Every conversation about AI data center power starts in the wrong unit. Square footage tells you almost nothing now. The number that governs the design, the capital plan, and the interconnection request is kilowatts per rack — and it has moved by an order of magnitude in about three years. This piece is the arithmetic, with every figure sourced and every wide range left visibly wide.
It is written for investors and developers rather than operators. The question is not how to run the facility. It is how much power to secure, what that power converts into in compute terms, and which of your assumptions is most likely to be the one that breaks the model.
What is the short answer, in numbers?
Four figures carry most of the answer.
- A conventional enterprise rack draws roughly 7 to 10 kW. That was the planning assumption most existing floor space was built to.
- A current-generation AI rack is rated at 132 kW. That is NVIDIA's GB200 NVL72 — 72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled 19-inch frame.
- One megawatt of IT load holds about 7.6 of those racks, or roughly 545 GPUs.
- A single modern AI building commonly exceeds 100 MW, and announced campuses reach gigawatt scale.
The rack figure comes from Uptime Institute Intelligence, where research director Daniel Bizo writes that the NVL72 “is rated at 132 kW” (Uptime Institute, 25 June 2025). The traditional-rack range, the AI range, and the facility scale come from a September 2025 review of AI data center grid impacts by Chen, Wang, Colacelli, Lee and Xie, which puts traditional racks at “7 kW - 10 kW/rack”, AI racks at “30 kW - over 100 kW/rack” with averages above 60 kW, and modern hyperscale facilities at “power demand exceeding 100 MW” (arXiv:2509.07218, 29 September 2025).
The macro picture those add up to is the part already in the public record. The IEA estimated global data centre electricity consumption at around 415 TWh in 2024, about 1.5% of global electricity, and projects around 945 TWh by 2030 in its Base Case, just under 3%, with accelerated servers accounting for almost half of the net increase (IEA, Energy and AI, 2025).
How much power does one AI rack actually draw?
Start with what is installed, not what is announced. The Uptime Institute's 2025 Global Data Center Survey — more than 800 owners and operators, fielded April to May 2025 — found “greater adoption of racks in the 10–30 kW range” and stated plainly that “few facilities exceed 30 kW, and extreme densities are as yet rare” (Uptime Institute, 30 July 2025). In the 2024 edition, only about 1% of operators reported any rack above 100 kW.
So the 132 kW rack is not the industry average. It is the leading edge, and the gap between it and the installed base is the whole engineering problem.
| Density tier | Rack power | Racks per MW of IT load | What that tier represents | Source for the density figure |
|---|---|---|---|---|
| Conventional enterprise | 7–10 kW | 100–143 | General-purpose servers. The assumption most existing floor space was designed to. | Chen et al., arXiv, Sept 2025 |
| What operators actually run today | 10–30 kW | 33–100 | Mixed enterprise plus some acceleration. Few facilities exceed 30 kW. | Uptime Institute survey, 2025 |
| AI, de-escalated on purpose | 20–90 kW | 11–50 | Fewer GPUs per rack, deliberately, to stay inside an existing power and cooling envelope. | Uptime Intelligence, June 2025 |
| Current-generation AI, full rack | 132 kW | 7.6 | GB200 NVL72: 72 GPUs, 36 CPUs, liquid-cooled, one frame. | Uptime Intelligence, June 2025 |
| Next generation, published target | up to 1,000 kW | 1.0–1.7 | NVIDIA's stated design target for 800 VDC racks from 2027. | NVIDIA Technical Blog, May 2025 |
The third row is the one most often left out of vendor material, and it is the row a developer should read twice. Uptime's analysis describes operators deliberately reducing density — 64-GPU racks held “within 80 kW to 90 kW”, four 8-GPU frames per rack brought to “under 50 kW”, further configurations at roughly 26 kW sustained, and lower-powered GPU options taking “rack power below 20 kW”. Density is a design choice constrained by the building, not a specification handed down by the chip vendor. A site that cannot support 132 kW can still run accelerated compute; it runs less of it per cabinet.
