Data center rack DC power distribution and the move to higher-voltage HVDC

The rack DC power shift, and why it is still emerging

Higher-voltage DC rack power is the practice of distributing direct current at a few hundred volts or more to a busbar inside the rack, fed by a centralized power shelf, instead of running many alternating-current branch circuits to a strip and letting each server rectify its own AC. The change matters because AI racks have pushed past 100 kW, and at that load the traditional AC approach runs into too much copper, too many conversion stages, and too much wasted heat.

Be clear about where this stands. Most of the installed base still feeds the rack with AC, through a rack PDU and per-server power supplies, and that does not change overnight. The high-voltage DC move is emerging, driven by the hyperscalers and the Open Compute Project, and the specifications are still being written as of 2026. This guide explains the reasoning, the architecture, and the parts that are not nailed down.

Two sibling guides sit next to this one. The data center power distribution chain guide walks the AC path from the utility to the rack, and the AI and GPU rack readiness guide covers sizing the power, cooling, and floor for a high-density rack. This guide is the rack-level DC shift, so it points to both rather than repeating them.

Why does a 120 kW AI rack break the traditional AC approach?

A 120 kW rack breaks the traditional AC approach because the current it pulls gets too large to carry on a reasonable amount of copper, and every server rectifying its own AC stacks up conversion losses the rack can no longer absorb. A legacy general-purpose rack drew 5 to 12 kW on one or two branch circuits. A GB200 NVL72 class rack is specified near 120 kW and has been measured drawing 130 to 132 kW under load, and the announced next generation runs 180 to 220 kW on the path to the 1 MW rack the industry is now planning around.

Walk the current to see the problem. At 415 V three-phase, a 120 kW rack draws on the order of 167 A, which a heavy busway tap still handles. Push toward a 1 MW rack and the AC distribution gets impractical, and a rack-internal bus at the old 48 to 54 V would have to carry tens of thousands of amps for the same megawatt. The only sane way out is to raise the distribution voltage so the current, and the copper, come down.

The power-density numbers, the cooling that comes with them, and the readiness assessment all live in the AI and GPU rack readiness guide. The point here is narrower: the load has grown to where the power delivery method itself has to change, not just its rating.

AC distribution versus distributing DC to the rack

The traditional chain delivers AC almost all the way to the chip and converts it several times near the end. Utility power steps down through transformers, the UPS conditions it, a floor PDU and a rack PDU break it into AC branch circuits, and then each server power supply rectifies that AC to DC and steps it down again for the board. The full path is covered in the data center power distribution chain guide; the relevant part here is how many conversions sit at the bottom of it.

Distributing DC changes the bottom of the chain. A central power shelf rectifies the incoming AC once, at the rack or in a nearby power rack, and sends DC down a busbar to the servers, which only have to step it down rather than rectify it. Fewer conversions means less hardware repeating the same work, and the efficiency case is real, though the published figures are still vendor and project specific.

This is not a wholesale replacement of the AC plant. The facility still generates, conditions, and delivers AC to the row in nearly every design. What moves is the last conversion, from happening dozens of times inside the rack to happening once at a shelf. Treat it as a change to the final distribution leg, not a rebuild of the whole chain.

Where the conversion losses stack up

Every AC-to-DC and DC-to-DC conversion turns a few percent of the power into heat, and in the traditional rack those conversions repeat in every server. The UPS converts, the server power supply rectifies AC to an intermediate DC, and on-board regulators step that down again to the chip voltages. None of the steps is wasteful on its own. Multiplied across a rack full of servers and then across a hall, the stacked loss is a number worth chasing.

Cutting conversion stages is most of the efficiency argument for rack DC. The OCP Mt. Diablo work, the disaggregated AC-to-DC power approach co-authored by Microsoft, Meta, and Google, reports roughly a 3 percent end-to-end efficiency gain from converting once at a shelf rather than in every server. NVIDIA, promoting its 800 VDC direction, has quoted up to a 5 percent end-to-end improvement. Read both as manufacturer and consortium figures tied to a specific architecture, not as a guaranteed number for any build.

The honest framing is that the savings are real but bounded, and they compound with the copper savings rather than standing alone. A few percent on a multi-megawatt hall is meaningful money over its life, which is why the hyperscalers are pursuing it. Confirm the efficiency claim against the actual equipment and the actual operating point before you put it in a business case.

Why does higher voltage mean less copper?

