The rise of generative AI has triggered a seismic shift in data center design. What began as a hardware evolution – driven by Nvidia’s latest GPU platforms – has quickly become a full-stack infrastructure challenge. Hyperscalers are racing to deploy AI supercomputers with rack densities exceeding 100kW, and the ripple effect is now reaching colocation providers and enterprise data centers.

This isn’t just about more power or faster chips. It’s about rethinking how we cool, scale, and deploy infrastructure in an era where workloads are more demanding, deployment timelines are tighter, and flexibility is paramount.

The density dilemma

Traditional colocation environments were built around five to 10kW racks cooled by air. That model is no longer viable. AI workloads – especially those involving large language models (LLMs), computer vision, and real-time inference – require tightly packed GPU clusters that can push rack densities well beyond 100kW. Some projections suggest densities could reach 1MW per rack within the next few years as multi-rack GPU pods become more tightly integrated and liquid cooling becomes standard.

This density explosion creates a thermal challenge that air cooling alone cannot solve. Liquid cooling is emerging as a necessity, not a luxury. But retrofitting entire rooms for liquid cooling can be expensive, disruptive, and slow.

Rack-level containment: A smarter approach

One of the most promising strategies is rack-level containment – a design approach that isolates hot and cold zones at the rack level, enabling ultra-high-density cooling without requiring full room redesigns. This method allows operators to deploy high-performance infrastructure in existing facilities, accelerating time-to-compute and reducing capital expenditure.

Rack-level containment also supports hybrid cooling architectures, where air and liquid cooling can coexist. This is critical for operators who need to support mixed workloads or transition gradually to liquid-cooled systems.

Designing for flexibility and safety

Beyond cooling, colocation providers must consider how their infrastructure supports rapid deployment, modularity, and safety. AI companies are moving fast, and they expect infrastructure partners to keep pace.

Key design considerations include:

• Modular deployment options: Roll-in or mounted racks that can be installed quickly in greenfield or retrofit environments.
• Environmental resilience: Cabinets rated for indoor and outdoor use (e.g., NEMA 3R) to support edge deployments or unconventional locations.
• Integrated fire suppression: As power densities rise, so do safety risks. Built-in fire suppression systems are becoming essential for high-density environments.
• Future-proofing: Infrastructure must be adaptable to evolving GPU architectures, power delivery standards, and cooling technologies.

The path forward

AI is not just changing what we compute – it’s changing how we build and operate the infrastructure that supports it. For colocation providers, the opportunity is clear: evolve quickly, embrace modular and rack-level cooling strategies, and design for flexibility, safety, and scale.

The next generation of data centers won’t be defined by square footage or legacy cooling systems. They’ll be defined by how well they support the workloads of tomorrow – starting today.