What is NVIDIA 800V HVDC? The direct answer: it is the power architecture NVIDIA announced in 2025 to feed the next generation of AI data centers, where a single rack now draws as much power as 1,500 homes. Instead of the old 48-volt AC chain that loses energy at every conversion step, NVIDIA 800V HVDC converts 13.8 kV medium-voltage AC straight to 800V DC at the data center perimeter, then distributes that DC directly to the rack. NVIDIA's 800V HVDC architecture roadmap states the system will support 1-megawatt IT racks and beyond starting in 2027, cutting copper requirements by about 45% and total cost of ownership by up to 30%, by transmitting up to 85% more power through the same conductor size.

The reason NVIDIA pushed for 800V HVDC is simple math. AI rack power jumped from 10 kilowatts to 1 megawatt. At 48 volts, feeding one megawatt means pushing 20,800 amps, which needs a copper busbar weighing about 200 kilograms inside one rack. At 800 volts, the same megawatt needs only 1,250 amps, and line losses drop to about 1/256 of the 48V case. The old low-voltage AC architecture physically cannot keep up, so NVIDIA, Google, and Meta have set 800V HVDC as the standard for megawatt racks, with the Open Compute Project co-defining the spec. Infineon, NVIDIA's named semiconductor partner, confirms AI data centers will need one megawatt or more per rack before the decade is out.

This guide answers what is NVIDIA 800V HVDC in full: the four-layer power chain, the Solid State Transformer, the SiC and GaN devices that make it work, the supply chain bottlenecks, how it handles backup power, cooling, and green energy, and where the market is headed. It also covers the classical HVDC transmission background, since the 800V data center architecture borrows the same DC principle that has carried bulk power across continents for 70 years.

At COBTEL, we have spent more than 20 years in optical communication and network cabling, building the transceivers, MPO patch cords, and rack infrastructure that AI factories run on. We watch the power side closely, because the fiber and copper we ship only matter if the rack stays lit.

Before diving into NVIDIA's 800V data center architecture, a short note on HVDC itself. HVDC, or High Voltage Direct Current, is a way of transmitting electric power as one-way direct current at high voltage instead of as alternating current. Classically, it means long-distance grid transmission: a converter station turns AC into DC at one end, the DC travels over an overhead line or submarine cable, and another converter station turns it back to AC at the far end. Wikipedia notes that today's systems run at 100 kV to 800 kV, with ultra-high-voltage links reaching ±1,100 kV.

DC wins over distance because it has no skin effect and no reactive power, so it carries only real power with low loss. National Grid's HVDC technical information puts AC cable losses at about 0.7% per 100 kilometers, which is why HVDC becomes cheaper past roughly 600 kilometers for overhead lines and 50 kilometers for cables. The two main converter types are LCC (thyristor-based, for bulk long-distance transfer) and VSC (IGBT-based, for offshore wind and weak grids), which Hitachi Energy sells as HVDC Classic and HVDC Light.

NVIDIA's 800V HVDC takes this same DC principle and scales it down from hundreds of kilovolts across continents to 800 volts across a single data center hall, or even a single rack. The physics is identical: higher voltage means lower current, lower current means less heat and less copper, and DC eliminates the reactive power and skin effect that plague AC. The rest of this guide covers that 800V application in depth.

Before the architecture, a short refresh on the parts that make 800V HVDC work at the rack scale, since the rest of this guide leans on them.

Capacitors store charge and stabilize voltage. They sit in parallel with the devices they protect, smoothing ripples and absorbing spikes so the rail never sags. Their golden rule: voltage across them cannot jump instantly.

Inductors do the opposite. They store energy in a magnetic field and stabilize current, sitting in series with their load. Their golden rule: current through them cannot jump instantly. Break an inductor's circuit suddenly and it generates a huge voltage to keep current flowing, a root cause of arcs.

Direct current arcs are the danger that makes HVDC engineering hard at any voltage. Pull apart two contacts under load and the gap ionizes into a glowing plasma column, an arc. AC arcs self-extinguish because the current hits zero twice per cycle. DC current never hits zero, so a DC arc, once started, will not go out on its own as long as the system voltage beats the arc's sustaining voltage. This is why 800V HVDC relies on purpose-built DC circuit breakers and fuses: breakers forcibly crush the arc environment and can reset after a trip, while fuses melt once to sacrifice themselves.

