800 VDC Power Distribution: A Full Guide to How AI Data Centers Leave 48-Volt Behind

800 VDC is not a new way to make electricity, and it is not a cooling loop. It is not the silicon carbide MOSFET or the gallium nitride HEMT inside a power supply. It is the electrical-distribution architecture: convert medium-voltage AC once, at the perimeter of the hall, into roughly 800 volts of DC, run that DC down a busway instead of a 415/480 VAC feeder, and convert once more at the tray into the 54 V, 12 V, or 6 V the GPU can actually use. The product the operator buys is still a rectifier, a sidecar, a busway, a DC breaker, a UPS, or a rack-level DC/DC shelf. The thing that changed is the voltage the hall is wired at.

That split exists because the alternative ran out of copper. For twenty years, server power was a 12-volt, then a 48- or 54-volt, problem solved with in-rack CRPS or Open Rack v3 shelves, a 415 or 480 VAC row, and a transformer-and-UPS room at the other end of the feeder. At 10 kW a rack, that stack is boring and cheap. At 100–120 kW it is a busbar with a cooling problem. At a megawatt it is a physics argument NVIDIA has already named: 800 VDC, starting in 2027 with Kyber-class systems, not as a 2026 bill of materials.

What changed in 2026 is not that high-voltage DC was invented. Telecom has run −48 VDC for decades; 380 VDC data-center pilots have come and gone; the EV industry industrialized 800-volt traction inverters and made the connector, the breaker, and the isolation practice that data-center vendors are now borrowing. What changed is that AI racks crossed the current density where a 54-volt busbar stops being a product and starts being a foundry for copper, that NVIDIA, Google, and Microsoft put an 800 VDC architecture through the Open Compute Project, and that the 2026 volume path is still a 50-volt bus fed by 5.5–12 kW shelves. The useful question is which of those pieces is a shipping sidecar, which is a GTC demo, and which is still a 2028 facility drawing.

This is the dedicated electrical-distribution guide. The transistors that make 800-volt conversion efficient live in the SiC and GaN power electronics full guide. Generation — including the small modular reactors hyperscalers keep announcing — is a different object. Heat is a liquid-cooling object. We do not re-derive them here.

Key takeaways

• 800 VDC is distribution, not generation, not heat removal, and not the FET. A reactor, a gas turbine, or a utility interconnect makes the megawatts. A CDU takes the heat back out of the GPU and the converter. SiC and GaN are the switches inside the rectifier and the tray converter. This guide is the bus: facility medium voltage in, 800 VDC down the hall, one late-stage conversion at the compute. Backside power delivery is a fourth, unrelated conversation — millivolt grids on a GAAFET die, not kilovolt grids in a rack.
• 48 V and 54 V did not fail. They ran out of copper at 100–200 kW, and they become geometrically illegal at a megawatt. Power is voltage times current. A 120 kW GB200-class rack at 54 V is about 2,200 A. A 1 MW rack at 54 V is about 18,500 A. NVIDIA’s own 800 VDC architecture note puts the busbar in that 1 MW rack at up to 200 kg of copper — 200 tonnes of rack busbar copper in a 1 GW campus — and says a 54 V Kyber would consume up to 64 U of power-shelf space, leaving no room for compute. At 800 V the same 1 MW is 1,250 A.
• NVIDIA’s architecture is a facility argument, dated to 2027, with a 2026 hybrid on-ramp. Convert 13.8 kV AC toward 800 VDC at the perimeter with industrial rectifiers (or, later, a solid-state transformer). Distribute 800 VDC on two conductors. Convert once at the tray. NVIDIA quotes up to 5% end-to-end efficiency versus today’s 54 V path, 85% more power through the same conductor versus 415 VAC, and 45% less copper. Treat those as NVIDIA’s. Treat the date as a product-roadmap date: native 800 VDC and 1 MW IT racks coincide with Kyber in 2027. The 2026 object is an MGX-compatible 800 VDC power rack that slots into existing AC.
• Three layers will coexist, and the sidecar is the 2026–2027 one. NVIDIA’s August 2026 DSX language, Vertiv’s PowerDirect 5000, and Schneider’s own FAQ all describe the same sequence: a rack-adjacent sidecar that keeps the building on AC (H2 2026 commercialization, 2027 deployments); a row power center on an overhead 800 VDC busway, up to 2 MW per row, expected 2027; a facility DC power block that converts MV AC to 800 VDC once, for greenfield AI factories later. Vertiv’s Scott Armul, May 2026: “DC power in the rack moving to DC power in the pod.”
• Who actually ships is switchgear vendors and PSU OEMs, not a GPU press release. Vertiv: PowerDirect 5000 sidecar, 400–900 kW, portfolio commercial in H2 2026, customer pilots and ramp in 2027. Eaton: 2 MW, 15 kV-class medium-voltage solid-state transformer rated >97% at 800 VDC output, plus the Beam Rubin DSX grid-to-chip platform. Schneider Electric: Galaxy VXL UPS at 400–1,250 kW and a sidecar retrofit path they themselves cap at about 200 kW per rack. Delta: a 660 kW in-row 800 VDC rack with 480 kW of embedded battery backup, shown at GTC 2026. Those are real products and demos. They are not a 2026 GB200 hall.
• The unsolved problem is not the converter efficiency number. It is DC protection, busway and connector ratings, arc-flash procedure, a technician who can service an 800-volt rack alive, energy storage that actually damps GPU step-loads, and a multi-vendor stack that has been qualified together. NVIDIA’s own architecture post flags overcurrent protection as a remaining innovation item. A 98% PSU does not write that procedure.

