Capacity planning / Expansion feasibility
AI Data Center Expansion: Power and Cooling Feasibility
A spare 2 MW utility allocation does not mean the facility has 2 MW of usable rack capacity. Screen the bottlenecks before approving the build.
Sections in this article
A facility has room for forty new racks, a contract showing 2 MW of unallocated utility power and a cooling plant with a spare chiller. Can it accept 40 racks at 50 kW each? Not necessarily. Rack power, available protected IT power, liquid-loop capacity, remaining air cooling, A/B distribution, structural limits and interconnection constraints need to agree. The first decision is which of those limits is actually binding.
Start with a constrained capacity screen, not a simple floor-space count. The calculator below tests three quantities you can supply today: planned IT demand, verified spare protected IT power, and verified usable thermal capacity. It deliberately does not convert unknown nameplate ratings into usable capacity or certify redundancy, pipework or electrical protection.
Separate available power from useful IT capacity
What remains after existing loads and facility electrical constraints, under required failure conditions?
What remains at relevant supply-water, airflow and outdoor design conditions?
Which floor locations pass structural, maintenance, power and pipe-routing review?
A contractual utility figure is only the start. There may be UPS loading restrictions, switchboard limits, redundancy obligations, generator performance requirements, or additional mechanical power that consumes capacity upstream of the IT boundary. Use the same defined measurement boundary for all electrical input figures in any calculation.
Likewise, “500 kW spare cooling” can mean nominal chiller capacity, primary-loop capability or heat rejection actually deliverable to the racks. Those values are not interchangeable. If the project depends on two different cooling systems, show separately how much load each is designed to capture.
Interactive GPU expansion capacity screen
Enter a proposed rack count and target IT kW per rack. Then enter spare protected IT capacity and usable cooling capacity at the same defined load boundary. The tool identifies whether the simple steady-state arithmetic fits. All entries are hypothetical until an engineer validates the available capacity.
Screen: protected IT power is insufficient by 200 kW. Rework the expansion before detailed engineering.
This calculator intentionally assumes the additional IT power becomes an approximately equal IT heat load within the facility boundary, before accounting separately for any different cooling capture arrangements. For direct-to-chip systems, the liquid and air fractions require their own equipment-specific checks. The results must not be treated as an electrical design, cost estimate or safety approval.
A 2 MW example that fails in only one dimension
In the initial example, forty racks at 50 kW create 2,000 kW of intended IT demand. The proposed site has 1,800 kW spare protected IT electrical capacity but 2,300 kW of usable cooling capacity under the assumed operating condition. Thermal headroom is positive 300 kW; protected IT headroom is negative 200 kW. Buying a new CDU would not remove this electrical constraint.
Electricity: 1,800 − 2,000 = −200 kW. Cooling: 2,300 − 2,000 = +300 kW. One failing gate is enough to pause the current proposal.
If the IT electrical limit were later verified at 2,200 kW, both numerical inputs would pass, but that would only authorize further engineering study. A facility may still fail structural checks, transient performance, network connectivity, pump-flow constraints or maintenance and failure-state capacity.
For liquid-cooled equipment with an assumed 80% of rack heat entering the coolant, this 2 MW example would suggest 1.6 MW to the liquid-side equipment and 0.4 MW to the room air. That 80/20 split is only an illustrative planning assumption; the approved server's actual heat-removal split and the facility's system boundaries must be used before procuring any CDU or CRAH.
The engineering gates the calculator intentionally cannot solve
| Gate | Evidence required | What a pass would mean |
|---|---|---|
| Electrical supply | One-line diagram, A/B allocation, approved load and protection studies | Capacity is available on the required distribution path and failure state |
| Cooling and water | Temperature/flow design, plant and CDU capacity curves, redundancy checks | Full load can be rejected under seasonal and maintenance conditions |
| Structure | Rack wet mass, point loads, access-floor and slab review | Proposed positions and installation route meet approved structural limits |
| Network | GPU fabric design, rack-to-switch reach, cable routes and uplinks | The proposed service has a verifiable connectivity plan |
| Delivery schedule | Utility upgrades, equipment lead times, maintenance outage constraints | Critical prerequisites arrive before committed commissioning dates |
Consider physical installability too. GPU rack point loads can exceed what a simplistic kg/m² calculation suggests; use the rack structural checks for the distinctions between equipment mass, footprint and contact points. Cooling pipe cross sections and pumps need the same discipline: flow feasibility requires hydraulic loss and suitable equipment curves, not only a liquid heat-load calculation.
The project should also include a credible capacity reservation policy: existing tenants and future migrations consume headroom. If a nominal 2 MW of site power is already promised to another customer, it is not available just because measured load is low today.
Use a short decision memo before capital approval
- Definition of proposed service, rack count and peak or committed IT load.
- Verified available protected power and cooling at matching system boundaries.
- Named binding constraint, owner and required remediation.
- Structural, network, routing, safety, resilience and delivery gates.
- Budget, contingencies and the next verification milestone.
- Specific evidence required before authorizing equipment purchase.
Reject the idea that an attractive payback period alone proves technical feasibility. Claims that all direct-to-chip installations drop PUE from 1.40 to 1.15, for example, do not establish measured energy savings for the proposed site. Treat the expansion as a capital project with capacity evidence, not a generic cooling-upgrade advertisement.
The best preliminary study does not predict the exact outcome of every design decision. It prevents the team from spending on the wrong solution before identifying which resource actually constrains the new racks.
Sources and further reading
Primary frameworks and reference materials. Engineering decisions require site-specific validation; legal classifications require assessment of the actual use case.