Higher voltage means less copper because power is voltage times current, so for the same kilowatts a higher voltage carries a lower current, and the conductor is sized to the current. Double the voltage and you roughly halve the current for the same power. The resistive heating in the conductor, the I squared R loss, falls with the square of the current, so halving the current cuts that loss to about a quarter for the same wire.

This is the same physics the utility uses to move power across the country at high voltage, applied inside the rack. A rack-internal bus at 48 to 54 V was already a step up from the older 12 V for exactly this reason: lower current, less copper loss, and a smaller busbar. The push to 400 V and 800 V class is the next step of the same idea, taken because the megawatt rack makes even a 48 V bus carry an unworkable current.

The voltage progression worth carrying in your head is 415 V AC three-phase at the rack today, an emerging 400 V class DC, and an emerging 800 V class DC for the densest racks. Each step up trades a higher voltage and its insulation and safety demands for a lower current and far less copper. The exact voltages are set by the specification and the manufacturer, and they are still moving.

400V-class DC and the plus/minus 400V busbar

The 400 V class is the current concrete step toward rack DC, and the reference point is the OCP Mt. Diablo project, also called Diablo 400. It defines an AC-to-DC power rack where rectifier shelves take three-phase AC routed from AC PDUs and output plus and minus 400 V DC, which is then brought to the compute racks over high-voltage DC cables. The plus/minus 400 V scheme references each rail 400 V about a center point, which is 800 V rail to rail.

The choice of 400 V as the nominal rail was deliberate. The specification points to the electric-vehicle supply chain at that voltage, so the connectors, the semiconductors, and the manufacturing already exist at scale rather than having to be invented for data centers. That borrowed ecosystem is part of why 400 V class arrived first.

Hold this loosely. The Diablo 400 base specification was published at an early revision, on the order of version 0.5 in 2025, which is a pre-release draft, not a finished standard. The voltages, the connectors, and the mechanical interfaces can still shift before it settles. Design to the current published revision and the manufacturer's documents, and assume details will change.

800V DC for the highest-density racks

The 800 V class is the direction for the very densest racks, and the loudest voice behind it is NVIDIA, which has laid out an 800 VDC architecture aimed at 1 MW racks and beyond, with deployment framed around 2027. The argument is the same physics taken one step further: at 800 V the current for a megawatt rack drops to roughly 1,250 A, where the same megawatt on a 54 V rack bus would mean something like 18,000 to 20,000 A, which is not buildable as a rack busbar.

An ecosystem is forming around it rather than a single product. Texas Instruments, STMicroelectronics, Schneider Electric, Hitachi, and Flex have all announced 800 VDC power components or reference designs tied to the NVIDIA architecture, and the work has been presented through OCP. The plus/minus 400 V approach and the 800 V approach are related, since plus and minus 400 V about a reference is 800 V rail to rail, so read the marketing voltage carefully against the actual rail arrangement.

This is the part to hedge hardest. As of 2026 the 800 V class is an announced and demonstrated direction, not a shipping standard installed at scale. The voltages, the protection scheme, and the connectors are still being defined, and the rollout timeline is a manufacturer projection. Treat 800 VDC as an emerging architecture to plan toward and confirm, not a settled answer to specify today.

What is a power shelf?

A power shelf is a centralized rack unit, a rectifier shelf, that converts the facility AC feed into the DC the rack runs on, replacing the individual power supplies that used to sit in every server. In the current Open Rack v3 architecture the power shelf rectifies AC to a 48 V busbar and feeds the whole rack from there. In the higher-voltage DC architectures the same idea moves up to 400 V or 800 V class output.

The shelf is built from hot-swappable rectifier modules, usually run with one more module than the load strictly needs, so an N+1 arrangement keeps the rack up when a module fails or is pulled for service. Concentrating the rectification in one shelf is what lets the servers shed their own AC-to-DC supplies and take DC straight off the bus.

The shelf can live in the compute rack or in a separate power rack feeding several compute racks, which is the in-rack versus sidecar choice in the next section. Either way it is the single conversion point the whole rack DC approach is built around, so its rating, its redundancy, and its protection are where the design attention goes. The specifics are defined by the OCP specification and the manufacturer, and they vary by architecture.

Sidecar power and the separate power rack

Sidecar power puts the rectification in its own rack next to the compute racks instead of inside them. The sidecar takes facility AC, converts it to DC, and feeds one or more compute racks in the same row over a high-voltage DC busbar or cabling. The OCP Mt. Diablo work is a sidecar architecture, and NVIDIA demonstrated an 800 V sidecar at GTC to power a Rubin Ultra class rack.

