The rack runs out of power long before it runs out of space
We hold verified specifications for every current Dell PowerEdge, HPE ProLiant and Lenovo ThinkSystem server — 108 models, taken from the manufacturers' own technical documents. This study asks one question of all of them: how much power does a rack have to deliver, and what does that do to the number of servers you can actually install?
Updated 12 September 2026 · every figure computed from the manufacturers’ own documents · method and dataset below
Power is the binding constraint at every realistic rack budget
Scroll the chart sideways to see all of it →
Median across every rack model with node-level power supplies. “What the height allows” assumes 42U of usable space; real deployments lose several U to switching, patching and PDUs, so the gap is wider in practice.
| Rack budget | Models that fit at all | Median that fit on power | Median the height allows | Power-bound |
|---|---|---|---|---|
| 6 kW | 62 of 81 | 2 | 21 | 100% |
| 10 kW | 66 of 81 | 3 | 21 | 100% |
| 15 kW | 75 of 81 | 5 | 21 | 100% |
| 30 kW | 81 of 81 | 10 | 21 | 95% |
| 60 kW | 81 of 81 | 21 | 21 | 69% |
Scroll the table sideways to see every column →
From 500 to 3750 watts per rack unit
Scroll the chart sideways to see all of it →
Hover any point for the model. Dot size is the number of accelerators the chassis supports; colour is the cooling the manufacturer specifies.
| Model | Height | Supplies | Provisioned | Per rack unit | Accelerators | Cooling |
|---|---|---|---|---|---|---|
| PowerEdge XE9680L | 4U | 5 live of 6 × 3000 W | 15.0 kW | 3750 W | 8 | liquid |
| PowerEdge XE9685L | 4U | 5 live of 6 × 3000 W | 15.0 kW | 3750 W | 8 | liquid |
| HPE ProLiant Compute XD685 | 5U | 6 live of 12 × 3000 W | 18.0 kW | 3600 W | 8 | liquid |
| HPE ProLiant Compute DL384 Gen12 | 2U | 3 live of 4 × 2200 W | 6.6 kW | 3300 W | 2 | air |
| PowerEdge XE7740 | 4U | 4 live of 8 × 3200 W | 12.8 kW | 3200 W | 16 | air |
| PowerEdge XE7745 | 4U | 4 live of 8 × 3200 W | 12.8 kW | 3200 W | 16 | air |
| HPE ProLiant Compute DL380a Gen12 | 4U | 4 live of 8 × 3200 W | 12.8 kW | 3200 W | 10 | air |
| HPE ProLiant Compute XD230 | 1U | 1 live of 2 × 3200 W | 3.2 kW | 3200 W | — | air |
| PowerEdge XE9680 | 6U | 5 live of 6 × 3200 W | 16.0 kW | 2667 W | 8 | air |
| ThinkSystem SR780a V3 | 5U | 4 live of 8 × 3200 W | 12.8 kW | 2560 W | 8 | air |
Scroll the table sideways to see every column →
The ten densest machines in the catalogue. Every one of them is an accelerator platform, and every figure comes from the model’s own manufacturer document — the same documents behind our 108 data sheets.
Per rack unit it is a nudge; per machine it is four times the load
Scroll the chart sideways to see all of it →
The practical consequence: you do not plan for AI by finding space. You plan for it by finding a feed, and usually by finding a hall that offers 30 kW or more per rack — which is why colocation enquiries now start with power density rather than floor area. Our rack and power planner works the same arithmetic against your own estate.
Liquid cooling is 8% of the catalogue, and it sits at the top of the chart
| Cooling | Models | Median W/U | Highest W/U | Highest provisioned load |
|---|---|---|---|---|
| Air-cooled | 77 | 1600 W | 3300 W | 19.2 kW |
| Liquid-cooled | 4 | 3750 W | 3750 W | 19.2 kW |
Scroll the table sideways to see every column →
What was counted, and what it does not claim
How these figures were produced
- Population: all 108 current server models from Dell PowerEdge, HPE ProLiant and Lenovo ThinkSystem for which we hold a verified specification extracted from the manufacturer's own technical document. 81 of them are rack-mount with power supplies of their own and are included in the density figures.
