AI Rack Power & PDU Guide: From kW Math to 800VDC

Powering a high-density AI rack is a sizing problem before it is a shopping problem. Start with the load: a GB200 NVL72 cabinet draws roughly 120–140 kW, an order of magnitude above the 5–15 kW a conventional enterprise rack was built for. That single fact dictates the rest of the chain. You feed the rack with higher-amperage three-phase PDUs (415V/400V, 60A and up), you convert to a low-voltage busbar with 80 PLUS Titanium CRPS power supplies or OCP power shelves, and you increasingly look at 800VDC distribution as racks scale past several hundred kW. Redundancy (N+1, 2N, or shelf-level BBUs) sits over all of it. This guide walks the “last 30 metres” of copper in order — do the kW math, pick the PDU, choose the conversion tier, then decide redundancy — so you can spec a rack that powers up cleanly and has headroom for the next GPU generation. Figures are indicative and as of 2026-08; confirm against your platform’s published draw before ordering.

Step 1: Do the kW math

Size to real draw with margin, not nameplate alone. Sum the per-node power (a GB200 NVL72 compute tray is roughly 5.4–7 kW), add switching and cooling overhead, then leave headroom for the next refresh. A rack landing near 132 kW is not an outlier; it is the current baseline for rack-scale accelerators.

Two constraints follow immediately. First, current per circuit: at 415V three-phase, a 60A feed carries roughly 34–43 kW, so a 130 kW rack needs several such feeds or a busbar-fed shelf. Second, phase balance: spreading GPU nodes evenly across phases avoids tripping one leg while others sit idle — a reason per-outlet metering is worth paying for.

Step 2: Choose the rack PDU

For GPU rows, the practical default is a three-phase 415V/400V rack PDU at 60A, giving ~34.6 kW at 415V or ~43.5 kW at 400V per unit. Match the intelligence tier to how you operate:

• Monitored — inlet-level reporting only; cheapest, for lights-out rows you rarely touch.
• Metered — per-outlet or per-circuit measurement, often billing-grade; essential for phase balancing and chargeback.
• Switched — remote per-outlet on/off, so you can power-cycle a hung node without a truck roll.

Tier-1 switched-and-metered 34.6 kW units list in the ~$3,700–$5,800 range; intelligent 415V units from other vendors sit around $2,700–$4,000. China factory-direct metered/bypass PDUs start far lower (tens of dollars) but lack the outlet-level control dense GPU racks usually want. Compare tiers on our power & PDU page against your metering and switching needs.

Step 3: Convert power — CRPS PSUs vs OCP power shelves

Below the PDU, AI hardware runs on a low-voltage DC rail, and there are two ways to make it.

CRPS titanium PSUs are the per-server route. Common Redundant Power Supply modules at 3,000–3,200 W, 80 PLUS Titanium, feed a 48/54VDC bus with hot-swap 1+1 redundancy and PMBus digital monitoring. Entry modules run ~$125 and certified 3,200 W units ~$256 factory-direct; 4,000 W AI-server PSUs reach ~$900 in channel.

OCP power shelves move conversion out of the servers and into the rack. A Delta 72 kW ORV3 shelf uses six 12 kW PSUs to turn 480VAC into 50/48VDC across a busbar; a 108 kW HVDC shelf uses six 18 kW units. Stackable designs (33 kW per 1U) reach 132 kW per rack. Shelves are RFQ-only, sold project-by-project, but they are the cleaner path at rack-scale density.

Step 4: Watch the 800VDC shift

The industry is moving from AC-plus-48VDC toward 800VDC (and other HVDC) distribution. As racks pass several hundred kW, higher distribution voltage cuts current and copper loss, and Solid-State Transformers now convert medium-voltage AC to 800VDC at high efficiency. This is reshaping the power-shelf SKU landscape in 2026, so verify the input voltage your platform expects before committing to shelves — a shelf built for 480VAC is not a drop-in for an 800VDC design.

Step 5: Build in redundancy

Redundancy is layered. At the PSU, 1+1 CRPS covers a module failure. At the rack, dual PDUs on independent feeds (2N) or shelf-level battery backup units (BBUs) ride through disturbances — Delta pairs its shelves with 72 kW BBU options. Rack-level online double-conversion UPS (10 kVA class Tier-1, or lower-cost LiFePO4 units) protects smaller rows. Decide the target — N+1 for cost, 2N for critical — before you fix the PDU and shelf count, since it changes both.

One volatility note: high-power PSU BOM cost is exposed to gallium and SiC supply, which has been volatile since 2024 export curbs. Lock specs and pricing at RFQ.

FAQ

How much power does an AI rack draw? A GB200 NVL72 cabinet draws roughly 120–140 kW, versus 5–15 kW for a conventional rack. Roadmaps point at 240–600 kW, with hyperscalers planning toward 1 MW racks.

What is the difference between metered, monitored and switched PDUs? Monitored reports at the inlet; metered adds per-outlet or per-circuit measurement, often billing-grade; switched adds remote per-outlet on/off. Dense GPU racks usually want metering plus switching.

Why 415V three-phase instead of single-phase? A 60A three-phase 415V feed carries roughly 34–43 kW versus a few kW single-phase — more power per circuit at manageable current, so fewer, safer copper runs.

What is an OCP power shelf? A rack-mounted unit converting AC (or HVDC) to a shared 48/50VDC busbar. ORV3 shelves reach 72–108 kW; stackable designs exceed 130 kW per rack.

Is 800VDC replacing 48VDC in AI racks? It is emerging, not universal. As racks pass several hundred kW, 800VDC cuts current and copper loss. The SKU landscape is shifting in 2026 — confirm your platform’s input voltage first.

Bottom line

Size the rack in kW first, feed it with three-phase 415V metered-and-switched PDUs, convert with titanium CRPS or OCP power shelves, watch the 800VDC transition, and layer redundancy to your uptime target. We broker PDUs, CRPS PSUs and power shelves direct from Taiwan and China factories. See our power & PDU category and request a quote with your rack density and redundancy target.