How many GPUs fit within a 100 MW data center depends first on whether 100 MW means total facility power or IT power after cooling and other facility overhead. In the explicit steady-state scenario used here, a rack with 72 GPUs and approximately 120 kW, a facility-to-IT power ratio of 1.20, and a 10% allowance for IT outside those racks leave 75 MW for GPU racks: about 625 racks and 45,000 GPUs. Allocating all 100 MW to those racks instead produces a continuous-capacity estimate of about 60,000 GPUs, one-third above the scenario result; neither figure forecasts an actual deployment.[1][2]
Sources checked on 2026-10-19 22:40 Beijing time (Asia/Shanghai, UTC+8). This is an evergreen analysis of data-center power definitions, using NVIDIA hardware documentation marked updated September 17, 2026, and the US Department of Energy's energy-use definition. No securities prices are used, so trading sessions and quote delays do not apply. The 100 MW, 1.20 ratio and 10% allowance below are analytical assumptions, not parameters disclosed by an operator.
Separate facility power from rack power
NVIDIA's DGX GB rack-scale hardware guide specifies 18 compute trays per NVL72 rack, with four Blackwell GPUs per tray, or 72 GPUs. Its power-shelf section states approximately 120 kW of rack power consumption.[1] This article uses DGX GB200 as an example and adopts that approximate rack figure. It is not a measured draw for every workload or a specification for all generations and vendors.
Rack power already covers the corresponding in-rack system; adding the power of those 72 GPUs again would double-count it. Storage and other networking outside the racks may consume additional IT power. This example reserves 10% of the IT budget for that equipment, without counting in-rack components twice. Cooling and power-distribution losses belong to facility overhead, not to this 10% allocation.
The Department of Energy defines PUE as annual total facility energy use divided by annual IT-equipment energy use.[2] An annual energy ratio cannot automatically be treated as the power ratio at the hottest operating moment. This example additionally assumes a facility-to-IT power ratio of 1.20 at the steady-state operating point being analyzed. That assumption permits dividing total power by 1.20 for this budget; an annual PUE figure alone does not establish peak power capacity.
Reconcile 100 MW to 75 MW
Dividing 100 MW by 1.20 leaves approximately 83.33 MW for IT. Reserving 10% of that amount for IT outside the GPU racks uses approximately 8.33 MW, leaving 75 MW for the racks.
Convert 120 kW per rack to 0.12 MW. Then 75 / 0.12 = 625 racks, and 625 × 72 = 45,000 GPUs. The full formula is total facility MW / facility-to-IT power ratio × (1 − outside-rack IT share) / MW per rack × GPUs per rack. It translates a power budget under stated assumptions; it does not incorporate floor space, cooling delivery or equipment availability.
Skipping facility overhead and outside-rack IT gives 100 / 0.12 × 72 = 60,000 GPUs. On this continuous-capacity basis, the difference is 15,000 GPUs. The adjusted result is 25% below the shortcut, while the shortcut is 33.33% above the adjusted result. Those percentages use different denominators and are not interchangeable.
If only whole racks are counted, the shortcut becomes 833 racks and 59,976 GPUs; the adjusted scenario remains 625 racks and 45,000 GPUs. Because 120 kW is itself approximate, the integer arithmetic does not establish any site's deployment capacity to the nearest GPU.[1]
The same 100 MW label can imply a different result
If a project explicitly discloses 100 MW of IT load, dividing by 1.20 again would be inappropriate. Keeping the same assumed 10% outside-rack IT allocation leaves 90 MW for GPU racks: 750 racks and 54,000 GPUs. That is 20% above the 45,000-GPU result for 100 MW of total facility power, entirely because of the capacity definition, without any improvement in GPU performance.
If 100 MW is only a future plan, this budget applies only once that capacity can actually operate. It does not establish that the site is energized today, that servers are installed, or that GPUs have paying customers. Each requires separate project-progress or operating evidence.
Lower facility overhead has a calculable boundary
Hold total facility power at 100 MW and the outside-rack IT share at 10%. Reducing the steady-state facility-to-IT power ratio from 1.20 to 1.10 raises the GPU-rack budget to approximately 81.82 MW. Continuous capacity becomes approximately 49,091 GPUs, up 9.09% from 45,000. Rounding down to whole racks gives 681 racks and 49,032 GPUs, an increase of 8.96%. This is a conditional capacity change under the same power budget, not a 9.09% cost reduction or a 9.09% revenue increase.
Different rack power, redundancy policies or a larger outside-rack IT share require replacing the inputs and recalculating. Better annual PUE may still leave a peak cooling or electrical constraint unresolved. A lower advertised PUE therefore needs an accompanying measurement boundary and workload description.[1][2]
For readers assessing AI-infrastructure announcements, the useful next evidence is the capacity definition, the commissioning schedule and the allocation between in-rack and outside-rack IT. Those boundaries should precede estimates of deliverable compute. Multiplying megawatts straight into GPU procurement spending can otherwise allocate the same power budget more than once.
AlphaGBM researches supply, costs and commercialization across the AI value chain. For the adjacent financing constraint, see our research on CoreWeave bond yields versus company interest costs. This article does not infer any company's debt-servicing ability or order changes from that discussion.
Sources and limits
[1] NVIDIA DGX GB Rack Scale Systems hardware guide, marked updated September 17, 2026, checked October 19, 2026. Supports 72 GPUs and approximately 120 kW per rack, not an actual project's installed equipment.
[2] US Department of Energy, Cooling Water Efficiency Opportunities for Federal Data Centers, checked October 19, 2026; page metadata marks modification on April 28, 2026. Supports the annual PUE definition. The 1.20 and 1.10 ratios and 10% allowance are not project data supplied by that page.
Capacity estimates, differences and sensitivities are AlphaGBM's independent calculations. No securities prices, portfolio information or trading plans are used. This research is not investment advice and promises no investment returns.