The reason to move power out of the compute rack is space. Pulling the power shelves and per-server supplies out frees rack units for accelerators, and the Mt. Diablo work reports room for on the order of 35 percent more compute per rack as a result. When the rack is worth what an AI rack is worth, that recovered space is a real gain.

The trade is a row-level dependency and a high-current DC run between the power rack and the compute racks. That run is the part that has to be protected, supported, and made safe to service, and it is new territory for crews used to AC whips. Whether a given build uses an in-rack shelf or a sidecar depends on the architecture and the manufacturer, and both patterns are in play.

The rack DC busbar and the blind-mate

The rack DC busbar is the high-current conductor that runs the height of the rack and carries the shelf's DC output to every tray. Servers and trays connect to it through a blind-mate connector, so a node slides into its bay and lands on the bus without anyone running a cord. The bus is sized for the rack's full current, which at high power and low rack voltage is a very large number, and is the reason the architecture pushes the voltage up.

Open Rack v3 established the centralized power shelf and the common busbar pattern at 48 V, and the higher-voltage DC architectures keep the same shape with a higher-voltage bus. The blind-mate approach is what makes the rack serviceable: there are no whips to terminate at the node, and the connection geometry is defined so trays from different vendors land on the bus the same way.

A busbar carrying that much current is unforgiving about a poor connection. Every blind-mate joint is a place that has to seat fully and stay tight, because a high-resistance joint on a high-current bus makes heat exactly where it is hard to see. Torque, seating, and thermal scanning of the bus joints matter as much here as they do on overhead busway.

The copper and I-squared-R win

The copper savings are the second half of the case for rack DC, and they are mostly physics rather than marketing. Lower current for the same power means a smaller conductor, less copper by weight, and a lower resistive loss in the bus and the cabling. On a megawatt rack the difference between a 48 V bus and an 800 V bus is the difference between a busbar that cannot be built and one that can.

Weight and cost track the copper. A high-current low-voltage bus is heavy and expensive in metal, and it has to be supported by the rack structure that already carries dense, liquid-cooled hardware. Raising the voltage shrinks that conductor, which takes weight off the structure and metal cost out of the build, on top of the resistive loss it saves over the life of the site.

The win is real but it is not free. The higher voltage that saves the copper raises the insulation, clearance, and safety requirements, and it brings the DC arc-flash and protection issues covered below. The physics favors higher voltage; the engineering bill comes due in the protection and safety design, so weigh both, not just the copper.

Moving rectification out of the server

In the traditional rack, every server carries its own power supply that rectifies AC to DC, and a rack of forty servers carries forty of them, each converting independently. The rack DC approach moves that rectification into the shared power shelf, so the server no longer rectifies AC at all. It takes DC off the busbar and only steps it down to the board voltages it needs.

Pulling the conversion out of each server has a few effects at once. There are fewer power supplies to build, power, and fail, the shelf can be sized and made redundant as one unit rather than as dozens, and the rack space the supplies used to occupy is recovered for compute. The conversion still happens, but once, in a place built and maintained for it.

This is the change that ripples into the server design itself. A server built for this takes a DC input and expects the bus to be there, which is a different machine from one with an AC cord and an internal supply. That coupling between the rack power architecture and the server is part of why the standardization work matters, and part of why the move is happening at the hyperscale and OCP level first.

Is DC power more dangerous than AC?

DC is not automatically more dangerous than AC, but a DC fault is harder to interrupt, and that difference is the safety story crews have to absorb. Alternating current crosses zero twice every cycle, and an arc tends to self-extinguish at that zero crossing, which is part of how an AC breaker clears a fault. Direct current has no zero crossing. A DC arc is continuous, it does not get the natural moment to go out, and it carries higher continuous incident energy as a result.

That makes a DC arc both harder to clear and potentially more punishing if it is not cleared. A breaker designed for AC may fail to extinguish a DC arc at all, because it was counting on the zero crossing the DC fault never gives it. The hazard is not exotic, but it is different from what an AC-trained crew has internalized, and the higher voltages in play raise the stakes.

Treat the higher-voltage DC bus as a serious hazard and work it under a DC-specific safety basis, not by AC habit. Verify de-energization, follow the manufacturer's procedures for the specific architecture, and do not assume AC-rated tools, gloves, or breakers carry over. This is the area where the field most needs to retrain, and it is blunt for a reason.

