Rubin is already in production. The question is whether your infrastructure can handle 200-kilowatt racks when it arrives.
Every few years, NVIDIA releases a new generation of graphics processing units that raises the ceiling on what AI systems can do. Each release has followed a familiar pattern: more compute power, higher memory bandwidth, better performance on the tasks AI models care about most. Organizations upgrade their infrastructure, performance improves, and the cycle continues.
NVIDIA Rubin, the successor to the current Blackwell generation, is no longer a roadmap item. The Vera Rubin platform entered full production in early 2026, with volume shipments from partners beginning in the second half of the year and early systems already running at major cloud providers. The gap between what Rubin requires from the infrastructure around it and what most current facilities were designed to provide is wider than in any previous generation. Understanding what is changing, and why, matters for anyone making infrastructure decisions that will still be relevant in two or three years.
Why Each New GPU Generation Creates an Infrastructure Problem
A graphics processing unit is, at its core, a chip that performs calculations in parallel at very high speed. The faster and more capable the chip, the more electricity it needs to run, and the more heat it generates in the process. These two factors, power draw and heat output, are what drive infrastructure requirements.
Early GPU systems used by AI researchers drew perhaps 200 to 300 watts per chip. The chips in current high-performance AI servers draw anywhere from 700 watts to well over 1,000 watts per unit. A server with eight such chips, plus the surrounding components, can easily draw 10 to 15 kilowatts. Rubin-class systems push significantly beyond that, with per-chip power consumption continuing to climb.
The physical space that holds these chips has a limited capacity for both power delivery and heat removal. When the chips inside the rack generate more heat than the cooling system can remove, performance degrades and hardware fails. This is why infrastructure has to evolve alongside the chips themselves.
The Rack Density Challenge
Traditional data center racks were designed around power densities of roughly 5 to 10 kilowatts per rack. AI infrastructure has already moved far past that number. Current Blackwell rack-scale systems operate at roughly 120 to 130 kilowatts per rack. Rubin-class racks are specified at approximately 190 to 230 kilowatts, and NVIDIA’s published roadmap for 2027 points to rack designs in the range of 600 kilowatts.
This matters because the physical structure of a data center (its floors, its power distribution systems, its cooling infrastructure) was typically designed for a specific density range. A building that works well at 10 kilowatts per rack begins to face serious constraints at 60 kilowatts. At 130 kilowatts and beyond, most general-purpose facilities simply cannot support the load without significant modification.
Organizations planning to deploy next-generation AI hardware need to understand whether the facilities they are considering can actually handle the density requirements of that hardware. This is not a minor technical detail. It determines whether a planned deployment is physically possible.
Cooling Is the Constraint That Gets Underestimated
Heat is the practical limit that power density creates. More electricity flowing through a rack means more heat coming out of it. Traditional data centers remove that heat by pushing cold air through the facility. At low densities, this works well. At the densities that next-generation AI hardware requires, it breaks down entirely.
Air simply does not transfer heat efficiently enough to keep ultra-dense racks within safe operating temperatures. The physics of air cooling impose a ceiling. Above that ceiling, no amount of additional airflow solves the problem. The heat has to be removed by a medium that is better at carrying it: liquid.
Liquid cooling systems route cooled water or other fluids directly to the components generating heat, either through cold plates that sit directly on the chips or through full immersion systems where hardware is submerged in a non-conductive liquid. These approaches are more complex and more expensive than air cooling, but they are no longer optional. Rubin-class systems are fully liquid-cooled by design. There is no air-cooled configuration.
For organizations that have not yet deployed liquid cooling, the transition requires planning well in advance. It involves changes to the physical facility, new infrastructure procurement, and updated operational procedures. It is not something that can be done quickly in response to a hardware delivery that arrived last week.
Networking Has to Scale With the Hardware
A GPU cluster depends on its components communicating with each other at extremely high speeds. The bandwidth requirements of next-generation AI hardware are increasing in parallel with compute performance. InfiniBand and high-speed Ethernet connections that are adequate for current deployments may become bottlenecks when Rubin-class hardware is running workloads that require constant, high-volume data exchange between chips.
Planning for next-generation GPU deployments means evaluating not just whether the physical facility can handle the power and cooling requirements, but whether the networking infrastructure can move data fast enough to avoid constraining the hardware it connects.
Why Infrastructure Decisions Need to Be Made Now
GPU hardware has long lead times. Planning, procurement, facility preparation, and deployment all take time. With Rubin systems already in production and volume shipments underway, organizations that have not yet evaluated their infrastructure against next-generation requirements are not ahead of the curve. They are behind it.
The facilities that will be ready to deploy next-generation AI hardware as it becomes broadly available are the ones being designed and built now. This means evaluating current infrastructure honestly against what Rubin-class systems require, identifying gaps, and addressing them before they become constraints.
At Vertical Data, infrastructure design for current and next-generation compute density is central to how we approach every deployment. The difference between a facility that is genuinely ready for what comes next and one that is optimized for today’s hardware is wide enough that it will determine which organizations can move quickly as the next generation arrives and which ones cannot.