- Excluded: blades, multi-node trays, Synergy modules and rack-scale sleds. Their power supplies belong to an enclosure, not to the node — a Lenovo SC750 V4 tray lists 15,000 W because that is its N1380 enclosure shelf, and counting it as “15 kW in 1U” would be wrong. Two models have no supplies at all because the chassis powers them.
- Provisioned power = the largest supply the vendor lists for the model × the number of supplies that must be live in the most demanding documented mode (the N of N+1 or N+N). Where a vendor documents no redundancy mode and the model takes two supplies, 1+1 is assumed; where it takes more than two and states no mode, half are assumed live. Every per-model figure, including the supply count and redundancy wording, is in the dataset so the arithmetic can be checked.
- Chassis with separate power zones are read from their own zoned statement, not from the summary line. The Dell XE7740 and XE7745 list “7+1 redundant” across eight bays, but the technical guide defines full redundancy as 1+1 in the CPU zone plus 3+3 in the GPU zone — four supplies live, not seven. Reading the summary line alone overstated those two machines by 9.6 kW each, and they were the top of this table until it was corrected on 12 September 2026.
- This is a facilities ceiling, not a measurement. It is what a rack feed must be able to deliver to a fully-configured machine. Actual draw depends on the configuration and the workload and is lower — usually much lower.
- Rack arithmetic assumes 42U of usable height and no derating of the power budget. Real installations lose height to switching and patching, and most electrical designs leave headroom, so the power constraint bites sooner than shown here, not later.
- No prices appear anywhere in this study, and no vendor sponsored it.
Free to reuse with attribution to Servnet and a link to this page. If you are citing a single model’s figure, cite its data sheet too — each one names the manufacturer document and revision it was taken from.
Server power density, answered
How much power does a modern server need per rack unit?
Across 81 current rack servers from Dell, HPE and Lenovo, the median is 1600 W per rack unit of provisioned power — the capacity a rack feed has to be able to deliver to a fully-loaded example. The range is 500 W/U at the bottom to 3750 W/U at the top, a spread of about 8×.
How many servers fit in a 6 kW rack?
On power rather than height: a median of 2. A 42U rack has room for far more — the median model would fit 21 times over by height — but 100% of models run out of power first, and 19 of the 81 models cannot be powered at all from a 6 kW feed.
Do AI servers really need that much more power?
Yes, but not in the way people assume. Per rack unit the difference is modest — the median machine with eight or more accelerators provisions 1900 W/U against 1600 W/U for one with no accelerator support. It is the whole machine that changes: 12.8 kW against 3.2 kW, because the chassis is four to ten rack units tall and carries up to 12 power supplies.
At what point do you need liquid cooling?
9 of the 108 models in this study are described by their manufacturer as liquid-cooled or direct-liquid-cooled — 8% of the catalogue. Their median density is 3750 W/U against 1600 W/U for air-cooled models, so on this evidence the practical crossover sits above roughly 2.5 kW per rack unit rather than at a fixed GPU count.
Is provisioned power the same as what the server draws?
No, and the difference matters. Provisioned power is what the feed must be able to deliver to a fully-configured machine: the largest supply the vendor lists, times the number of supplies that have to be live in the most demanding documented mode. A typical configuration running a typical workload draws less. Facilities are sized on the former; electricity bills are paid on the latter.
Why is rack space no longer the constraint?
Because power per rack has not grown as fast as power per server. A 42U rack is the same height it was twenty years ago; the servers in it now provision up to 15.0 kW each. At every budget below 30 kW in this study, the overwhelming majority of models exhaust the power budget while the rack still has empty U.
Work the numbers on your own estate
Talk to a UK specialist
Get expert advice or a no-obligation quote — servers, storage, networking, maintenance, finance and cloud. We reply the same working day.