Back to News & Events A new GPU tray passes every bring-up test at idle. Under a real training job, the load jumps from a fraction of its rating to full power in milliseconds, the 48V bus sags, and the node resets. The specification said N+1, Titanium efficiency, and 48V input. None of those words described the load step.
That gap between a requirements list and the behavior of the load is where most AI platform power problems begin. This article walks through data center power requirements for AI racks, from the shelf that converts AC to the 48V DC-DC converters feeding each board, and shows how OEM engineering and procurement teams can specify power supplies for data centers so the first training run is not the first real test.
The Rack Outgrew Its Feed
Industry reporting puts today’s AI racks at roughly 150 kW, with next-generation designs aimed at 400 kW and, on some roadmaps, 1 MW. Conventional AC distribution does not scale to those levels without an unwieldy number of feeds and cables. That is why AI data center power requirements are being rewritten at the rack, not at the facility.
Density changes every requirement downstream. Conversion losses become heat measured in kilowatts per rack. Busbar cross-section and connector current ratings turn into design constraints. Redundancy has to be sized for modules that each carry a meaningful share of the rack. We cover the efficiency economics in our piece on high-efficiency power supplies for data center infrastructure. Here the focus is on what each converter in the chain has to do.
A typical AI rack power chain runs like this:
AC feed → rack power shelf (AC-DC, 48 to 54 V) → busbar → tray-level 48V DC-DC converters → point-of-load stages → GPU core rail
Every arrow after the AC feed is a conversion stage with its own losses. Three stages at 97% each deliver about 91% end to end. Efficiency has to be specified per stage, because a single headline number hides where the heat comes from.
What the Power Chain Has to Survive
Efficiency at the load points that matter. Titanium-grade shelves reach 96% or better near mid-load, but an AI rack spends its life wherever the job queue puts it. A requirement written as one peak number hides the curve. Specify efficiency at 20, 50, and 100% load, since at 100 kW of IT load every percentage point of conversion loss is roughly a kilowatt of heat to remove.
Redundancy that survives a service event. Shelf-level N+1 (3+1 and 6+1 are common) or N+N where uptime demands it, built from hot swappable power supplies with active current sharing and fault decoupling. For server rack power supplies, the test that matters is pulling a module at full load without a bus sag, and a shorted module not dragging the rail down with it.
Load-step response and hold-up. GPUs can move from idle to peak within milliseconds, and training jobs repeat that swing continuously. Specify step amplitude, slew rate, permitted droop, and repetition rate, plus hold-up time in milliseconds. “Fast transient response” describes an intention, and a test plan needs a number.
Thermal conditions as installed. Inlet temperature, chassis airflow, and the cold-plate or liquid-cooling interface all change what a converter can deliver. Derating data taken on an open bench says little about a tray with GPUs exhausting beside the converter.
Compliance and telemetry. IEC/UL 62368-1 for safety, FCC Part 15 Class A and CISPR 32 for emissions, and IEC 61000-3-2 for harmonics on AC-input units. PMBus telemetry belongs on this list too. Fleet software that can see current, temperature, and fault flags can schedule a swap before the fault becomes an outage.
None of this is unique to server halls. Efficiency, transient response, redundancy, and thermal limits colliding inside one enclosure is the compound-requirement problem behind every program we see.
The same discipline applies across custom power supply solutions for industrial applications, and AI raises the density and speeds up the pace.
48V DC-DC Converters: The Stage Most Often Underspecified
Start with why 48V. At the same power, a 48V bus carries a quarter of the current of a 12V rail, and resistive loss in busbars and connectors scales with the square of current, so it falls by a factor of 16. 48V also stays inside extra-low-voltage limits, which keeps serviceable hardware simple. That is why it has been a telecom standard for decades.
The price is another conversion stage near the load, and that stage carries a large share of the efficiency and density burden. For 48V DC-DC converters in an AI tray, the requirements that decide the design are efficiency across the whole load range, an input window that tolerates bus excursions between 48 and 54 V, inrush and hot-plug behavior that does not disturb neighboring trays, transient response matched to the point-of-load stages downstream, and a thermal path into the tray’s cold plate or airflow. Whether the stage needs isolation depends on where the architecture places the safety barrier, so the RFQ should state it and not leave it to assumption.
The 800V question deserves a line in every platform roadmap review. Reported proposals move the highest-density racks to 800V DC, Pilots are expected to start as sidecar racks within about a year, and the expectation is that 800V serves only the highest-density workloads over the next three to five years, not all data center power. For the converter stage, the practical step is to ask suppliers for a roadmap, so a design that works at 48V today can follow the platform. Our DC-DC power converters engineering and procurement guide covers topologies and sourcing fundamentals.
Server Rack Power Supplies: Catalog, Semi-Custom, or Custom
Catalog server rack power supplies fit when the rack format is fixed (CRPS or Open Rack style), volumes are standard, and the OEM accepts the vendor’s lifecycle terms. Plenty of platforms never need more than that, and our guide to off-the-shelf versus custom power supplies explains when that holds.