800 VDC at a glance

What 800 VDC actually is

Start with the thing it is not. It is not an SMR. A reactor makes megawatts at 13.8 kV or whatever the interconnect is; this architecture does not. It is not direct-to-chip liquid cooling. A cold plate removes heat after the GPU and the converter have already turned watts into waste. It is not SiC or GaN. Those are the switches inside the boxes this guide is about. It is not backside power. PowerVia moves a millivolt-class grid to the back of a logic die. An 800-volt busway in a hot aisle is a different object, in a different package, on a different voltage, bought by a different team.

An electrical-distribution architecture is a voltage, a conversion schedule, and a protection scheme. Today’s AI hall takes utility medium voltage (often 13.8 kV in North America, 10–35 kV elsewhere), steps it down to 415 or 480 VAC, conditions it through an AC UPS, runs three-phase AC down a busway or a set of PDUs, and converts to ~50–54 VDC in power shelves inside the rack. NVIDIA’s GB200 NVL72 user guide is the existence proof at approximately 120 kW: eight shelves, six air-cooled 5.5 kW PSUs per shelf, a nominal 50–51 V DC bus, N+N. That is a 48/54 V world. It is the 2026 volume path.

800 VDC deletes stages. The target end-state, in NVIDIA’s May 2025 architecture blog and the October 2025 ecosystem follow-up, is: convert 13.8 kV AC to 800 VDC once, at the data-center perimeter, with industrial-grade rectifiers; distribute 800 VDC through the hall and the row; accept two-conductor 800 V at the IT rack; convert DC-to-DC in the compute rack down to 54 V or 12 V and then to core voltage. The Kyber-specific version puts a late-stage, high-ratio 64:1 LLC converter next to the GPU and steps 800 V to 12 V in one stage, occupying 26% less area than the multi-stage path, per NVIDIA. Three output conventions — 50 V, 12 V, and 6 V — will coexist on the tray; that fight lives in the device guide. This one stops at the bus.

±400 VDC is a topology fork, not a different architecture. Vertiv, Infineon, and ST all say they can serve a bipolar ±400 V bus with 650 V devices more comfortably than a monopolar 800 V bus. It is the same copper argument with a different isolation and grounding drawing. Do not treat it as a rival standard until a hyperscaler PO says so.

Why 48 V and 54 V ran out of copper

The congestion is not a slogan. It is current, busbar cross-section, and a 42U hole.