DC PDUs (Power Distribution Units) are the rack-level "power strips" that take one high-voltage DC input and split it into many protected outputs. VRMs (Voltage Regulator Modules) are the near-chip circuits that step the 12V or 48V bus down to the 0.5V to 1.3V, hundreds-of-amps supply a GPU core actually needs.

 

 

Here is the pattern you will see throughout: capacitors go in parallel to stop voltage drops and filter noise, breakers go in to stop arcs, fuses go in to stop overcurrent, and every stage is decoupled so no stage fights the next. For a deeper companion read, our 800V DC/DC power architecture guide goes hands-on with the rack-level conversion stages.

Adopting a brand-new power architecture never happens for fun. It happens because the old one breaks. To understand why 800V HVDC took over the conversation, you have to see what changed inside AI data centers and where the legacy "low-voltage AC distributed" design finally hit a wall.

Power density exploded. If one word sums up the last five years of AI compute, it is density: compute density, heat density, and the most overlooked one, power density. A traditional rack ran 3 to 10 kilowatts with a flat load. An AI rack runs 50 to 300 kilowatts today, headed for a megawatt, with a load that swings violently. Behind those numbers, global data center electricity use is racing from about 415 terawatt-hours in 2024 toward an estimated 945 terawatt-hours by 2030, roughly the yearly draw of Japan, according to the IEA's Energy and AI report.

The math forces a voltage jump. Power equals voltage times current (P = U × I). Deliver the same power at higher voltage and the current drops in proportion. Loss from resistance equals current squared times resistance (W = I²R), so halving the current cuts heat loss by a factor of four. Step from 48V to 800V and the current falls to about one sixteenth, which means resistive losses fall by roughly 278 times, as SemiAnalysis spells out in its 800VDC industry analysis. The practical payoff: line losses drop to about 1/256 of the 48V case, and facility-level power consumption falls around 5%.

Real-World Example: Deliver 1 megawatt at 48V and you get 20,800 amps. At a typical 6 amps per square millimeter of copper, that needs 34 square centimeters of cross-section, a busbar roughly 5 centimeters wide by 7 centimeters tall, weighing about 200 kilograms inside one rack. Delta's 800V PDB design paper scales that up starkly: a 1-gigawatt data center of such racks would need on the order of 500,000 tons of copper just for rack busbars. At 800V, the current is only 1,250 amps and the copper shrinks dramatically.

Load swings turned violent. Training holds a cluster near full load, which is stable. Inference does not. GPUs pulse from idle to full load in milliseconds, and a large synchronized cluster can swing grid-scale power in an instant. A joint study by NVIDIA, Microsoft, and OpenAI found power utilization can jump from 30% to 100% in milliseconds, a "power pulse" that stresses the grid and every upstream protection device.

Electricity is now the dominant operating cost. The industry measures efficiency with PUE, Power Usage Effectiveness: total facility power divided by IT equipment power. PUE of 1.0 is the impossible ideal where every watt reaches compute. Traditional air-cooled halls run 1.4 to 1.6; AI halls with liquid cooling target 1.15 to 1.3. Electricity already eats 40% to 70% of operating spend, and large training clusters sit at the high end. Hardware depreciation is a one-time capital hit; power decides whether the project breaks even. For the full picture on measuring and trimming that bill, see our guide to data center power consumption and cooling.

Reliability bars went stratospheric. A traditional data center targets 99.9% to 99.99% availability, allowing 8.76 to 52.6 hours of downtime a year. An AI training cluster demands 99.999% ("five nines"), under 5.26 minutes a year, because a multi-week training run that crashes from a power blip can erase a fortune in compute. Our breakdown of how to design and build a data center covers the structural side of hitting those targets.

Here is the side-by-side reality check:

The advantages of 800V HVDC pile up fast:

It supports megawatt racks. NVIDIA, Google, and Meta set the future AI factory ceiling at 1 megawatt per rack. The 48V architecture cannot reach it, and 800V is the only standardized path, co-defined by NVIDIA and the Open Compute Project.