The decision for server rack power supplies moves toward semi-custom when the tray envelope, rail set, telemetry map, input range, or cooling interface does not match a standard module. Semi-custom adapts a proven platform, so NRE is a fraction of a ground-up development and the schedule runs in months. Full custom is for architectures no platform reaches: liquid-cooled integrated shelves, 800V-ready designs, unusual envelopes.
Most hyperscalers design their own power. The buyers for whom this choice is live are server and accelerator OEMs, system integrators, rack-level builders, and edge or telecom equipment makers, who need purpose-built performance without a hyperscaler’s program cost. That is the audience for custom and semi-custom power supplies for data centers.
AI Data Center Power Procurement: What the RFQ and NRE Proposal Must Pin Down
Platform cycles run a year or two, ramps are steep, and a power supply that misses its qualification date holds up a whole rack shipment. Procurement here is a program decision more than a price comparison.
An RFQ that returns accurate quotes states the power requirements for data centers at three levels. At the input: AC voltage, phase, and frequency, or the DC bus range. At the load: per-shelf or per-tray power, continuous and peak, plus step amplitude, slew rate, permitted droop, and hold-up in milliseconds. At the system: efficiency at stated load points and input voltages, redundancy scheme and hot-swap requirement, inlet temperature and airflow or liquid-cooling interface, form factor or standard, the PMBus telemetry map, the standards list, and program data covering prototype quantities, ramp profile, annual volume, and platform lifetime.
The NRE proposal that comes back should itemize load-step characterization against your actual load profile, efficiency measurement at your load points, fault-injection testing of redundancy and hot-swap, chassis thermal validation, EMC pre-compliance, and the documentation package, with named milestones. Three red flags are worth knowing: “Titanium” quoted with no load points, “fast transient response” with no step amplitude, and no capacity commitment for the ramp.
Lifecycle terms close the loop: configuration control so qualified and production units stay identical, PCN notice before any change affecting form, fit, function, or compliance, a qualified second source or drop-in replacement path, and obsolescence management over the platform’s life. Teams that treat server rack power supplies as a commodity line item and skip these tend to repeat the common RFQ mistakes that delay power supply programs.
FAQ: Data Center Power for AI Racks
What are the main data center power requirements for AI workloads? Six things: rack density rising from about 150 kW toward 400 kW and beyond, efficiency measured at stated load points, redundancy with hot-swap serviceability, load-step response and hold-up sized to GPU behavior, thermal performance as installed, and compliance plus telemetry. Written as numbers (step amplitude, droop, milliseconds, load points) they become testable requirements. Written as adjectives they become disputes at acceptance.
Why do AI racks distribute power at 48V instead of 12V? At equal power, 48V carries a quarter of the current, and resistive loss in busbars and connectors scales with current squared, so it drops by a factor of 16. 48V also stays within extra-low-voltage limits, which simplifies serviceable hardware. The cost is an extra conversion stage near the load, which is why the efficiency and transient response of 48V DC-DC converters carry so much weight.
Will 800V DC replace 48V in AI data centers? Not across the board, on current reporting. 800V DC is proposed for the highest-density racks, with pilots expected as sidecar deployments and broader use for top-end AI workloads over three to five years. Most 48V platforms will keep shipping for years. Ask converter suppliers for an 800V roadmap so today’s design does not become a dead end.
How should server rack power supplies be specified for redundancy? State the scheme (N+1, N+N, or shelf-level 3+1 and 6+1), the hot-swap requirement, and the failure cases to survive: one module removed under full load, a shorted module, and loss of one input feed. Require current sharing and fault decoupling, and ask for fault-injection test evidence instead of a design-intent statement.
When does an OEM need semi-custom instead of catalog? When the tray envelope, rail set, input range, telemetry map, or cooling interface does not match a standard module, or when volume and platform life justify configuration control. Semi-custom adapts a proven platform at a fraction of full-custom NRE, and full custom is reserved for architectures no platform reaches. If a standard module meets every requirement with margin, catalog is the right answer.
What should procurement include beyond unit price? Capacity commitments for the volume ramp, an itemized NRE proposal with a test plan and milestones, configuration control and PCN notice periods, a qualified second source or drop-in replacement path, and obsolescence management across the platform’s life. With platform cycles of a year or two, a late power supply holds up the whole rack shipment, so schedule risk belongs in the evaluation next to price.
Power for an AI platform is easier to get right when the load profile is on the table from the first conversation. Horizon PSS, a custom power supply manufacturer working with OEMs on custom and semi-custom power, starts each program with a feasibility review of the rack architecture and load steps, and recommends catalog when catalog is enough. If your platform’s data center power requirements have moved beyond a standard module, send the load profile, bus architecture, and thermal conditions at whatever maturity they exist, or a full RFQ if you have one.
