North American data center vacancy has remained around 1 percent while AI demand continues to absorb new capacity.
Industry reporting cited approximately 25 GW of absorption in the first half of 2026, with a large share of capacity under construction already precommitted. Those numbers are being used in the broader debate over whether AI infrastructure is being overbuilt.
For data center operators, there is a more immediate question.
If external capacity is scarce, expensive and increasingly constrained by power availability, how much additional capacity can be recovered from the infrastructure already in operation?
The answer is often larger than a simple rack count suggests.
Empty U space is not the same as available capacity
A common capacity view begins with rack occupancy.
If a rack has ten empty U positions, it appears to have room for more equipment. But those ten U positions may not be deployable.
Power could already be near the rack limit. Cooling conditions may not support another high density server. The available electrical circuits may not provide the required redundancy. Network ports may be unavailable. The rack may sit in a zone with limited cooling headroom.
That means the facility contains several different kinds of capacity.
There is physical space.
There is power capacity.
There is cooling capacity.
There is network capacity.
There is operational reserve.
Deployable capacity is the intersection of those constraints.
For AI environments, this distinction becomes critical because a relatively small number of GPU servers can consume substantial power and create concentrated heat.
Stranded capacity is becoming financially important
Stranded capacity is infrastructure that exists but cannot currently be used because another dependency is limiting it.
A rack may have space but no power.
A data hall may have electrical capacity but insufficient cooling.
A facility may have both, but its existing asset data may be too inaccurate to support confident deployment decisions.
In a market where external data center capacity is scarce, every stranded resource becomes more valuable.
Recovering one rack of usable capacity can delay a new build. Improving power distribution can unlock equipment placement. Correcting asset data can reveal space that was thought to be occupied. Rebalancing workloads may release underused GPU resources.
This is why capacity optimization is increasingly becoming an operating discipline rather than an annual planning exercise.
Accurate asset data is the starting point
Capacity decisions are only as reliable as the inventory underneath them.
If rack location, U position, equipment configuration or power information is outdated, the calculated capacity will also be wrong.
Manual asset records struggle in environments where equipment is added, moved, replaced and upgraded continuously.
Operators therefore need continuous reconciliation between the physical environment and the management system.
Sensaka DCOS is designed to support automatic physical asset discovery, rack and U position management, equipment inventory and infrastructure monitoring. This provides a stronger baseline for capacity planning than a static spreadsheet.
The objective is simple.
Before deciding where new infrastructure can go, the operator needs to know what is actually there.
Power may be the real capacity unit
For many AI deployments, electrical capacity matters more than floor space.
A facility can have room for another row of racks and still be unable to energize them.
This changes the meaning of data center capacity.
Instead of measuring only racks or square meters, operators increasingly need to understand usable kilowatts and megawatts, with the appropriate redundancy and cooling support.
Rack power density should therefore be visible continuously.
Operators should be able to compare current consumption with configured limits, understand available headroom and identify where power is unevenly distributed.
That information can inform equipment placement.
A new GPU server should not simply be assigned to the nearest rack with open U positions. It should be placed where space, power, cooling and connectivity align.
Cooling creates another layer of stranded capacity
Cooling can strand capacity even when power is available.
This is becoming more common as high density AI systems are introduced into facilities designed around lower density enterprise equipment.
One area of the data hall may have electrical headroom but insufficient thermal margin. Another may support higher density because of containment, airflow design or liquid cooling infrastructure.
Without thermal visibility, both areas may appear identical on a rack map.
This is why capacity planning needs to incorporate real operating conditions.
Temperature, cooling zones, rack density and equipment health should inform deployment decisions. In liquid cooled environments, CDU and cooling loop status also become part of the capacity picture.
The operator needs to know whether a location can support the equipment over time, not only whether the equipment can be installed today.
Intelligent predeployment reduces expensive mistakes
When capacity is scarce, placement mistakes become more costly.
Moving equipment after deployment creates labor, risk and downtime. Discovering after installation that a rack cannot support the expected power or cooling load can delay projects. Deploying into a location with poor network connectivity can create another bottleneck.
A better process is to evaluate the deployment before the equipment arrives.
An intelligent predeployment workflow can compare the server requirements with available rack space, power, cooling and other infrastructure conditions.
This makes capacity planning closer to constraint solving.
The system can identify locations that satisfy the deployment requirements and reject locations that appear physically available but fail another condition.
The result is a more defensible answer to the question, “Where should this server go?”
Capacity should also include the compute already installed
Physical facility capacity is only part of the picture.
AI operators should also look for stranded compute.
A GPU pool can be fully allocated on paper while containing resources that are idle or poorly utilized. Fragmentation can leave capacity unavailable to large workloads even when total free accelerator count appears sufficient.
This creates a parallel between facility capacity and compute capacity.
At the facility level, power and cooling can strand rack space.
At the compute level, allocation and fragmentation can strand GPUs.
A modern AI data center therefore needs both views.
The operations team should be able to see the physical capacity supporting the compute and the actual utilization of the compute itself.
Scarcity changes the business case for better operations
When vacancy is high and new power is easy to obtain, imperfect capacity management is inconvenient.
When vacancy is around 1 percent and new data center power can take years to secure, the economics change.
Finding additional deployable capacity inside an existing facility becomes strategically valuable.
This does not mean every data center can avoid expansion. Many organizations genuinely need more infrastructure.
It means expansion should begin with a more rigorous question.
How much capacity is already present but hidden by poor asset data, uneven power distribution, cooling constraints, conservative rack rules or underutilized compute?
Sensaka's role is to help make those constraints visible.
DCOS provides physical infrastructure visibility across assets, racks, power, cooling and hardware. iDCOS can add configuration relationships and operational workflows. For AI infrastructure, broader resource management can connect physical capacity with compute pools and workload demand.
The result is a capacity model based on what can actually be deployed.
The most valuable new capacity may not be new
The debate over AI overbuild will continue.
Vacancy, precommitment and hyperscaler spending will be interpreted differently by investors depending on their view of future demand.
Data center operators have a more practical problem.
Capacity is scarce today.
That makes every unused U position, every unbalanced rack, every underutilized GPU and every unrecognized constraint worth investigating.
The first step in the next expansion project should therefore be a capacity audit of the infrastructure already in service.
Before building the next megawatt, find out whether part of it is already there.
Sources
- JLL: Data center demand exceeds expectations in H1 2026
- J.P. Morgan: Financing AI infrastructure and data centers
- Channel Dive: Data center vacancy and record construction
Originally published on the Sensaka blog.