DC protection, breakers, and fault current

Protection is where DC distribution diverges most from AC practice. Because a DC arc does not self-extinguish at a zero crossing, the protective devices have to force the arc out using magnetic or thermal means, so DC-rated breakers and fuses are not the same parts as their AC counterparts and cannot be swapped in. A device rated for an AC fault may not clear the equivalent DC fault.

The fault current on a high-current DC bus is large and fast, and the available fault current has to be calculated for the DC architecture rather than carried over from an AC study. Isolation, fusing strategy, and the coordination between the shelf, the bus, and the trays all have to be designed for DC behavior. This is engineering work tied to the specific equipment, not a field judgment call.

There is also a standards gap to be honest about. The familiar incident-energy and arc-flash methods, including those in NFPA 70E, were developed largely for AC, and standardized DC clearing curves, PPE categories, and arc-flash boundaries for these architectures are still maturing. Until they settle, the protection and safety basis comes from the manufacturer and the engineer of record for the specific system, and should be confirmed, not assumed.

DC grounding and the plus/minus reference

Grounding a DC system is its own design decision, and the plus/minus 400 V scheme is a clue to how it is approached. Referencing the bus as plus and minus 400 V about a center point keeps each conductor 400 V from the reference while delivering 800 V across the pair, which holds the touch voltage to either rail lower than a single 800 V conductor to ground would. The reference can be solidly grounded, resistance grounded, or left floating with detection, and each choice changes the fault behavior and the touch-safety story.

The grounding choice drives what happens on the first fault. A solidly grounded reference clears a ground fault directly; a floating or high-resistance reference rides through a first ground fault but has to detect it and alarm, because a second fault then becomes a hazard. Which scheme a given architecture uses is set by the specification and the manufacturer, and it has to match the protection design rather than being decided in isolation.

The field lesson is that DC grounding is not a copy of AC grounding, and the reference is part of the safety case, not an afterthought. Confirm the grounding scheme for the specific architecture, confirm how a first ground fault is detected and handled, and confirm the touch-safety basis before anyone works on a live bus.

Redundancy at the power shelf

Redundancy in a rack DC architecture concentrates at the power shelf, because the shelf is now the single conversion point for the whole rack. The common approach is N+1 within the shelf, running one more rectifier module than the load needs so any module can fail or be pulled without dropping the rack. On more demanding designs the redundancy extends to two shelves or two power feeds, the same 2N idea the rest of the critical-power chain uses.

Concentrating the conversion is efficient but it changes the failure surface. In the old rack, a failed server power supply took down one server. In the new rack, the shelf feeds everything, so its redundancy has to be real and its modules have to be serviceable live. The A and B feed independence that runs through the whole power chain still applies, and it has to reach the shelf, not stop short of it.

How the redundancy is arranged, N+1 at the module level, dual shelves, or dual feeds, is set by the architecture and the project basis of design. The redundancy notations and the upstream A and B path are covered in the data center power distribution chain guide. The point specific to rack DC is that the shelf is the new place the redundancy has to hold.

The rack battery backup unit

A battery backup unit, the BBU, is a rack-level battery that rides the load through a power disturbance and the transfer to backup generation, placed in the rack rather than in a central UPS room. Open Rack v3 defines a BBU shelf that sits alongside the power shelf and feeds the same busbar. The published 48 V BBU shelf, for example, houses several modules in a 5+1 redundant arrangement and holds the bus up during the transfer between sources, with the busbar specified to stay above about 46 V through the dip.

Putting the battery at the rack shortens the ride-through path and ties it directly to the bus the load actually runs on. The energy is lithium chemistry at the rack, which brings its own placement, thermal, and fire-protection considerations that belong with the rack and the room design, not glossed over. The role is the same as a central UPS battery: bridge the gap until the generator carries the site.

Whether a build uses rack BBUs, a central UPS, or both depends on the architecture and the resilience target, and the higher-voltage DC architectures are still defining how rack-level energy storage fits at 400 and 800 V class. The ride-through sizing and the redundancy are project decisions. Confirm them against the basis of design and the specification rather than assuming the 48 V numbers carry up to the higher-voltage buses.