How many racks and GPUs fit in a megawatt?
Racks per megawatt is trivial arithmetic: 1,000 ÷ rack kW. At 8 kW that is 125 racks. At 132 kW it is 7.6. That is the whole calculation, and it is why converting existing floor space is rarely a matter of adding cabinets.
GPUs per megawatt is the more useful figure, and it behaves in a way that surprises people. Power per GPU sets it, not rack density. Work it through on two members of the same family:
- GB200 NVL72: 132 kW ÷ 72 GPUs = 1.83 kW per GPU
- GB200 NVL36: approximately 66 kW ÷ 36 GPUs = 1.83 kW per GPU
Both land in the same place, because they are the same silicon packaged differently. At 1.83 kW per GPU, one megawatt of IT load supports about 545 GPUs whether you rack them 36 or 72 at a time. The NVL36 figure is from the arXiv review cited above, which lists “GB200 NVL36 at ~66 kW, NVL72 at ~120 kW”.
Within a hardware generation, higher rack density buys floor space and interconnect locality. It does not buy more GPUs per megawatt. If you are underwriting an AI facility, the conversion rate from megawatts to compute is set by the chip, not by the cabinet — which means a density upgrade is a real estate and networking decision, and a generation change is the only thing that moves your compute-per-megawatt.
Sizing a site, or checking someone else's power model?
Book a Discovery Call →Worked example: how much power do 10,000 GPUs need?
A single worked example is worth more than any amount of prose, because it exposes exactly which assumption is doing the work. Target: 10,000 GB200-class GPUs. Every step below is our arithmetic on the sourced figures above, and every step is reversible if you disagree with an input.
| Step | Calculation | Result | Where the risk is |
|---|---|---|---|
| 01. Target compute | Stated requirement | 10,000 GPUs | None — this is the input. |
| 02. GPUs per rack | GB200 NVL72 configuration | 72 | Low. A published platform configuration. |
| 03. Racks required | 10,000 ÷ 72 = 138.9, rounded up | 139 racks | Low, though partial racks are wasted power envelope. |
| 04. IT load at rated power | 139 × 132 kW | 18,348 kW = 18.3 MW | Moderate. 132 kW is a design rating, not a measured average. |
| 05. Facility load at PUE 1.09 | 18.3 MW × 1.09 | 20.0 MW | High. 1.09 is a measured result from a hyperscale operator, not a default. |
| 06. Facility load at PUE 1.30 | 18.3 MW × 1.30 | 23.8 MW | The realistic band for a well-designed new build. |
| 07. Facility load at PUE 1.54 | 18.3 MW × 1.54 | 28.2 MW | The 2025 global average. Applies if you inherit an existing envelope. |
| 08. What you request | Facility load, plus redundancy, plus refresh headroom | > 28.2 MW | Highest. Headroom is the assumption nobody writes down. |
Read rows 05 through 07 together. Identical compute, three defensible efficiency assumptions, and an 8.2 MW spread — about 41% more grid capacity for the same 10,000 GPUs. In most markets that difference is not a line item. It is the difference between a connection you can get and one you cannot.
Row 08 is where models quietly fail. A 28 MW connection sized to 132 kW racks holds 212 racks of IT load. Sized to NVIDIA's published megawatt-class target — it states that 54 VDC distribution approaches its practical limit around 200 kW per rack and that its 800 VDC architecture is built for “1 MW IT racks and beyond, starting in 2027”, coinciding with production of its Kyber rack-scale systems (NVIDIA Technical Blog, 20 May 2025) — the same connection holds fewer than 20. The connection does not shrink; the rack count does. We have written separately about what developers do when the grid connection itself is the binding constraint — that is a different problem from this one and it has different answers.
What is the difference between nameplate and what the meter sees?
