Hyperscale colocation and AI data center capacity

What hyperscale colocation is

Hyperscale colocation is the top end of the wholesale market. Instead of a suite or a hall, the tenant leases a dedicated building or a phase of a campus, often tens of megawatts at a time, from a developer or operator that builds and usually runs the facility. The largest cloud providers built this market. They self-build some capacity and lease the rest, because leasing adds capacity faster and keeps their capital for servers and chips.

AI companies, GPU clouds and large enterprises now lease the same way. The commercial structure is familiar from wholesale leasing: rent per kW of critical IT capacity, power billed on top, long terms. What differs is the scale, how early the commitment is made, and the facility the tenant needs.

Who is leasing at campus scale

Cloud providers still take most of the capacity, but they are no longer alone. JLL estimates 2026 demand at 59% hyperscalers, 11% neoclouds (GPU cloud providers) and 7% pure-play AI companies, with the typical enterprise requirement at 500 kW to 3 MW (JLL, August 2026). CBRE expects demand to shift toward 25 to 75 MW sites for inference and enterprise AI, and reports growing interest in 2 to 20 MW urban infill sites (CBRE, H1 2026).

Leases disclosed in 2026 show what campus-scale AI commitments look like:

The pattern: 15 to 20 year terms with extension options, delivery one to two years after signing, and credit support where the tenant's balance sheet is thin. The largest tenants' commitments dwarf any single campus. Oracle reported $260 billion of leases signed but not yet started, substantially all data centers, running 15 to 19 years (Oracle 10-K), and CoreWeave reported $38.5 billion (CoreWeave 10-K). Those tenants are competing for the same pipeline you are.

What AI racks demand from a facility

A decade ago an enterprise rack drew 5 to 10 kW. Rack-scale GPU systems draw more than ten times that, and each generation goes higher:

At these densities air cooling is out. CBRE notes that 100 kW GPU servers can only be cooled by a closed-loop liquid cooling system, and that converting an air-cooled hall to liquid can delay fit-out by more than six months (CBRE, H1 2026). NVIDIA designs its Vera Rubin generation racks for 45°C warm-water inlet, which lets many sites reject heat with dry coolers instead of chillers (NVIDIA, March 2026). Most of the market is not there yet: the average rack in AFCOM's 2026 operator survey drew 27 kW, and 36% of operators had liquid cooling deployed (AFCOM, 2026).

For a tenant, that turns into a short list of facility questions that decide whether a hall can take the deployment at all:

• Power per rack and per row. The sustained limit for your contiguous footprint, not the building's best-case figure.
• Liquid cooling. Whether direct-to-chip cooling is installed or only designed for, who owns the coolant distribution units, the facility water supply temperature and flow per rack, and how much heat can still go to air.
• Structure. Floor loading for dense, liquid-filled racks, and clear height for overhead busway, cable trays and piping.
• Electrical design. Busway capacity, distribution voltage, and whether redundancy can be reduced for training halls to save cost.
• Network. Space and pathways for the dense fiber that GPU clusters need between racks and to the outside world.
• The next generation. Whether the hall can take a denser system in two years without a rebuild.

Preleasing: capacity is committed before it is built

Large blocks of finished capacity barely exist. Almost everything a campus-scale tenant can lease is still under construction, and most of that is already committed:

• Primary markets: 80.4% of the 7,481 MW under construction was preleased in the first half of 2026, and less than 1,500 MW of future capacity remained unclaimed, about six months of demand (CBRE, August 2026).
• North America overall: JLL counts 66 GW under construction, 95% pre-committed, and says most tenants signing today are contracting for 2028 deliveries (JLL, August 2026).
• What is left in finished buildings: across the Americas, the largest single availability in an operating facility was 15 MW, and only 28 availabilities exceeded 5 MW (Cushman & Wakefield, H1 2026). Worldwide, only seven operating colocation facilities had 20 MW or more available, and none of them was in the Americas (Cushman & Wakefield, May 2026).

So a search for 10 MW or more is really a search of the construction pipeline. The question is not only which building, but which phase, which delivery date, and how firmly that date is committed.

For a tenant, three things follow:

1. Start earlier than feels necessary. Work backward from the date you need power energized, not from the date you want to sign.
2. Commit in stages. A first phase with rights of first offer on later phases secures growth without paying for it before you need it.
3. Price certainty against delivery risk. A future phase may cost less than energized space today, but only if the delivery date is enforceable.

Power is the binding constraint

Every hyperscale search ends up as a power search. Cushman & Wakefield puts the average time to secure power for new large loads at 5.0 years in the Americas (May 2026), and CBRE says power procurement takes more than ten years in some markets (CBRE Midyear Review, August 2026).

Utilities and regulators are also changing the terms of a large connection. Those obligations reach tenants through the lease:

The response from developers is to bring power with them. CBRE reports developers weighing behind-the-meter generation where utility delays run past contracted delivery dates (CBRE, H1 2026). Cleanview counts 59 announced behind-the-meter data center projects totaling about 90 GW, mostly gas, of which about 2 GW was operating in mid-2026 (Cleanview, May 2026). Announced is not operating. If a campus depends on onsite or bridge power, ask who owns and fuels it, what it costs per kWh, what emissions permits it needs, and exactly how and when it hands over to the grid.

Lease terms specific to AI capacity

Density and cooling commitments

Contracted kW per rack and per row, liquid cooling capacity, supply temperature and flow, with remedies if the facility cannot deliver them.

Cooling equipment ownership

Who buys, owns, maintains and replaces the coolant distribution units and secondary loop, and who is liable for a leak.

Delivery tied to hardware

A delivery date and ramp that match your GPU shipments, with rent credits for delay. Idle hardware waiting for a hall is expensive.

Redundancy and pricing

Lower redundancy for training halls should carry lower rent. Get the redundancy level and the price for each hall in writing.

Technology refresh

The right to upgrade to denser systems, and the operator's obligation to support higher power and cooling when you do.

Credit support

Letters of credit, deposits or prepayment scaled to your credit profile. Newer AI companies often face higher security requirements, which are negotiable.

Sublease and assignment

The right to sublease capacity if demand shifts, so a long commitment does not become a stranded one.

Frequently asked questions