A facility can have plenty of total substation capacity and still be unable to accept a new GPU rack — because density isn't just a total-power problem, it's a per-square-foot and per-circuit problem.
Facility power capacity is the total electrical load a data center's utility service, generators, and UPS plant can deliver, usually expressed in megawatts (MW) for the whole site or hall. Rack power density is a completely different measurement: how much of that power is concentrated into each individual rack's footprint, expressed in kW per rack. A hall can have 10 MW of total facility capacity and still be structurally unable to host a row of 100 kW GPU racks, because that capacity has to arrive through a finite number of branch circuits, PDUs, and — just as importantly — through cooling infrastructure sized for a much lower kW-per-rack design point. Facility capacity answers "how much power exists on site." Rack density answers "how much power can any one rack actually draw, and can everything around that rack — wiring, breakers, airflow — actually support it."
A facility's total MW capacity is a single aggregate number that says how much power the whole site can pull from the grid (or generators) at once. Rack density is a local constraint about how that power gets concentrated into one physical footprint — roughly 24–30 square feet for a standard rack. A hall could theoretically have 10 MW of facility capacity spread evenly across 2,000 legacy racks at 5 kW each, with nothing left over to give any single rack more without either taking capacity from elsewhere or expanding the substation. Raising kW per rack instead requires larger branch circuits and PDUs to that specific rack, and — usually the harder constraint — enough cooling capacity delivered specifically to that rack's row (via CRAH capacity, liquid manifolds, or CDU sizing) to remove that concentrated heat. This is why retrofitting an existing legacy-density hall for GPU racks is rarely just "pull more power to it" — it typically requires touching electrical distribution, cooling distribution, and sometimes structural floor loading (since a fully populated 100+ kW liquid-cooled rack can weigh substantially more than a legacy air-cooled rack) all at once.
Total available megawatts is a necessary condition, but not a sufficient one. Those 2 MW still have to physically reach the specific racks that need it through existing branch circuits, RPPs, and PDUs sized for a much lower density — and, just as often the actual limiting factor, the cooling system serving that specific row or aisle has to be able to remove that much concentrated heat without overheating neighboring racks that share the same CRAH unit or containment zone. A facility can have abundant total capacity and still require a substantial (and expensive) retrofit — new branch circuits, upgraded PDUs, added CDUs or in-row cooling, sometimes reinforced flooring — before a handful of high-density GPU racks can actually be installed in an existing hall built around a much lower design density.
Explains the difference between facility power capacity (total site-level megawatts available from utility, generators, and UPS) and rack power density (kW concentrated into a single rack's footprint), and why a facility with abundant total capacity can still be unable to host new high-density GPU racks without a targeted retrofit.
Facility power capacity is the aggregate electrical load a data center site or hall can support, set by the utility interconnection size, backup generator capacity, and UPS plant sizing, typically expressed in megawatts. It answers a top-level question: how much total power exists on site. It says nothing on its own about how that power is subdivided across the hall's racks.
Rack power density measures how much of that total power is delivered into a single rack's physical footprint (commonly 24-30 square feet), expressed in kW per rack. Legacy enterprise racks commonly draw 3-8 kW; modern AI/GPU racks routinely draw 30-100+ kW as accelerator counts and per-GPU power both climb. Rack density is constrained locally by branch circuit ampacity, PDU rating, and — usually the binding constraint — how much cooling capacity (air or liquid) can be delivered to that specific rack or row.
Total facility capacity being sufficient does not automatically mean any individual rack can be upgraded to high density — that power still has to travel through existing branch circuits and PDUs sized for a lower design point, and be matched by cooling capacity delivered specifically to that rack's location. Retrofitting an existing hall for GPU-class density typically requires simultaneous upgrades to electrical branch circuits, cooling distribution (often adding CDUs or in-row units), and sometimes structural floor loading, since populated high-density liquid-cooled racks can weigh considerably more than legacy air-cooled equivalents.
When planning new AI/GPU capacity, engineers typically size a hall's electrical and cooling infrastructure around a target kW-per-rack design density from the outset, rather than treating total facility megawatts as the only planning number — a hall designed around 8 kW/rack racks cannot simply "absorb" a request for 60 kW/rack racks without a substantial, often costly, redesign of both power and cooling distribution down to the rack level.
Design targets vary by facility, but many new-build AI/GPU halls are being designed around 40-100+ kW per rack, compared to historical enterprise design points often in the 3-8 kW per rack range — an order-of-magnitude increase driven largely by GPU accelerator power and rack-scale system architectures like NVIDIA's NVL72.
Not strictly at every density level, but practically, yes, above a certain point. Air cooling alone becomes increasingly difficult to justify economically and thermodynamically beyond roughly 20-30 kW per rack, which is why most modern high-density GPU racks pair with direct-to-chip or immersion liquid cooling rather than pushing air cooling further.
Yes — this is a recognized industry problem sometimes called stranded capacity, where a facility has purchased or built more total electrical capacity than its cooling infrastructure, floor loading, or circuit distribution can actually deliver to individual racks, leaving some contracted power effectively unusable without further infrastructure investment.
Higher rack density concentrates more heat into a smaller footprint, which can actually improve PUE if it's matched with efficient liquid cooling (which typically consumes less energy per kW of heat removed than air cooling at the same density), but can worsen PUE if the facility tries to remove that concentrated heat with air-cooling infrastructure not designed for it, forcing inefficient over-cooling of the whole space.