TL;DR
- Space Is Not the Constraint, Usable Space Is: A 5 MW requirement for rack-scale GPU systems went to more than ten operators. Nearly all had space; almost none could support the density on the timeline
- Liquid-Cooling Ready Is Not Liquid-Cooled: The date that matters is when facility cooling reaches the white space, and it often sits well behind the building
- Credit Can Matter as Much as Capacity: Two operators raised, unprompted, whether the tenant was investment grade
- Buyers Have More Flexibility Than They Use: Phasing, multiple sites, air-cooled platforms and an earlier start all widen the field
# # #
By James Mercer, Principal, Metro Colo Advisory
North American data center vacancy has never been lower. CBRE put it at a record 1.4 percent across primary markets in the first half of 2026. It is easy to read that as “there is no space.” For AI buyers, the more accurate reading is harder: there is space, but very little of it can run what they are trying to deploy.
This year we took a requirement to market that tested that distinction directly: 5 MW of capacity for rack-scale GPU systems, liquid-cooled, commissioning in early 2027, with a path to roughly twice that within a year. We took it to more than ten operators across North America. Nearly every operator had space. Almost none could support the requirement. Here is why, and what buyers can do about it.
The Gap Between Space and Usable Space
Current rack-scale systems such as NVIDIA’s GB300 NVL72 draw 140 to 160 kW per rack. Most of the market was not built for that. According to the Uptime Institute, rack density growth is concentrated between 10 and 30 kW, about one facility in eight reports any racks at 30 to 59 kW, and racks above 100 kW remain rare.
Three things decide whether a hall can take this hardware, and none of them can be changed quickly:
Power density per rack. A hall designed for 15 kW racks cannot be averaged up to 150. The electrical distribution to each position has to exist.
Where the cooling loop ends. Direct liquid cooling needs facility water delivered to the rack row. A building with chilled water at the plant room but not in the white space is not a liquid-cooled building.
Floor loading. A fully populated rack-scale system weighs about 3,000 pounds. Many raised floors were never rated for it.
An operator can have megawatts of vacant, powered space and still have nothing that clears all three.
“Liquid-Cooling Ready” Is Not Liquid-Cooled
The most common answer we heard was a variation of “our facilities are liquid-cooling ready.” In practice that meant the building could support liquid cooling once equipment was procured and installed, and one operator put the lead time for facility-side cooling equipment at nine to twelve months. In one case, a building was due in early 2027 while its liquid-cooled space was not due until months later.
It also helps to separate two kinds of equipment that often share a name. The coolant distribution units inside the racks usually ship with rack-scale systems. The facility-side units that connect the building’s water to the rows belong to the operator. When an operator cites a long lead time, it is worth asking which of the two they mean.
The Gate Nobody Mentions: Credit
Two of the operators that engaged most seriously raised the same question, unprompted: was the tenant investment grade? One made a meeting about its future capacity conditional on the answer.
That makes sense from the operator’s side. A multi-year, multi-megawatt commitment is a large receivable, and in a market this tight an operator can choose its tenants. But it means some declines framed as “no capacity” may be credit decisions. AI buyers without investment-grade credit should expect the question, and arrive with an answer: a parent guarantee, a deposit, or a stronger entity on the lease.
What Buyers Should Ask Before Shortlisting a Facility
- What density can you support per rack, sustained, in the hall you are offering? Not the campus maximum.
- On what date will facility cooling be live at the rack row? Not the building handover date.
- What is the floor rating in that hall?
- What is the expansion path, and is it approved or on a roadmap? Many operators can start a deployment. Far fewer can commit to doubling it.
- What will you need to see on credit? Asking early saves weeks.
Where the Flexibility Is
Buyers have more room than the headline requirement suggests. Starting smaller with a committed expansion path opens more facilities than insisting on the full load on day one. Splitting a deployment across two sites works when the workloads are independent. And where the workload allows, air-cooled GPU servers at 30 to 50 kW per rack fit a far larger share of the market than rack-scale systems do.
The biggest lever is time. Capacity for 2027 is being committed now: CBRE found 80.4 percent of capacity under construction already preleased. A buyer who starts twelve to eighteen months ahead, with credit support lined up and flexibility on phasing, is choosing from a market. A buyer who starts six months ahead is choosing from whatever is left.
# # #
About the Author
James Mercer is Principal of Metro Colo Advisory, an independent data center advisory practice based in New York. He places colocation, wholesale and AI infrastructure requirements with operators across North America, and advises developers and investors on how facilities will compete for tenants.
