Tools · Infrastructure · 17

Find what actually limits an AI datacenter.

Set the total facility power envelope, PUE, rack count, physical density and cooling. The explorer calculates how many accelerators can be active while keeping physical capacity, training throughput and inference throughput as separate questions.

Facilitytotal MW → IT
Rackpower + cooling
Trainingpeak × MFU
Inferencemeasurement only

Facility

Includes IT and facility support. The IT budget is approximated from representative PUE.
PUE is an energy ratio; applying it to power assumes it represents this operating point.
Headroom intentionally left unallocated by the calculated workload.

Racks

Use the effective thermal limit of the rack or the cooling system serving it.

Hardware and compute

Approximate CPU, memory, network and storage attached to each accelerator.

Optional inference mapping

0 disables the estimate. Enter a measurement from your model, runtime, batching policy and target SLO; it is not inferred from FLOPs.
Advanced accelerator assumptions
Result

Capacity is the minimum of independent limits.

Active acceleratorsmaximum simultaneous
Bottlenecklimiting constraint
Facility drawwith PUE applied
Training computedense peak × MFU

Four limits on the same fleet

Each bar shows how many accelerators that constraint would support on its own. The publishable capacity is the smallest value, so available power is never mistaken for physical or cooling capacity.

Facility

Usable IT capacity:

IT reserve:

Total-power headroom:

Total-envelope utilization:

Rack

Maximum active per rack:

IT load at that maximum:

Electrical headroom per rack:

Thermal headroom per rack:

Training

Active-fleet dense BF16 peak:

Sustained compute at MFU:

Sustained compute per day:

Physical density

Installed slots that can actually be used:

Low utilization here does not imply poor efficiency. Power or cooling may simply bind before physical space does.

Inference: use a serving measurement, not peak FLOPs

Aggregate throughput
Approximate completions
Method

Keep facility, rack and service separate.

1 · Facility

Total facility power is divided by a representative PUE to approximate IT capacity. The operational reserve is then removed.

2 · Rack

Each accelerator adds average GPU power plus associated IT. That footprint is checked independently against installed slots, rack electrical power and rack heat-removal capacity.

3 · Compute and serving

Training uses dense BF16 peak × MFU. Inference is calculated only when you enter measured sustained throughput from your stack; FLOPs are never converted into tokens/s.

Limits

This is a capacity model, not an MEP design or a serving benchmark.

PUE is an energy ratio and can change with load and climate. TDP is not average power. The associated-IT approximation is linear even though CPU, network and storage do not always scale that way. Thermal capacity must come from the actual rack and cooling design. MFU depends on workload and software. Inference throughput depends on model, precision, context, batching, runtime and SLO, which is why this explorer only aggregates a throughput measurement that you provide.

Compare with the Training Compute / Energy Estimator →