Three separate numbers get collapsed into one in most conversations, and keeping them apart is the difference between a model that survives diligence and one that does not.
The design rating
132 kW for a GB200 NVL72. It is a thermal design figure used to size power distribution, busbar, and cooling capacity. It is not a prediction of average consumption. Worth noting honestly: NVIDIA's own public product page for the platform publishes GPU and CPU counts and confirms the rack-scale liquid-cooled design, but does not publish a kW rating at all. Every citation of 132 kW, including ours, traces to secondary technical analysis of vendor documentation.
The facility-planning figure
The peer-reviewed literature uses roughly 120 kW for the same rack. That is an 11% spread on the single most-quoted number in the field, between two credible sources, and the correct response is to carry both ends in the model rather than pick the one that suits the conclusion. On 139 racks it is the difference between 16.7 MW and 18.3 MW of IT load.
What the meter records
Lower than the rating on average, and far more volatile than either figure suggests. The arXiv review reports that training workloads can produce “power fluctuations of hundreds of megawatts within only seconds”, with large GPU clusters swinging “by tens to hundreds of megawatts within sub-second intervals”. Average draw governs your energy bill. Peak and ramp rate govern what the utility and your own electrical design have to tolerate.
The practical consequence: average consumption sizes the operating cost, and instantaneous behaviour sizes the infrastructure. An AI facility is a worse citizen on the grid than its annual energy figure implies, and that gap is a real diligence item rather than an operational footnote.
How much does cooling and overhead add to the number?
Power usage effectiveness is total facility power divided by IT power, and it multiplies the entire load rather than a portion of it. Which is why it is the highest-leverage single assumption in the whole exercise.
| PUE | Where the figure comes from | Facility power for an 18.3 MW IT load | Extra grid capacity vs. 1.09 |
|---|---|---|---|
| 1.09 | Google's 2025 fleet-wide trailing-twelve-month average across large-scale data centers, all seasons, all sources of overhead. | 20.0 MW | — |
| 1.08–1.09 | Meta's 2024 reported average, per the arXiv review. Comparable operator, comparable result. | ~19.8–20.0 MW | — |
| below 1.30 | The band the literature attributes to modern large-scale facilities generally. | up to 23.8 MW | up to +3.8 MW |
| 1.50–1.60 | Typical enterprise facilities, per the same review. | 27.5–29.3 MW | +7.5 to +9.3 MW |
| 1.54 | The Uptime Institute's 2025 global weighted average — flat for the sixth consecutive year. | 28.2 MW | +8.2 MW |
Two things about that table deserve emphasis. First, the global average has not moved in six years, which means efficiency is not something a market delivers on its own — it is designed in or it is absent. Second, 1.09 is a measured operating result from a company that builds its own facilities at enormous scale. Assuming it for a first-of-kind development is not optimism, it is a modelling error.
Where the overhead goes is documented. The IEA's component breakdown puts servers at “around 60% of electricity demand in modern data centres”, storage at “around 5%”, networking at “up to 5%”, and cooling at anywhere “from about 7% for efficient hyperscale data centres to over 30% for less-efficient enterprise data centres” (IEA, Energy and AI, 2025). That 7%-to-30% range for cooling is the PUE spread, restated as a component. It is also why the cooling architecture is a capital-planning decision rather than a mechanical one — we treat it that way in our data center architecture work, and the thermal design choices themselves are covered in our post on water-free cooling and self-generated power.
How large is a site, and what does that mean for the grid?
Facility scale has detached from the enterprise era entirely. The arXiv review describes modern hyperscale facilities with “power demand exceeding 100 MW” and campuses “planned to scale to the gigawatt (GW) level”. Applied to the arithmetic above, a 100 MW building at PUE 1.30 carries roughly 77 MW of IT load, about 583 GB200 NVL72 racks, and on the order of 42,000 GPUs. A gigawatt campus is that figure times ten.