Current. P = V × I. Raise voltage, drop current, drop I²R, drop copper. The integers:

A 2,200 A 54 V busbar at 120 kW is a product. A 7,400 A busbar at 400 kW is a foundry. Schneider Electric’s own AI-factory FAQ is blunt: above 400 kW, traditional 54 VDC “becomes physically impossible.” Vertiv puts the break between 350 and 400 kW, where “connector sizes, busbars, and the sheer volume of copper needed to carry that much current start to run out of room.” NVIDIA’s number is the one to tattoo: a single 1 MW rack on 54 VDC requires up to 200 kg of copper busbar. Scale that by a gigawatt of IT and the rack busbars alone are 200,000 kg. That is not a sustainability slide. It is a copper-mine and a crane.

Volume inside the rack. GB200 NVL72 and GB300 NVL72 already use up to eight power shelves. NVIDIA’s architecture post is explicit about the next step: the same 54 VDC distribution at Kyber megawatt scale would consume up to 64 U of rack space — the entire cabinet — leaving no room for compute. The alternative shown at GTC 2025 was an 800 V sidecar feeding 576 Rubin Ultra GPUs in one Kyber rack, or a dedicated power-supply rack for every compute rack. Sidecar versus sister-rack is a floorplan argument. Both are admissions that the PSU no longer fits next to the GPU.

Conversion count. Today’s path is MV AC → LV AC (transformer) → UPS AC → row AC → rack AC/DC → tray DC/DC → core. Every stage has a loss, a fan, a failure mode, and a maintenance ticket. NVIDIA’s claim is that converting once at the perimeter and once at the tray cuts that chain, improves end-to-end efficiency by up to 5% versus 54 V, and (in the same post) cuts TCO by up to 30% through efficiency, reliability, and architecture. The 5% is the number the industry has repeated; the 30% is NVIDIA’s bundled TCO framing. Use the first as a physics-scale claim. Use the second as a vendor claim.

Why 48/54 V still won the last decade. Open Compute and the 48 V rack were themselves a copper argument against 12 V. They worked. A 120 kW NVL72 is the proof. The wide-bandgap content in those 5.5–12 kW shelves is how you keep 50 V legal at 120 kW; it is not how you keep 50 V legal at 1 MW. Mixing those two sentences is how a 2026 hall gets specified for the wrong busway.

415/480 VAC, 48/54 V, and 800 VDC in one table

Two comparisons get sold as one. Facility AC versus facility DC is one. Rack 54 V versus rack 800 V is the other. They stack.

NVIDIA’s row-level claim is the AC-versus-DC sentence: switching distribution from 415 VAC to 800 VDC “enables 85% more power to be transmitted through the same conductor size” because higher voltage drops current, and DC drops skin effect and reactive power. Copper requirement “by 45%.” Schneider repeats the 45% copper and 5% efficiency figures on its AI-factory power page. Those are the same ecosystem numbers. They are not a third-party audit of a built hall.

The 54 V path is not “AC.” It is DC at the wrong voltage. A 415 VAC hall feeding 54 V shelves is the 2026 default. An 800 VDC hall feeding 800 V trays is the 2027 target. A 415 VAC hall feeding an 800 V sidecar feeding 800 V trays is the bridge. Three drawings. One press cycle.

From medium voltage to the rack

Follow the current, because the slides will not.

Today (shipping). Utility MV (NVIDIA’s example: 13.8 kV; NVIDIA’s later ecosystem post also writes 35 kV-class) lands on a substation transformer. Low-voltage AC, typically 415 V in IEC land and 480 V in much of the US, feeds an AC UPS — Schneider’s Galaxy VXL is the AI-density SKU, 400–1,250 kW in 1.2 m², validated at 125% overload for ten minutes and 150% for one minute against real hyperscale AI load profiles at 40 °C. UPS output rides a busway or a set of PDUs down the row. At the rack, CRPS or ORv3 shelves convert to ~50–54 V. Inside the tray, an intermediate bus converter and a swarm of point-of-load bucks make core, HBM, and NIC rails. Battery backup, if it is at the rack, is an ORv3-style BBU on the 48/54 V bus.

