a traditional ten or 15kW rack; it can require a fundamentally different approach to heat removal.

That's why liquid cooling has moved from a differentiator toward a basic requirement for many current AI deployments. But the important real-estate question is what must change around the cooling technology itself: structural capacity, water distribution, heat rejection, leak detection, drainage, electrical infrastructure and space for the equipment that supports the liquid loop.

Design guidance for liquid cooling keeps maturing, but a building permitted five years ago reflects the guidance of five years ago. A facility may have been designed and permitted around one set of assumptions and then be asked to accommodate a tenant whose hardware roadmap has moved several generations ahead over the term of a typical lease.

Fire protection presents a similar challenge. NFPA 75 governs the protection of information technology equipment spaces, and NFPA 855 addresses stationary energy storage systems, including lithium-ion batteries. But today's facilities can combine much higher rack power, concentrated heat, liquid-cooling equipment and lithium-ion battery systems in ways that require careful, application-specific engineering. What matters is whether the facility's original fire-protection strategy still matches the equipment being proposed.

Structural and electrical requirements face the same basic issue. A single GB200 NVL72 rack weighs roughly 3,000 pounds (about 1.36 metric tons), before counting the coolant distribution units, piping and fluid that come with it. Floors designed around legacy rack weights need to be checked, not assumed. Codes and standards provide a framework for safe design, but many of the industry’s major codes and standards – including the National Electrical Code – follow multi-year revision cycles, while local jurisdictions may adopt those editions on different schedules. NVIDIA, by contrast, now introduces a new data-center platform roughly every year.

Power distribution is changing too. NVIDIA's 800 VDC architecture, which the company says cuts copper requirements by roughly 45 percent compared with conventional 415 VAC distribution, is one example of the industry's move toward higher-voltage power. Higher-voltage distribution can reduce the amount of conductor material and space required as power increases, but NVIDIA itself acknowledges that it introduces new challenges in safety, standards and workforce training, and for projects, new questions around equipment, protection, maintenance, listings, and permitting. Projects will have to work through those requirements rather than assume a traditional data-center distribution model will simply scale upward.

Why this is a business and real estate problem, not just an engineering one

When a landlord asks me whether their shell can support a tenant's proposed IT load, I'm not only evaluating whether the mechanical engineer can design a solution. I'm evaluating whether the building's structure, utility infrastructure, code-compliance pathway and insurance posture can absorb equipment that may be substantially different from what the facility was originally designed around. That's a fundamentally different question than "Can we cool this?" – and it is often asked too late, after a lease is signed and a tenant's hardware roadmap has already changed.

This matters because code compliance is a baseline, not a guarantee that a building is ready for a tenant's future deployment. Permitting authorities, insurers, lenders, and investors may each look at different aspects of the risk. A facility can be compliant with today's requirements and still require significant investment to support tomorrow's IT load.

That distinction is becoming increasingly important. The question for owners and investors should not simply be, "Is this facility code compliant?" It should also be, "What is this facility actually capable of supporting rack density increases?"

What owners and investors can do now

Owners and investors don't need to wait for the next NFPA, ASHRAE or code revision to act. By the time a standard is updated, the hardware may have moved again. Three questions belong in every acquisition, lease and development decision now:

• What rack densities can this facility realistically support in three to five years?
• Which assumptions in the original design are based on legacy density and cooling requirements?
• How would insurers, lenders, permitting authorities and tenants evaluate the facility if AI workloads materially increased the projected thermal and electrical load?

The goal is not to predict the future perfectly. It is to treat the pace of technology change as a design input, just as owners already treat power availability, utility interconnection and construction schedules as design inputs, and to understand how much flexibility the building has when the future arrives.

That means engineering to the density roadmap a tenant is likely to need, not the density it deploys today. It means testing insurance, code compliance, and infrastructure against actual thermal and electrical loads. And it means "the mechanical engineer says it works" is the start of the conversation, not the end.

The industry has spent years focused on power availability and interconnection queues. Those constraints are real, but another sits underneath them: the hardware roadmap can move faster than the codes, standards and existing-building assumptions around it. Closing that gap will take more than updated codes. It will take owners, operators, developers, and investors making decisions today based on tomorrow's density, cooling, and infrastructure requirements.

As AI continues to reshape infrastructure needs, the winners will be planning not for today's workloads, but for those still to come. At JLL, we bring together data center strategy, technical due diligence, infrastructure planning, and real estate expertise to prepare portfolios for what's coming next, not just what's deployed today.

Your next step: If you're evaluating whether your facilities and infrastructure are ready for the next wave of AI workloads, JLL can help assess opportunities, identify constraints and develop a roadmap for future growth. Contact us and we'll work through it together.