National-level projections are where the honest uncertainty is widest, and it is worth showing rather than smoothing:
- Lawrence Berkeley National Laboratory found US data centers consumed 176 TWh in 2023, 4.4% of US electricity, and projected 325–580 TWh, or 6.7% to 12%, by 2028, with demand having “more than doubled” between 2017 and 2023 “largely due to the growth in AI servers” (Berkeley Lab, 2024 United States Data Center Energy Usage Report, December 2024).
- EPRI, fourteen months later, put 2024 consumption at 177–192 TWh, roughly 4–5% of US electricity, and projected 9% to 17% by 2030 at approximately 380–790 TWh — about 60% above its own 2024 estimate (EPRI, Powering Intelligence 2026, February 2026; figures confirmed via Data Center Knowledge and E&E News, as EPRI's report site does not render for automated retrieval).
Those two ranges do not reconcile, and pretending otherwise would be the easier thing to write. A 6.7–12% forecast for 2028 and a 9–17% forecast for 2030 describe a field where a leading national laboratory's high case became another institution's low-to-middle case in a little over a year. EPRI attributes its own revision to the volume of announced and under-construction projects in the preceding eighteen months, and names the reasons it may still be wrong: completion rates, AI workload trajectories, and supply chain, labour and permitting constraints.
The planning implication is not that one number is right. It is that demand forecasts are revising upward faster than grid capacity can be added, which is what turns a power question into a schedule question and then into a financing question.
Where are these numbers genuinely uncertain?
A page of figures should name its own weak points, so here are ours.
| The figure | Why it is softer than it looks |
|---|---|
| 132 kW per rack | NVIDIA does not publish a kW rating on its public product page. The figure is well corroborated in technical analysis, but it is a design rating restated by third parties, and credible sources use roughly 120 kW for the same rack. |
| Installed rack densities | Uptime's data is self-reported by more than 800 survey respondents, weighted toward established operators. It describes the installed base well and the frontier poorly, which is the correct bias for capital planning and the wrong one for forecasting. |
| Megawatt-class racks by 2027 | A vendor's published design target, not a shipped product. NVIDIA has strong reason to state it and a track record of hitting roadmaps, and it is still a roadmap. |
| National demand forecasts | LBNL and EPRI disagree materially, and EPRI revised its own estimate up 60% in about two years. Treat any single figure in this category as a scenario. |
| PUE | Not standardised in practice. Boundary definitions, seasonality, and whether a facility has reached stable operations all move it. Google states its boundary conditions explicitly; most operators do not. |
| Per-site MW figures | We have deliberately excluded the widely repeated “100 to 750 MW per site” range. We could not trace it to a primary source and are not going to launder a vendor figure into a citation. |
The pattern worth taking from that table: the settled numbers are the ones about hardware you can buy today, and the unsettled ones are every projection built on top of them. Underwrite the first category. Stress-test the second.
We design compute infrastructure with the energy position treated as part of the architecture rather than a utility bill — the reasoning behind that is on our AI data center architecture page, and the approach we take to sizing before specifying anything follows the same order as the rest of our method: map first, then control, then leverage, then strategy. On a power question that order is not a preference. Sizing a building before you have written down your kW-per-GPU and PUE assumptions is how a 20 MW model becomes a 28 MW connection request that nobody budgeted for.
Frequently Asked Questions
It is set by rack density, not floor area. A conventional enterprise rack draws roughly 7 to 10 kW; a current-generation AI rack such as NVIDIA's GB200 NVL72 is rated at 132 kW in a standard 19-inch frame. One megawatt of IT load therefore holds about 100 to 140 conventional racks but only 7 to 8 of those AI racks. A single modern AI building commonly exceeds 100 MW of power demand, and some campuses are planned at gigawatt scale. To get the number you actually request from a utility, take the IT load and multiply it by the facility's power usage effectiveness, which ranged from 1.09 at the measured best to a 1.54 global average in 2025.