Bridge (H2 2026 commercialization, 2027 deployments). Keep the building on AC. Put a three-phase AC-to-800 VDC conversion in a sidecar or an in-row power rack next to the compute. Vertiv’s PowerDirect 5000 is the named object: 400 kW to 900 kW of 800 VDC into the GPU rack over a busbar, returning the 8–16 U the power shelves used to occupy. NVIDIA’s August 2026 DSX blog calls the same object an “MGX-compatible 800 VDC power rack, arriving in the second half of 2026,” designed to “slot into existing AC infrastructure and deliver 800 VDC to compute racks within the row — no changes to the building’s electrical system required.” Delta’s GTC 2026 in-row 660 kW rack (six 110 kW shelves, 80 kW BBU per shelf, 480 kW of embedded backup, 18.5 kW AC/DC PSUs at up to 98%) is the PSU-OEM version of the same sentence. Schneider’s FAQ is the honest constraint: facilities that currently support up to 200 kW per rack can often take a DC sidecar; above 400 kW you are in purpose-built or heavily modified infrastructure.

Native (2027+ greenfield). Convert MV AC to 800 VDC once, at the perimeter. NVIDIA’s “DC power block” is that unit. Eaton’s MVSST is one way to build it: 15 kV class, 2 MW, >97% at 800 VDC, 12 isolated ports, pad-mount, and a spec sheet that also lists EV charging — because the industrial base is the EV industrial base. Vertiv describes two parallel facility paths: an MV DC UPS (transformer plus rectifier) as the higher-readiness first mover, and an SST whose supply chain “still has ground to cover.” NVIDIA, Google, and Microsoft published a joint 800 VDC white paper through OCP in March 2026 and an LVDC SST specification v0.3 in July 2026; NVIDIA’s August 2026 telling is “more than 80” companies building to it. A spec and eighty logos are a supply chain forming. They are not a commissioned hall.

At the tray. Direct 800 V into the compute rack. NVIDIA Kyber: high-ratio 64:1 LLC, 800 V to 12 V adjacent to the GPU. The device-level 800-to-50/12/6 V boards — Infineon, ST, Navitas, EPC — are the SiC/GaN story. The distribution story is that the tray is no longer allowed to hide an AC/DC stage. That is the whole point of moving conversion out.

Storage sits on the DC bus, or it is in the way. AI training is a synchronous load. NVIDIA, Microsoft, and OpenAI documented grid-scale oscillations from synchronized GPUs; NVIDIA’s ecosystem post says a rack can swing from ~30% “idle” to 100% and back in milliseconds, and that hundreds of megawatts ramping in seconds is a utility-interconnect problem. The 800 VDC architecture is supposed to make storage a low-pass filter: capacitors and supercapacitors at the rack for milliseconds-to-seconds, facility BESS at the interconnect for seconds-to-minutes. Eaton’s 2025 reference architecture literally puts supercapacitors in the drawing. Delta puts aluminum capacitors in the 18.5 kW PSU and 80 kW BBUs in the shelf. This is not UPS-as-insurance. It is UPS-as-filter. A hall that copies the 800 V busway and skips the storage is a hall that will fail a step-load test and then a grid-code test.

NVIDIA’s architecture, in the order it actually arrives

NVIDIA is not a switchgear company. It is the load. The May 2025 architecture post (updated July 2025) is the physics: 1 MW IT racks in 2027, industrial rectifiers at 13.8 kV → 800 VDC, up to 5% end-to-end efficiency, up to 70% lower maintenance from fewer in-rack PSU failures — that 70% is NVIDIA’s, not an operator study. Silicon, PSU, and facility partners fill the rest of the page; the October 2025 ecosystem post added ABB, GE Vernova, Hitachi Energy, Mitsubishi, Siemens, and the Kyber 64:1 LLC.

The staged product (August 2026). Existing factories do not have to wait for a greenfield DC block.

1. Power rack / sidecar, H2 2026. MGX-compatible, hybrid, existing AC, 800 VDC to the compute rack in the row. This is the object Vertiv, Delta, and Schneider are all describing. It is how you put Kyber-class voltage onto a floor that was poured for 480 VAC.
2. Row power center, 2027. Centralized power station for a full row, overhead 800 VDC busway, “up to 2 megawatts per row.” This is Vertiv’s “DC power in the pod.” It is also where DC breakers, busway ampacity, and a service procedure stop being a rack problem and start being a hall problem.
3. DC power block, new facilities. Medium voltage to 800 VDC in one step. SST or industrial rectifier. The decade-ahead drawing.

