- BTIG initiated “Buy” ratings for Digital Realty and Equinix citing AI investment as the primary driver of data center market expansion.
- Nvidia’s GB200 NVL72 systems require 120kW per rack, according to the company, well beyond air cooling’s practical ceiling of roughly 15 to 20kW, making direct-to-chip liquid cooling a requirement rather than an option.
- Modular data center deployments from vendors including WWT and MDCS.AI can add AI-ready capacity in months, compared to years for ground-up facility builds.
BTIG’s decision to initiate “Buy” ratings on Digital Realty and Equinix wasn’t driven by traditional enterprise storage demand, AI infrastructure investment was the call. The analyst move reflects a broader reckoning in the colocation market: facilities built for conventional IT workloads are structurally mismatched to what GPU clusters actually need.
Rethinking Rack Power and Cooling
Nvidia’s GB200 NVL72 systems already require 120kW per rack, according to the company. Air cooling stops working well above roughly 15 to 20kW per rack, at 120kW, it simply isn’t viable.
Direct-to-chip liquid cooling is the approach most providers are settling on. It circulates coolant directly over processors rather than relying on airflow, cutting energy draw from fans and air conditioning, reducing noise and allowing more GPUs in the same footprint compared to immersion cooling. Digital Realty and DataBank are among the providers redesigning facilities around hybrid cooling architectures that incorporate these systems, building toward densities that would have seemed implausible five years ago.
Interconnection Becomes Central
Power density is only half the problem. AI training workloads move enormous volumes of data between distributed nodes, sustained traffic can exceed 400Gbps between servers during large model training runs, and inference for real-time applications like autonomous vehicles demands sub-millisecond latency. Neither tolerance is forgiving.
Equinix, through its Platform Equinix service, and Digital Realty, through its ServiceFabric platform, both emphasise private, high-capacity connectivity within their campuses and to cloud on-ramps, according to the companies. The pitch to enterprise customers is straightforward: keep AI traffic off the public internet, cut egress costs and maintain the latency headroom GPU clusters need. Some specialised facilities now offer minimum 4x400GbE connectivity per rack, with support for both InfiniBand and high-throughput Ethernet fabrics.
Modular Deployments Accelerate Scale
Retrofitting a legacy data center for 100kW-plus racks is slow and disruptive. Modular data centers sidestep the problem. Pre-engineered units arrive with power, cooling, networking and storage already integrated and can be deployed in months rather than the years a ground-up facility build requires. For colocation providers under pressure to add AI-ready capacity quickly, that timeline difference is the whole argument.
WWT, MDCS.AI and Advanced Giga Inc. are among the vendors offering modular systems purpose-built for AI, with density ratings often exceeding 100kW per rack and integrated liquid-to-chip cooling. It’s worth comparing this infrastructure push with how hyperscalers are approaching the same cooling and density challenges with custom silicon.
Sustainability and Ecosystem Integration
AI’s power appetite is putting real pressure on grid capacity and carbon commitments. Liquid cooling helps here too: replacing fans and air conditioning with coolant loops cuts overall energy draw and improves Power Usage Effectiveness (PUE), the ratio of total facility power to IT equipment power. A lower PUE means less energy wasted keeping the building cool rather than running compute.
Beyond the physical layer, providers like Equinix and Digital Realty are building out managed services for GPU deployments, on-site technical support, customisable configurations and data residency controls for customers with compliance requirements. Renewable energy procurement is part of the picture too, though how aggressively individual providers are moving on this varies. The colocation market is selling a platform that AI workloads can actually run on, not floor space. Intel’s recent push into AI inference infrastructure reflects the same pressure from a different angle. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