OCP, Open Rack, and the standards picture

The standardization is happening mostly inside the Open Compute Project, and naming the pieces helps keep them straight. Open Rack v3 is the established framework that defines the centralized power shelf, the BBU shelf, and the common busbar at 48 V. Mt. Diablo, also called Diablo 400, is the disaggregated AC-to-DC sidecar work co-authored by Microsoft, Meta, and Google that takes it to plus and minus 400 V. NVIDIA's 800 VDC architecture is the direction for 1 MW racks, with components and reference designs contributed by several vendors and presented through OCP.

The reason the work is open is interoperability. The hyperscalers and the silicon and power vendors are aligning on voltage ranges, connector interfaces, and safety practices so that racks, shelves, and servers from different sources fit together, which is what lets the ecosystem move at all. The whitepapers and specifications coming out of the OCP summits are where the current direction is documented.

Be honest about maturity. The Open Rack v3 48 V architecture is shipping and real. The 400 V class specification is at an early, pre-release revision. The 800 V class is an announced and demonstrated direction with an ecosystem forming around it, not a finished standard installed at scale. The whole area is evolving cycle to cycle, so confirm the current revision of any specification before you build to it.

The facility still delivers AC to the row

For all the talk of DC, the facility upstream of the rack is still an AC plant in nearly every design. The utility brings in medium voltage, transformers step it down, the UPS conditions it, and AC distribution carries it across the hall to the row. That path is unchanged by the rack DC shift and is covered end to end in the data center power distribution chain guide.

What changes is where the AC-to-DC conversion happens. Instead of each server doing it, a power shelf or a sidecar power rack does it once at the row or the rack, taking the facility AC feed and producing the DC bus. The facility delivers AC to the doorstep of the rack; the rack DC architecture takes over from there.

This matters for how you scope a project. The medium-voltage gear, the generators, the UPS plant, and the busway to the row are the same conversation as a conventional high-density build. The rack DC architecture is a change to the final leg, so design the upstream chain as you would for any AI hall and confirm where the AC hands off to DC.

This is for new halls, not a legacy retrofit

Rack DC architecture is a greenfield play. It assumes racks, servers, power shelves, and a bus all designed together for DC, plus a row layout and protection scheme built around it. You do not drop a 400 V or 800 V DC rack into a legacy AC hall and wire it to the existing rack PDUs, because almost nothing carries over: the servers are different, the distribution is different, and the protection and safety basis are different.

The realistic path is the same one the AI and GPU rack readiness guide lays out for high density generally. New high-density halls get designed for the architecture from the start, while legacy halls host AI at lower density on conventional AC distribution, or get rebuilt rather than retrofitted. The rack DC shift sits at the leading edge of that, where the density and the new-build commitment are both highest.

So the decision is not whether to convert an existing room to DC. It is whether a new high-density build adopts a rack DC architecture, and which one, knowing the standards are still forming. Most operators today are still building AC halls and watching the DC work mature. Decide deliberately and confirm the architecture is ready for the timeline you are building to.

The open questions and what is not settled

The honest summary is that the higher-voltage DC rack is promising and early, and several things are not settled. The 400 V class specification is at a pre-release revision. The 800 V class is an announced direction with components and demonstrations but not a deployed standard. The DC arc-flash, protection, and safety standards are still maturing, with the familiar AC methods not yet fully translated to these voltages and architectures.

The ecosystem is also still assembling. Connectors, the exact voltages, the grounding schemes, the rack-level energy storage, and the cross-vendor interoperability are all being worked out in the OCP process, and details will change between revisions. A design committed too early to a moving specification can end up orphaned when the spec settles somewhere else.

None of this means the direction is wrong. The physics behind it, lower current and less copper at higher voltage, is sound, and the hyperscalers are committing real money to it. It means the field should treat rack DC as an emerging architecture: worth understanding now, worth planning toward for new high-density builds, and to be confirmed against the current specification and the manufacturer at the moment of design, not taken as a settled answer. Most of the industry is still AC, and that is a reasonable place to be while this matures.

Commissioning rack DC power

Commissioning a rack DC system proves the shelf, the bus, and the protection before live load arrives, and it leans harder on the manufacturer's procedures than a conventional AC fit-out does, because the equipment and the hazards are new. The work runs from checking each rectifier module and the shelf redundancy, through verifying the busbar connections and the blind-mate seating, to proving the DC protection and the grounding scheme behave as designed under fault.

Two things deserve extra attention because they are different from AC commissioning. First, the DC protection and isolation: confirm the breakers and fuses are DC-rated, that they clear the calculated DC fault, and that the grounding and first-fault detection work as specified. Second, the shelf redundancy and the BBU transfer: pull a module under load, fail a feed, and confirm the bus holds through the transfer to backup as the specification claims.