DSX is the reference-design wrapper around those three. OCP is the interoperability wrapper. Vladimir Troy, NVIDIA VP of data center infrastructure, August 2026: “Through OCP, NVIDIA is working with more than 80 ecosystem companies to give AI factories a practical path forward — not just a future vision.” A practical path is a product you can order, a spec a second vendor can build to, and a date. Two of those exist. The date on native halls is still 2027.

What NVIDIA is not claiming, if you read it cold. It is not claiming GB200 halls are 800 VDC. It is not claiming the 30% TCO or 70% maintenance figures have been measured on a production AI factory. It is not claiming the SST is the 2026 default versus a transformer-plus-rectifier. It is not claiming 800 VDC creates megawatts at the utility. Inference TCO is still the accelerator, the HBM, and utilization. A 1–5% conversion win is real hall-scale money. It is not a substitute for a fuller rack or a new interconnect.

Who actually ships

A GPU architecture is a load spec. A shipping electrical product is a SKU, a date, and a service organization. Four names sit on every NVIDIA 800 VDC slide. They are not interchangeable.

Vertiv is the sidecar company in public. The 13 October 2025 note moved 800 VDC from alignment to design maturity, commercial in H2 2026, ahead of Rubin Ultra / Kyber in 2027. The July 2026 “rack to data hall” piece is the honest one: PowerDirect 5000 at 400–900 kW; deployments beginning 2027; ±400 VDC as well as 800 V; a 1.5 MW AI-pod validation lab; customer engagements rather than a booth tour. Take H2 2026 as a portfolio date. Take 2027 as the deployment date Vertiv itself wrote.

Eaton is the grid-to-chip company with a datasheet. The MVSST is specified, not sloganed: 2 MW, 15 kV class, >97% at 800 VDC, 12 isolated outputs, pad-mount, liquid-cooled, and also an EV-charging product — the dual-use NVIDIA keeps citing. Beam Rubin DSX, unveiled for GTC 2026, wraps SST or MV transformer, UPS, switchgear, and a 1–2 MW “Beam Pod” that takes ~400 VAC (sidecar to 800 VDC) or native 800 VDC. The 2025 reference architecture is the parts list: supercapacitors, ORv3 busbar, DC connectors. That is further along than a slide. It is not a commissioned hall.

Schneider Electric is the UPS-and-switchgear incumbent. The AI-factory power page is the primary: 5% end-to-end, 45% copper, sidecar for existing ≤200 kW/rack sites, purpose-built above 400 kW, Galaxy VXL as the AI-tolerant UPS. Schneider will sell you the MV gear, the busway, the meters, and the digital twin. Ask which of those SKUs is DC-rated at 800 V and which is the 480 VAC plant they already ship.

Delta put a numbered 800 VDC in-row rack on a GTC floor in March 2026: 660 kW, 480 kW BBU, 98% AC/DC, 1RU 90 kW DC/DC for MGX. IEEE Spectrum’s 24 March 2026 DC-transition piece used that rack as evidence the power-delivery community is catching up. A GTC rack is engineering. It is not a hyperscale process of record.

Others sit in the same drawing — ABB, Siemens, Hitachi Energy, GE Vernova on the MV side; LITEON, Flex, Lead Wealth, Megmeet on the shelf; Infineon/ST/Navitas/EPC/Wolfspeed on the FET — and they matter. The four above are enough to show the pattern: nobody credible is shipping a native 800 VDC data hall as a 2026 catalog building. Everybody credible is shipping, demonstrating, or dating a sidecar, a UPS, or an SST that makes 2027 possible.

Switchgear, busway, safety

800 VDC in a serviceable rack is a safety product before it is an efficiency product. NVIDIA’s architecture diagram puts overcurrent protection — “fuse-and-disconnect combinations or emerging solid-state devices or safety breakers” — at the power-room/hall, hall/row, and row/rack boundaries, and then admits that “innovations will still be required to fully deliver possible overcurrent protection reliability and maintenance gains.” That sentence is the long pole.

