HOME / BLOG / RACK POWER DENSITY PLANNING
POWER PLANNING · SEP 9, 2026
How to Plan Power Density for an AI Rack
Density planning is arithmetic, not preference. This is the sequence we run on every rack RFQ that crosses the engineering desk: establish the load, convert it to amperes at the actual distribution voltage, size the protective device against the continuous-load rule, balance the phases across an outlet map, and only then decide what hardware goes in the rack. Every number below is reproducible with a calculator.
STEP ONE
Start with the IT load, not the breaker.
Load list first, distribution second
Write the rack down before you price anything: accelerator trays, head nodes, switches, storage, and the in-rack conversion and buffering hardware that rides along with them. A modern AI rack spans roughly 30 kW to 142 kW depending on generation, against a 7–10 kW legacy enterprise design point — a tenfold range that no single distribution answer covers. The load list is what tells you whether this rack is a 22 kW problem or a 142 kW problem.
Nameplate is a ceiling, draw is a curve
Every server nameplate is the worst case the vendor will defend, not the number the rack actually pulls. Training load also steps: GPU power can move by tens of kilowatts in under a second while the utility chain answers on a 1–90 minute clock. Plan the feed against nameplate, then instrument at the outlet so the operating curve you invoice and cool against is measured rather than assumed — ±1% billing-grade metering at the outlet is what makes that comparison honest.
STEP TWO
kW to amperes: the three-phase arithmetic.
For a three-phase feed, current per phase is I = P ÷ (√3 × V). IT power supplies present a power factor close enough to unity that the simple form is the right planning tool — and the numbers below are the ones our PDU and busway quotes are built from.
| Rack load | Amperes at 400V 3φ | Amperes at 415V 3φ | Where this class sits |
|---|---|---|---|
| 11 kW | 15.9 A | 15.3 A | Single 3φ PDU, enterprise compute |
| 22 kW | 31.8 A | 30.6 A | One 3φ 32A PDU at rating |
| 44 kW | 63.5 A | 61.2 A | Dual 3φ 32A PDUs, A/B fed |
| 132 kW | 190.5 A | 183.7 A | NVL72-class, low end |
| 142 kW | 205.0 A | 197.6 A | NVL72-class ceiling we build to |
Why the same kW reads differently on two voltages
Current is the number that decides conductor size, breaker frame and connector rating, so a 415V hall feed draws about 3.6% less current than an identical load on 400V — enough to move a 142 kW rack from a 300 A frame to a 250 A frame at the protective device. Confirm the actual distribution voltage before you approve a single-line: the difference is small in volts and material in copper.
STEP THREE
Breaker sizing and the 125% continuous rule.
A data centre rack is a continuous load by definition — it runs for more than three hours at a time. That puts the protective device at 125% of the calculated load, then up to the next standard frame size. Skipping this step is the most common reason a rack trips at 80% of the nameplate it was sold.
| Rack load | Load current (400V 3φ) | × 1.25 | Standard device size |
|---|---|---|---|
| 11 kW | 15.9 A | 19.9 A | 20 A |
| 22 kW | 31.8 A | 39.7 A | 40 A |
| 30 kW | 43.3 A | 54.1 A | 60 A |
| 44 kW | 63.5 A | 79.4 A | 80 A |
| 132 kW | 190.5 A | 238.2 A | 250 A |
| 142 kW | 205.0 A | 256.2 A | 300 A |
The 142 kW case, worked
At 400V three-phase a 142 kW rack draws 205 A. Multiply by 1.25 and the device must carry 256 A continuously — the next standard frame is 300 A. Downstream, that same rack needs two 3φ PDUs at 63A inputs, each rated past 22 kW, with outlet metering accurate to ±1% so the split between the A and B sides is visible rather than theoretical.
Where the rack stops and the hall starts
The breaker you just sized usually lives upstream of the rack, at the busway tap-off or the panel. Row feeds run 160–800A on track busway with tap-off boxes from 80–400A and per-tap metering, which is how an 800A run serves an NVL72-class row around 132–142 kW. Write the tap rating into the rack spec: a 400 A tap cannot feed a rack that needs a 300 A device plus diversity.
STEP FOUR
Phase balance decides your usable capacity.
