The colocation industry has had a good run. For two decades, it sold a proposition that worked. Take enterprise IT off-premise. Put it in a professionally managed, carrier-neutral facility. Let the customer focus on running its business rather than running data centers. That proposition created a multi-hundred-billion-dollar industry. It also created an industry whose physical infrastructure, commercial models, and operational assumptions AI is, in turn, rendering obsolete.
The challenge is not that colocation operators have been caught off guard. Most of the major operators have announced AI-ready product lines, high-density zones within their campuses, and partnerships with GPU vendors. Those announcements are, however, in many cases grafted onto facilities and commercial structures built for workloads drawing 5 to 10 kilowatts per rack. Genuine AI infrastructure requires 100 kilowatts per rack and above. The gap between what most colocation facilities can deliver and what AI customers require is, consequently, wider than the industry’s marketing language suggests.
What AI Actually Requires From a Colocation Provider
AI training and inference workloads have physical requirements that differ from traditional enterprise IT in ways that go beyond power density. The power distribution architecture of a legacy colocation facility cannot simply be upgraded to serve racks drawing 100 kilowatts or more. Legacy facilities targeted racks drawing a few kilowatts each. The electrical infrastructure, busway systems, UPS configurations, and cooling infrastructure all need fundamental redesign. Incremental upgrade does not, in turn, get you there. A facility that installs a high-density zone in a corner of a campus built for 10-kilowatt racks is, specifically, not an AI data center.
Cooling is the most visible constraint. Air cooling becomes progressively less effective above 30 to 40 kilowatts per rack. The heat density that AI GPUs generate at 100 kilowatts per rack requires liquid cooling. Direct-to-chip, rear-door heat exchangers, and full immersion are the three main approaches. Retrofitting liquid cooling into a facility built for air cooling is expensive and technically complex. In some cases it is physically impossible without rebuilding the raised floor and power distribution from scratch. Most colocation operators have not, however, done that work at scale. The Long Read How Colocation Is Being Redefined by AI Workload Requirements mapped this constraint in detail. What has become clearer since is that the pace of colocation retrofit is falling further behind the pace of AI demand growth.
The Commercial Model Problem Is Harder Than the Physical One
The physical infrastructure problem is, at least, solvable with capital. The commercial model problem is, however, more structurally difficult. Traditional colocation is sold on a per-kilowatt or per-cabinet basis, with power as a separate line item billed at actual consumption. AI workloads break that model in two specific ways.
First, AI training clusters require dedicated, predictable power at densities that most colocation pricing structures were not designed to accommodate. A customer deploying a 50-megawatt AI training cluster is not, in other words, buying a few cabinets. They are consuming a significant fraction of a facility’s total power capacity on a dedicated basis. Build-to-suit or lease structures are, consequently, more appropriate than traditional colocation agreements. Second, AI inference workloads have highly variable power consumption profiles. They can spike and trough dramatically as query volumes fluctuate. Colocation facilities provisioned for steady-state loads struggle, in turn, to accommodate that variability.
The operators best positioned to address this are, notably, those that have moved toward wholesale or hyperscale-style structures rather than retail colocation. The Blog Colocation in the Age of Agentic AI: Why the Mid-Tier Operator Has a Window identified a specific opportunity for operators who move fast on power density and commercial flexibility. That window is, however, narrower than it looked twelve months ago. Hyperscalers and purpose-built AI infrastructure operators have been moving faster than most colocation operators anticipated.
Who Is Actually Winning the AI Infrastructure Business
The AI infrastructure business is, in practice, not flowing to mainstream colocation operators at the scale their announcements might suggest. It is going to three types of operators. Hyperscalers building their own campuses. Purpose-built AI infrastructure operators like CoreWeave, Nebius, and Nscale that were designed from the ground up for AI density. And a small number of colocation operators who made the genuine capital commitment to rebuild for AI workloads.
The mainstream colocation operator sitting in the middle is, in turn, competing for a smaller share of the AI market than its capacity suggests. Legacy infrastructure, retail pricing models, and a higher-density zone do not, consequently, add up to a competitive AI offering. That is not, however, a reason for despair. Enterprise customers still need managed data center infrastructure for workloads that are not AI at scale. The risk is, specifically, that the AI infrastructure revenue opportunity operators have been counting on does not materialise at the margins they have modelled.
The colocation industry has, ultimately, built enough infrastructure to weather a period of AI-driven disruption. What it has not done is reckon honestly with that gap. The sooner that reckoning happens, the better positioned the industry will be to close it rather than paper over it.