DC does not zero-cross. An AC arc is interrupted 100 or 120 times a second by the voltage passing through zero. A DC arc is a continuous plasma until something extinguishes it. That is why DC breakers, solid-state interrupters, and rated disconnects are a different catalog from the AC plant in the room today. Creepage and clearance at 800 V, touch-proof connectors, isolation (ST’s 800 V boards already spend transformer volume on 10 kV isolation), and a lockout/tagout a technician can perform on a live aisle are the unglamorous reasons a 98% converter is not a commissioned row.

Busway is the copper you can actually pull. 800 VDC busway is how the row power center reaches 2 MW without a cable tray full of 54 V. It is also a listed product with an ampacity, a tap-off, a fire rating, and a coordination study against the breaker upstream. Eaton putting “power distribution technology through busbar included in the ORV3 design” in a 2025 reference architecture is the starting point, not the coordination study. Vertiv listing “high-efficiency DC busways” as a 2026 portfolio item is the same sentence from the other OEM. Ask for the DC rating, not the AC cousin.

Arc flash and workforce. NVIDIA and partners are “actively studying the CapEx, OpEx, and safety implications” of transformer-based versus SST approaches. That is a polite way of saying the NFPA 70E / IEC 61660 / vendor-application-guide work is not done. 800 V is inside the EV-charger and traction-inverter experience base, which is why every vendor mentions EVs. An EV service procedure is not a data-center service procedure. A hall that copies the voltage and skips the training is a hall that will lose a technician.

Grounding and ±400 V. A monopolar 800 V bus and a bipolar ±400 V bus are different shock, insulation, and protection drawings. Vertiv is engineering both. Do not assume a 650 V GaN board that is happy on ±400 V is a drop-in on a monopolar 800 V busway, or the reverse.

It still dumps heat. Moving the PSU out of the compute rack moves the conversion loss into the sidecar and, in the native architecture, into the perimeter rectifier. That heat is smaller if the conversion chain is shorter — NVIDIA’s 5% — and it is still heat a CDU or a rear-door or a perimeter CRAH has to take. Vertiv is explicit: “every change in compute power becomes a change in heat,” and the 800 VDC path has to be designed with the thermal system, not after it. A sidecar that is not liquid-ready at 400–900 kW is a space heater with a busbar.

What still breaks

Protection and service, not the FET. The 800-to-50 V board is a solved-enough demonstration. The connector, the DC breaker, the isolation, the arc-flash study, and the procedure are why NVIDIA dated native halls to 2027 and why Vertiv dated deployments the same way. There is no dual-qualified 800 VDC busway. You are on AC, or you are on DC, or you are on a sidecar that isolates the experiment next to the rack.

Step-load and the grid. Storage is in the architecture because the load is synchronous. Skip it and the 800 V bus is a stiffer way to show the utility a 100-millisecond cliff. Interconnection queues are already the long pole on generation. A hall that cannot prove it will not oscillate does not get the next feeder, regardless of bus voltage.

Standards versus SKUs. An OCP white paper (March 2026) and an SST spec v0.3 (July 2026) plus eighty companies is the right way to build a market. It is not a UL listing on a particular tap-off box. Interoperability means a Vertiv sidecar, an Eaton breaker, and a Delta shelf share a voltage range, a connector, and a protection curve. Until those are tested together — Vertiv’s 1.5 MW pod lab is exactly that claim — “ecosystem” is a booth.

Brownfield ceilings. Schneider’s 200 kW sidecar / 400 kW purpose-built split is the number operators should carry into a 2026 design review. A 120 kW NVL72 hall can take a sidecar path for the next GPU generation. A 600 kW Kyber row cannot live on the 480 VAC gear that was sized for 50 kW racks. Pouring 2026 concrete for 50 V, then expecting 2028 to be a busway swap, is how you strand a building.

Copper and lead times still exist. 800 VDC reduces copper in the rack and the row. It does not reduce MV transformer or switchgear lead times, or the utility’s interconnect calendar. Eaton’s “up to 50% faster installation” on the MVSST is a vendor figure for that box, not a campus schedule.

It does not shrink the GPU bill. Tokens per watt is still the accelerator. Lawrence Berkeley’s 2025 update, the DOE-cited reference, puts U.S. data centers at 11.8% of national electricity by 2030 in the reference case (649 TWh; 9.5–15.3% / 521–843 TWh across scenarios). Conversion efficiency is a 1–5% lever on a bill dominated by compute utilization and cooling. It is worth taking. It is not the generation problem, and it is not a reason to skip a CDU.