Run the safety case as part of commissioning, not after it. Verify lockout and de-energization actually de-energize the bus, confirm the touch-safety basis, and make sure the crew that will service the system has been trained on the DC-specific hazards. The acceptance criteria come from the manufacturer and the project specification for the specific architecture, so commission against those documents rather than a generic AC checklist.

What to document

The record for a rack DC system has to capture the architecture, because nothing about it is the default a later crew would assume. The next person to work on the rack needs to know it is DC, at what voltage, with what protection and grounding, before they touch it. Capture the power architecture, the shelf and BBU configuration, the bus voltage and protection, and the as-built condition, keyed to the rack and the row.

FieldOS, or whatever field record system the operator runs, is where this belongs, tied to the rack position so the power architecture travels with the cabinet. The integrity rule is the same as the rest of the chain: a record that cannot be found later is a record that cannot be trusted when something fails.

Common mistakes

• Applying AC habits and assumptions to a DC system, from the work practices to the device ratings.
• Ignoring DC arc flash and the fact that a DC fault is harder to interrupt than an AC fault.
• Fitting AC-rated breakers or fuses where DC-rated protection is required, so a fault may not clear.
• Undersizing the busbar or the high-current DC run for the current at the chosen voltage.
• Building with no power-shelf redundancy, so a single conversion point can drop the whole rack.
• Trying to retrofit a legacy AC hall to rack DC instead of treating it as a new-build architecture.
• Treating the emerging 400 V and 800 V specifications as settled when they are still at draft or announced revisions.

Field checklist

Standards and references

The Open Compute Project is where the rack DC architecture is being standardized. Open Rack v3 defines the centralized power shelf, the BBU shelf, and the 48 V busbar that the higher-voltage work builds on. The Mt. Diablo, or Diablo 400, specification covers the disaggregated AC-to-DC sidecar at plus and minus 400 V, and was published at an early pre-release revision, so confirm the current version before building to it. NVIDIA's 800 VDC architecture and the vendor reference designs around it document the 800 V class direction, which is emerging rather than deployed at scale.

On the installation and safety side, the NEC, NFPA 70, governs the wiring, overcurrent protection, and grounding, including its provisions for DC systems, and the figures and article numbers shift between code cycles, so confirm them against the adopted edition and local amendments. NFPA 70E covers electrical safety and arc flash, with the caveat that its incident-energy methods were developed largely for AC and the DC equivalents are still maturing. IEC standards cover DC distribution and equipment internationally, and IEEE work covers protection and fault behavior. NETA gives the acceptance and maintenance testing basis for the gear.

Above all of these sits the manufacturer's documentation for the specific architecture and the engineer of record's design. Where a standard, a manufacturer, and a project specification disagree, the stricter and more specific one governs, and none of them replaces the calculations for the actual equipment. Because this area is evolving, verify the current edition of every reference before citing it, and do not treat an emerging architecture as a settled standard.

Units, terms, and abbreviations

The rack DC discussion mixes power, voltage, and architecture terms that get used loosely, and the same words mean different things between the AC plant and the rack. Pin the term to the layer before acting on it.

HVDC / rack DC power

Distributing direct current at higher voltage, here 400 V or 800 V class, to a rack busbar rather than AC to per-server supplies

Power shelf

Centralized rectifier shelf that converts facility AC to the rack DC bus, replacing per-server power supplies, usually N+1

Busbar

High-current conductor running the height of the rack that carries the shelf's DC output to the trays via blind-mate connectors

Sidecar power

A separate power rack that rectifies AC to DC and feeds one or more compute racks over a DC bus, freeing rack space for compute

Conversion stage

Any AC-to-DC or DC-to-DC step; each loses some power as heat, so fewer stages raise efficiency

Plus/minus 400 V

A DC bus referenced 400 V each side of a center point, 800 V rail to rail, which lowers the touch voltage to either rail

DC arc flash

An arc on a DC system; with no zero crossing it does not self-extinguish, so it is harder to interrupt and carries higher continuous energy

BBU

Battery backup unit, a rack-level battery that rides the load through a disturbance and the transfer to backup generation

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• How does power get from the utility to a server?
• How much power does an AI rack use?
• What is a rack BBU?
• What is a rack PDU?
• What is a rack unit (U)?