Between about 30 kW and more than 100 kW, with published averages above 60 kW for AI racks against 7 to 10 kW for traditional ones. The specific figure most often quoted is NVIDIA's GB200 NVL72 at 132 kW for 72 GPUs and 36 CPUs in one liquid-cooled rack. Be careful with that number in two ways. It is a design power rating, not a measured average draw, and NVIDIA's own public product page for the platform does not publish a kW figure at all, so every citation of 132 kW is tracing back to secondary technical analysis. Some peer-reviewed work uses roughly 120 kW for the same rack as a facility-planning figure. An 11 percent spread on the headline number is worth carrying explicitly in a model rather than resolving by preference.
Racks per megawatt is simply 1,000 divided by the rack's kW. At 8 kW that is 125 racks; at 132 kW it is about 7.6. GPUs per megawatt behaves differently and this is the part most sizing conversations get wrong. Power per GPU, not rack density, sets it. A GB200 NVL72 at 132 kW for 72 GPUs and a GB200 NVL36 at roughly 66 kW for 36 GPUs both work out to about 1.83 kW per GPU, so both deliver roughly 545 GPUs per megawatt of IT load. Within a hardware generation, higher density buys you floor space and interconnect locality. It does not buy you more GPUs per megawatt.
The site-level totals differ by an order of magnitude and the per-rack figures differ by roughly the same. Traditional facilities were planned at rack densities of 7 to 10 kW; the Uptime Institute's 2025 survey of more than 800 owners and operators found greater adoption of racks in the 10 to 30 kW range, with few facilities exceeding 30 kW and extreme densities still rare. A single AI rack at 132 kW is therefore not an increment on that installed base, it is a different power distribution architecture, a different cooling system, and a different structural load. The industry-wide consequence is visible in the electricity data: the IEA estimated global data centre consumption at around 415 TWh in 2024 and projects around 945 TWh by 2030 in its Base Case.
NVIDIA has published its own target, which is the most defensible figure available: it states that 54 VDC power distribution approaches its practical limit around 200 kW per rack and that its 800 VDC architecture is designed for 1 MW IT racks and beyond starting in 2027, coinciding with production of its Kyber rack-scale systems. The physical reason is instructive for anyone underwriting a build today. Supplying a 1 MW rack at 54 VDC would require up to 200 kg of copper busbar per rack, and up to 200,000 kg across a gigawatt facility. A megawatt-class rack means fewer than two racks per megawatt of IT load, so the design question shifts from how many cabinets fit to whether the electrical room, the floor, and the interconnection were sized for a generation that had not shipped when the building was permitted.
Because it multiplies the entire IT load, not a slice of it. Power usage effectiveness is total facility power divided by IT power. Google reports a 2025 fleet-wide trailing-twelve-month average of 1.09 across its large-scale data centers including all sources of overhead; the Uptime Institute's 2025 survey put the global weighted average at 1.54, effectively flat for the sixth consecutive year. Applied to an 18.3 MW IT load, that spread is the difference between a 20.0 MW facility and a 28.2 MW facility, or about 41 percent more grid capacity for identical compute. The IEA's component breakdown shows where it goes: servers average around 60 percent of demand in modern data centres, while cooling ranges from about 7 percent in efficient hyperscale facilities to over 30 percent in less-efficient enterprise ones.
Underwrite megawatts, then convert to compute at a rate you have written down and can defend. Three assumptions carry most of the risk and each should be a range rather than a point: the kW per GPU of the hardware generation you expect to install, the power usage effectiveness the design will really achieve rather than the one in the marketing material, and the headroom you are holding for a refresh cycle whose densities are already published and are higher than what you are building for. The thing we would flag hardest is that a fixed grid connection converts into a shrinking rack count with every generation, so a site sized only for the racks you can name today is a site that stops being useful before the debt is repaid.