A realistic timeline

One useful heuristic: ask what voltage the busbar in the rack actually is, whether the 800 V conversion is in a sidecar or at the perimeter, and whether the number you were quoted is a SKU with a DC listing or a GTC photo. Do not ask whether the vendor “supports 800 VDC.”

The bottom line

2026 is the year 800 VDC became the default answer for the next AI factory — and the year the hall you can actually buy is still a 50-volt, 120 kW, liquid-cooled cabinet on a 415/480 VAC feeder.

The physics is old. At 100–120 kW, 12 V is impossible and 54 V is a copper-and-volume compromise that still ships. At 350–400 kW the compromise dies. At a megawatt, NVIDIA’s own 200 kg busbar and 64 U power-shelf numbers are the eulogy. 800 VDC is how you cut current, stages, and some percentage of the electricity bill: convert MV AC once, distribute DC, convert once at the tray. Vertiv, Eaton, Schneider, and Delta are all, in different combinations, selling that split: a 400–900 kW sidecar, a 2 MW SST, a 1,250 kW UPS, a 660 kW in-row rack with a 480 kW battery. Those are real. They are not a 2026 800 VDC hall, and they are not a substitute for the reactors or the CDUs the hall also needs.

For anyone making decisions today, the split is clean. If the question is what is inside a GB200-class row, you are looking at 415/480 VAC and a ~50 V bus, with SiC and GaN already in the better shelves. If the question is what to design into a 2027 Kyber row, you are looking at a sidecar or a row power center, 800 VDC busway, DC protection, and storage that can swallow a millisecond cliff. If the question is whether a vendor’s “800 VDC for AI” story is real, ask which layer (sidecar, row, perimeter), which numbered kW, the name of the DC-rated SKU, and whether the thing is a shipping product, a H2 2026 commercial date, or a GTC rack. The voltage in the press release will not answer it.

Frequently asked questions

What is 800 VDC power distribution?

A data-center electrical architecture that converts medium-voltage AC to a nominal 800 volts DC at the facility perimeter or in a rack-adjacent sidecar, distributes that DC down the hall, and converts once more at the compute tray. It is a bus and a conversion schedule, not a generator and not a cold plate.

Why can’t 48 V or 54 V feed 100 kW+ AI racks forever?

They can, at 100–150 kW — that is a GB200 NVL72. Current scales as power over voltage. At 400 kW the copper, the connectors, and the in-rack PSU volume run out. At 1 MW a 54 V bus is ~18,500 A and, in NVIDIA’s telling, up to 200 kg of busbar in one rack. 800 V makes the same 1 MW 1,250 A.

How is this different from 415/480 VAC?

415/480 VAC is the facility feeder in almost every 2026 hall. 800 VDC replaces that feeder (in the native architecture) or sits downstream of it (in the sidecar architecture). DC at 800 V carries more power in the same copper than 415 VAC, drops skin effect and reactive power, and lets you delete rack-level AC/DC. It does not delete the utility.

Are 2026 AI racks already on 800 VDC?

No. NVIDIA’s DGX GB200 documentation describes ~120 kW racks with AC-to-~50 V power shelves. NVIDIA dated native 800 VDC and 1 MW IT racks to 2027 / Kyber. 2026 800 V work is sidecars and in-row racks entering commercialization in H2, OCP specifications, and GTC demonstrations.

Who actually makes the equipment?

Facility and rack power vendors, not GPU vendors. Vertiv (PowerDirect 5000 sidecar, H2 2026 commercial), Eaton (2 MW MVSST, Beam Rubin DSX), Schneider Electric (Galaxy VXL UPS, sidecar path), Delta (660 kW in-row rack). Silicon and PSU names — Infineon, ST, Navitas, LITEON, Flex — sit inside those boxes; they are covered in the SiC/GaN guide.

Is this the same as SiC/GaN, liquid cooling, or SMRs?

No. SiC and GaN are the transistors in the rectifier and the tray converter. Liquid cooling removes heat. SMRs (and other generation) make energy. 800 VDC is how that energy is distributed once it is already in the building.

