Colocation and hyperscale are the two ways most organisations get computing capacity without owning a building outright. Colocation rents space, power and cooling inside someone else's facility. Hyperscale is a single operator running enormous, standardised sites for its own workloads or for one large tenant. The difference comes down to scale, control and who carries the risk.
That sounds simple. In practice the line has blurred, because the same operator often sells both, and a hyperscaler will lease wholesale colocation when building its own site would be too slow. This guide sets out what each model is, who uses which, and how to read the trade-offs before you commit capital or capacity.
What is colocation?
Colocation is a service where a business rents physical space in a third-party data centre and installs its own servers there. The operator supplies the building, power, cooling, security and network access. You bring the hardware and keep control of it. You pay for the rack space, the power you draw, and the connectivity you use.
Colocation comes in two broad forms. Retail colocation rents by the rack or cage, suited to enterprises with a modest hardware footprint. Wholesale colocation leases whole data halls or megawatt blocks on longer terms, which is how larger tenants and cloud providers take space. The split matters because pricing, contract length and the buyer profile differ sharply between the two.
What is a hyperscale data centre?
A hyperscale data centre is a very large facility, usually run by a single operator, built to a standardised design and measured in tens or hundreds of megawatts. The term covers both the giant operators (the cloud and platform companies) and the purpose-built sites that serve them. These sites prioritise uniformity and power density over the flexibility a colocation tenant might want.
Hyperscale exists because a handful of workloads now dwarf everything else. Cloud platforms, large search and social services, and AI training clusters need capacity at a scale that retail colocation was never designed to deliver. There is no single legal definition, but a common rule of thumb puts a hyperscale site at around 40 MW of power and upwards, and the largest now run into the hundreds of megawatts.
Colocation vs hyperscale: the core differences
The two models diverge on a handful of practical points. Read them as trade-offs, not as a ranking.
Scale. Colocation deployments range from a single rack to a few megawatts. Hyperscale sites start where wholesale colocation ends and run far beyond it. A hyperscale campus can carry more power than an entire retail colocation building.
Control. In colocation you own and manage your hardware and your software stack. In a hyperscale model the operator controls the full environment. If you are a tenant buying cloud capacity, you rent an abstraction, not a cage.
Cost structure. Colocation is operating expenditure: a recurring fee for space and power, with limited upfront outlay. Building or anchoring a hyperscale site is capital-heavy, with large fixed commitments and long payback, commonly around 10 to 12 million US dollars per megawatt for standard capacity and 20 million or more for AI-ready, high-density halls.
Standardisation. Colocation accommodates mixed hardware and varied tenant requirements. Hyperscale runs near-identical halls so the operator can build, cool and maintain at scale. That uniformity is what makes the economics work.
Who carries the risk. A colocation tenant offloads building and facility risk to the operator. A hyperscale owner carries it directly, including the grid connection, the land and the energisation timeline.
Who uses colocation, and who builds hyperscale?
Colocation suits organisations that want professional-grade facilities without owning one. Enterprises moving out of an on-premise server room, software firms placing equipment near an internet exchange, and financial services needing low-latency interconnection all use it. Wholesale colocation, at the larger end, is taken by cloud providers and content platforms that need capacity faster than they can build it.
Hyperscale is the domain of the largest operators and the developers who build to suit them. The economics only stack up at volume, so the buyer pool is small but the cheques are enormous. A regional fund or a new entrant rarely builds hyperscale from scratch. It is more likely to buy a powered shell, a land position with a grid connection, or an operating asset that a hyperscaler will lease.
Where colocation and hyperscale overlap
The clean split between the two models has eroded. Several patterns now sit in the middle.
Hyperscalers lease wholesale colocation routinely, because a third-party developer can deliver a powered shell faster than the hyperscaler can secure land, power and consents itself. So a "hyperscale" workload often runs inside what is, contractually, a wholesale colocation deal.
Colocation operators, in turn, build hyperscale-grade campuses to win those tenants. The hardware density, the power blocks and the cooling are designed to hyperscale specifications even though the commercial wrapper is a lease. The result is a market where the same site can be described either way depending on who you ask.
How AI is reshaping both models
AI training and inference have pushed power density far past what either model was originally built for. Standard enterprise racks ran at a few kilowatts. AI clusters can demand 50 to 150 kW per rack, which forces liquid cooling and concentrates enormous load in a small footprint.
This changes the calculus on both sides. Colocation operators are retrofitting or building AI-ready halls with direct-to-chip or immersion cooling to stay relevant. Hyperscalers are racing to secure powered land and grid capacity years ahead of need, because the binding constraint is no longer space but deliverable power. Across most markets, the deal that decides an AI build-out is a power and land deal, not a rack-space conversation.
Which model is right for a buyer or investor?
If you need capacity for your own workloads, the question is control versus commitment. Colocation gives you flexibility and a predictable operating cost. A self-built or anchored hyperscale position gives you scale and long-term economics, at the price of capital, complexity and grid risk.
If you are an investor or developer, the more useful question is where the value sits. Increasingly it sits in the inputs: powered land, a grid connection in a viable queue, water access and a consent stage that lets a site energise on a timeline a buyer can underwrite. Those inputs surface in public filings long before any site is described as colocation or hyperscale. Reading them early is the edge.
Frequently asked questions
What is the difference between colocation and hyperscale?
Colocation rents space, power and cooling in a shared third-party facility while you keep your own hardware. Hyperscale is a very large, standardised site run by a single operator for its own or one tenant's workloads. The split is about scale, control and who carries the facility risk.
Is hyperscale a type of colocation?
Not exactly, though they overlap. Hyperscale describes the scale and design of a site. Colocation describes a commercial model of renting space. A hyperscale workload often runs inside a wholesale colocation lease, which is why the terms get used interchangeably.
Is colocation cheaper than hyperscale?
For most organisations, colocation has a lower entry cost because it is a recurring operating expense rather than a capital build. Hyperscale economics only beat colocation at very large, sustained volume, where the upfront capital, commonly 10 to 12 million US dollars per megawatt and more for AI-ready capacity, is justified by scale.
What counts as a hyperscale data centre?
There is no single legal definition. In practice the label is applied to very large sites, typically tens to hundreds of megawatts, built to a standardised design by or for a major operator.
Does AI need hyperscale or colocation?
Both. AI workloads need high power density and liquid cooling, which colocation operators and hyperscalers are both building. The harder constraint is deliverable power, so the AI question is usually about which sites have land and grid capacity available, not which commercial model is used.