Skip to main content

Energy layer

Cluster Buildout Sizer

Go from a power envelope (MW) or a GPU count to the reciprocal, plus cooling tonnage and 5-yr TCO — one chained buildout view.

The engineer question
If I have N MW (or want N GPUs), how big a cluster, how much cooling, and what 5-yr TCO?

Inputs

Pick the constraint you start from; the tool solves for the other.

Used when sizing by power. Facility load the substation contract delivers.

Used when sizing by GPU count. Total accelerators in the cluster.

Liquid hyperscale 2026: 1.10–1.25. Air colo: 1.4–1.6.

All-in delivered industrial / colo rate. US varies $0.04–$0.18 by state.

Result

GPUs supportedat nameplate; subtract ~20% for idle/between-job
32,051 × B200
Power envelope (input)
50 MW
IT load (accelerators + servers)+30% server overhead on 32.1 MW of accelerators
41.7 MW
Cooling loadheat ≈ IT load; 1 RT = 3.517 kW
11,847 RT
Recommended cooling
Direct-to-chip liquid (cold plate)
Acquisition capex32,051 × ~$40,000 (rough street price)
$1.28 B
Power cost (5-yr)@ 70% util, $0.08/kWh
$94.34 M
5-yr TCO (capex + power)excludes networking, storage, facility build — see assumptions
$1.38 B
$ / effective PFLOP-yrdense FP16, idealized — for relative comparison only
$5.5 k

Recommendation

50 MW fits a single hyperscale building; substation-upgrade lead-time 12–24 months in top markets (VA Loudoun, TX Dallas, OR Hillsboro, IA Des Moines). Cooling: direct-to-chip liquid (cold plate) — 1000W/accelerator forces liquid. The binding constraint at this scale is usually power-interconnect timing and networking (NVLink/InfiniBand) cost — neither of which this sizer prices.

Assumptions

  • · B200 (Blackwell): 1000 W TDP, ~$40,000 street price, 2250 TFLOPS dense FP16. TDP + overhead reconcile with the Datacenter Power Calculator; price + throughput with the GPU TCO Calculator.
  • · Power chain: facility W per accelerator = TDP × 1.3 server overhead × PUE 1.20. GPU↔MW conversion is reversible and assumes nameplate (not utilization-derated) power for facility sizing — the grid contract must cover peak.
  • · Cooling: heat to remove = IT load (≈100% of IT power becomes heat), tonnage at 1 RT = 3.517 kW (ASHRAE). The cooling regime is the per-accelerator-density default for the generation, not a per-rack CFD result.
  • · TCO: 5-yr energy = TDP × 70% utilization × PUE 1.20 × 8760 h/yr × 5 yr × $0.08/kWh. "Effective PFLOP-yr" = dense FP16 × utilization × count × lifetime — idealized, no MFU applied, so absolute $/PFLOP is optimistic (use for relative comparison).
  • · Street prices are ROUGH public estimates (NVIDIA publishes no list price); ±20–40% by volume, SKU, and quarter. Not an audited quote.
  • · EXCLUDED: networking fabric (NVLink/NVSwitch/InfiniBand — often 15–30% of a Blackwell rack), storage, host CPU/memory, datacenter build/lease, cooling capex (CDUs, plant), staff/ops, software, financing, resale, water (WUE), and grid-interconnect risk. Single-generation cluster assumed.

Worked example (default inputs)

Result

GPUs supportedat nameplate; subtract ~20% for idle/between-job
32,051 × B200
Power envelope (input)
50 MW
IT load (accelerators + servers)+30% server overhead on 32.1 MW of accelerators
41.7 MW
Cooling loadheat ≈ IT load; 1 RT = 3.517 kW
11,847 RT
Recommended cooling
Direct-to-chip liquid (cold plate)
Acquisition capex32,051 × ~$40,000 (rough street price)
$1.28 B
Power cost (5-yr)@ 70% util, $0.08/kWh
$94.34 M
5-yr TCO (capex + power)excludes networking, storage, facility build — see assumptions
$1.38 B
$ / effective PFLOP-yrdense FP16, idealized — for relative comparison only
$5.5 k

Recommendation

50 MW fits a single hyperscale building; substation-upgrade lead-time 12–24 months in top markets (VA Loudoun, TX Dallas, OR Hillsboro, IA Des Moines). Cooling: direct-to-chip liquid (cold plate) — 1000W/accelerator forces liquid. The binding constraint at this scale is usually power-interconnect timing and networking (NVLink/InfiniBand) cost — neither of which this sizer prices.

Assumptions

  • · B200 (Blackwell): 1000 W TDP, ~$40,000 street price, 2250 TFLOPS dense FP16. TDP + overhead reconcile with the Datacenter Power Calculator; price + throughput with the GPU TCO Calculator.
  • · Power chain: facility W per accelerator = TDP × 1.3 server overhead × PUE 1.20. GPU↔MW conversion is reversible and assumes nameplate (not utilization-derated) power for facility sizing — the grid contract must cover peak.
  • · Cooling: heat to remove = IT load (≈100% of IT power becomes heat), tonnage at 1 RT = 3.517 kW (ASHRAE). The cooling regime is the per-accelerator-density default for the generation, not a per-rack CFD result.
  • · TCO: 5-yr energy = TDP × 70% utilization × PUE 1.20 × 8760 h/yr × 5 yr × $0.08/kWh. "Effective PFLOP-yr" = dense FP16 × utilization × count × lifetime — idealized, no MFU applied, so absolute $/PFLOP is optimistic (use for relative comparison).
  • · Street prices are ROUGH public estimates (NVIDIA publishes no list price); ±20–40% by volume, SKU, and quarter. Not an audited quote.
  • · EXCLUDED: networking fabric (NVLink/NVSwitch/InfiniBand — often 15–30% of a Blackwell rack), storage, host CPU/memory, datacenter build/lease, cooling capex (CDUs, plant), staff/ops, software, financing, resale, water (WUE), and grid-interconnect risk. Single-generation cluster assumed.

Related tools in the Energy layer

Get notified when Cluster Buildout Sizer numbers update

We refresh the inputs as the market moves. One email when they change.

→ Get this data as JSON