What about ±400 VDC?

A bipolar bus at ±400 V that reuses 650 V devices more comfortably. Vertiv is engineering it alongside 800 V. It is a topology fork, not a second architecture, until a hyperscaler standard says otherwise.

Does a 5% efficiency gain solve data-center electricity growth?

No. LBNL’s 2025 update puts U.S. data centers around 11.8% of national electricity by 2030 in the reference case. Conversion efficiency is a 1–5% lever on a bill dominated by compute and cooling. NVIDIA’s “up to 5%” is worth taking. It is not the interconnect, and it is not a token-cost substitute for a better accelerator.

Is this the same as backside power delivery?

No. Backside power is an on-die wiring scheme for GAAFET logic. 800 VDC is a facility and rack bus at hundreds of volts.

Methodology

This guide is based on primary vendor, operator, and government disclosures available as of 26 August 2026. NVIDIA’s technical blog “NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories” (20 May 2025, updated 31 July 2025: 1 MW IT racks starting 2027; up to 64 U of power shelves and up to 200 kg copper busbar per 1 MW rack on 54 V; 13.8 kV AC to 800 VDC at the perimeter; 85% more power versus 415 VAC; 45% less copper; up to 5% end-to-end efficiency; up to 70% maintenance and up to 30% TCO as NVIDIA’s bundled claims; GTC 2025 sidecar for 576 Rubin Ultra GPUs) is the architecture primary. NVIDIA’s 13 October 2025 ecosystem post (Kyber 64:1 LLC, 26% less area; millisecond rack swings; multi-timescale storage), the 800 VDC product page, and the 25 August 2026 corporate blog (MGX-compatible 800 VDC power rack H2 2026 on existing AC; row power center 2027 with overhead 800 VDC busway up to 2 MW per row; DC power block; NVIDIA/Google/Microsoft OCP white paper March 2026 and LVDC SST specification v0.3 July 2026; 80+ companies; DSX) are the staged-product primaries. NVIDIA’s DGX GB Rack Scale Systems user guide (~120 kW; AC to nominal 50–51 V DC; six 5.5 kW PSUs in eight shelves, N+N) is the 2026 volume-rack primary. Vertiv’s 19 May 2025 and 13 October 2025 NVIDIA-alignment releases (portfolio H2 2026, Rubin Ultra / Kyber 2027) and the 17 July 2026 “From rack to data hall” note (PowerDirect 5000 sidecar 400–900 kW; deployments 2027; ±400 VDC; Scott Armul May 2026 “DC power in the rack moving to DC power in the pod”; MV DC UPS versus SST; 1.5 MW validation lab) are the Vertiv primaries. Eaton’s MVSST catalog page (2 MW, 15 kV class, >97% at 800 VDC, 12 isolated outputs, up to 1,500 VDC architectures) and Eaton’s 2025 800 VDC reference-architecture release plus 2026 Beam Rubin DSX announcement are the Eaton primaries. Schneider Electric’s U.S. “800 VDC AI Factory Power Train Architecture” page (400 kW–1 MW; 5% / 45% copper; sidecar up to ~200 kW/rack; purpose-built above 400 kW; Galaxy VXL 400–1,250 kW) is the Schneider primary. Delta’s 16 March 2026 GTC release (660 kW in-row rack, 480 kW BBU, 18.5 kW PSUs up to 98%, SST “as high as 98.5%”) is the Delta primary. IEEE Spectrum’s 24 March 2026 “Data Centers Are Transitioning From AC to DC” is the independent trade-press primary. Demand context: LBNL United States Data Center Energy Usage Report: 2025 Update (OSTI 3374245, 18 June 2026) — 11.8% of U.S. electricity by 2030 reference (649 TWh; 9.5–15.3%; 521–843 TWh). NVIDIA TCO, maintenance, and copper-mass figures, Eaton installation-speed percentages, and Delta/Eaton SST peak-efficiency numbers are company-stated, not independent audits. Unofficial BOM prices, unsourced “800 VDC is already in GB200 halls” claims, and any conflation of this architecture with SiC/GaN devices, SMRs, or liquid cooling are omitted or labeled. Device-level conversion, generation, and heat removal are cross-referenced, not re-derived.