A three-phase PDU does not fail when its total load reaches nameplate — it fails when the heaviest phase does. On a 400V WYE feed each phase sits at 230V line-to-neutral, so the per-phase current ceiling is a hard wall that skew in the outlet map reaches early.
| Load split across phases | Heaviest phase share | Usable PDU capacity (3φ 32A, 400V) | Capacity lost |
|---|---|---|---|
| 33 / 33 / 33 | 33% | 22.2 kW | none |
| 36 / 33 / 31 | 36% | 20.4 kW | 8% |
| 40 / 33 / 27 | 40% | 18.4 kW | 17% |
| 50 / 50 / 0 | 50% | 14.7 kW | 34% |
| 60 / 40 / 0 | 60% | 12.3 kW | 45% |
Balance the map, not the invoice
Phase rotation is decided when the outlet map is drawn, which is a documentation exercise, not a hardware one. During DFM we phase-balance the outlet map against your rack elevation and verify it in the pilot build, so the rack you receive already distributes load evenly instead of asking a commissioning engineer to re-cord it in a live hall.
What imbalance costs downstream
Skew does not only strand PDU capacity. It also loads the neutral, raises upstream losses, and hands the facility team a metering picture that disagrees with the load list. Per-outlet ±1% metering turns that into a number you can act on: three phase totals that should track within a few percent, and one that visibly does not.
STEP FIVE
Headroom rules worth writing into the spec.
Headroom is the difference between a rack that absorbs a tray upgrade and one that needs an electrician on a change window. These are the rules we apply when quoting AI rack power.
- Keep planned load at or below 80% of the protective device. A 250 A frame means 197.6 A of continuous load at 415V — which is exactly the 142 kW ceiling, with nothing left over. Above that, step to the next frame.
- Budget the conversion loss. At 96% Titanium efficiency, a 142 kW rack turns roughly 5.7 kW into heat inside the enclosure. That heat is a cooling and a derating input, and it does not appear on the IT load list.
- Reserve one branch circuit per PDU. Rack power changes arrive as an added tray or a faster node, and the cheapest place to find 3–5 kW is a spare branch circuit that was specified at design time.
- Size each A/B feed for the whole rack. A dual-cord rack is not a 50/50 split. Each side is rated to carry the full load so that one feed can be maintained without dropping the rack.
- Plan metering accuracy with the same rigor as ampacity. A 2% inlet-only reading on a 142 kW rack is nearly 3 kW of unexplained drift per hour. Outlet-level ±1% data is what keeps the power plan and the invoice in agreement.
- Do not exceed a tap-off rating to save a run. Tap-off boxes run 80–400A; if the rack device needs 300 A, the tap has to match before the busway layout is frozen.
WORKED EXAMPLE
A 142 kW rack, link by link.
Density planning is chain planning. The table below is the sizing sequence for an NVL72-class rack at the top of the range we build, with the number that decides each link.
| Link | Sizing decision | Deciding number |
|---|---|---|
| Primary feed | 415V AC today, 800VDC ready path | 197.6 A per phase at 142 kW |
| Overhead busway | 800A track with a 400A-class tap | Row feed 160–800A |
| Protective device | 250 A frame at 415V, 300 A at 400V | 125% continuous rule |
| Rack PDU | Dual 3φ 63A units, phase-balanced outlet map | 22 kW+ per unit, ±1% metering |
| Conversion | 54V shelves of 5.5kW+ modules, N+N | 96% Titanium, ~5.7 kW of loss |
| Buffering | 51.2V LFP BBU plus 48V LIC modules | Millisecond and seconds-to-minutes layers |
The plan is only finished when it is measured
A power plan that lives in a spreadsheet has one failure mode: reality. Commissioning closes the loop — per-outlet metering against a reference, phase totals that track, and a per-tap reading upstream that confirms the rack is drawing what the load list predicted. Anything that disagrees is a question for the same engineering desk that sized the chain, which is the practical argument for buying the busway, PDU, PSU and buffering layers from one factory instead of four.
RELATED SYSTEMS
The hardware behind this brief.
Rack PDU
Basic through switched+outlet tiers, 16–63A inputs, 3φ past 22 kW+, ±1% billing-grade metering and SNMPv3 / Modbus TCP / Redfish.
Explore rack PDUs →CRPS Server PSU
550W–5.5kW+ modules at 96% Titanium, 54V direct-to-rail for GPU shelves, N+N hot-swap, 100% burn-in before shipment.
Explore CRPS →Integrated Rack Power
Busway, PDU, PSU and buffering engineered as one tested system — one QA file, up to 142 kW per rack.
Explore integration →
DENSITY PLANNING