Full-chain efficiency climbs from 88% to 98%. Infineon reports its GaN intermediate bus converters now reach up to 98% efficiency per conversion stage.

Space opens up. Thick busbars leave the rack, freeing room for GPUs and liquid cooling, lifting compute density 3 to 8 times and cutting footprint by about half.

Copper drops 45% to 75%. At the extreme, nearly 80% savings means material cost collapses, not just dips.

Payback under two years. At 1 megawatt per rack, a site saves hundreds of thousands of kilowatt-hours a year, covering the 800V equipment premium inside two years.

It natively fits green power and is forward-compatible. The 800V DC bus connects straight to batteries and solar with no inverter, and the distribution can scale to 1000V or 1500V without rebuilding the room.

The takeaway is blunt: 800V HVDC is not a choice, it is the only road that fits where AI racks are going.

An 800V HVDC data center collapses the old five-stage AC chain into a mostly-DC path with two core conversions. Instead of repeatedly stepping voltage down through transformers, rectifiers, and inverters, it converts medium-voltage AC to 800V DC once, ships that DC around the hall, and steps it down at the rack and the chip.

Why 800V and not higher? There are real limits above 1000V. Cross that line and you enter extra-high-voltage DC territory, where insulation, creepage, partial-discharge, and safety rules all jump a tier and every component's voltage rating gets expensive fast. At 800V working voltage, the matching 1200V SiC devices have comfortable derating headroom, and 1200V SiC is the most mature, best-value grade in the supply chain today.

The cleverness of 800V HVDC is simple to state: raise the distribution voltage, drop the current proportionally, cut the heat, shrink the copper, and raise end-to-end efficiency.

The old chain (five conversions): 110/35kV AC comes in from the grid, a main transformer drops it to 10kV AC, a second transformer drops that to 0.4kV (380V) AC, a UPS rectifies it to DC and inverts it back to 380V AC, and finally the server power supply converts 380V AC to 12V DC for the CPU. Every conversion burns 8% to 12% in total, the low-voltage 380V lines run hot, and voltage regulation lags when load spikes.

The transitional scheme: 10kV switchgear plus a line-frequency transformer plus an 800V SiC centralized rectifier cabinet plus a DC static transfer switch. This is the bridge most sites build first, since it reuses the existing 10kV infrastructure.

The ultimate scheme (NVIDIA's end state): 10kV medium-voltage AC enters a Solid State Transformer (SST), which outputs plus and minus 400V DC (800V total) directly to the hall's DC bus, then to a Sidecar rack power unit, then to the server's onboard supply, then to the GPU core. Just two core power conversions. No line-frequency transformer, no UPS, no multi-stage low-voltage AC distribution. As Infineon's 800V HVDC partnership announcement puts it, this is a shift from decentralized power supplies to a centralized architecture built on silicon, SiC, and GaN.

 

The ultimate architecture has five layers, all purpose-built for 800V DC, none of which mix with legacy 48V or 400V AC parts.

Layer 1: 10kV AC to 800V DC. This is the Solid State Transformer's job. The SST takes 10kV AC and, through high-frequency isolated conversion, outputs a stable plus and minus 400V DC (800V across the poles). It does the work of the old transformer and rectifier in one solid-state box, with active voltage regulation, harmonic filtering, and reactive power compensation built in.

Layer 2: the 800V main distribution system. Once the SST outputs DC, this layer moves it around the hall and protects it. The SST sends bipolar DC (plus 400V, minus 400V, 800V pole-to-pole) out on silver-plated, low-inductance laminated busbars that suppress SiC switching spikes. A large 800V film capacitor (Cbus) sits across the bus to absorb high-frequency ripple, buffer GPU peak-load dips, and hold the hall's bus voltage steady. After the filter, an 800V solid-state DC breaker (QFdc) disconnects the SST on any short, leakage, or voltage fault in milliseconds, with real DC arc-extinguishing capability. Storage, solar, and insulation monitoring all tie in here: an 800V battery bank charges off-peak and discharges at peaks with no inverter, a solar DC array feeds the bus directly, and an insulation monitoring (IM) unit watches both rails to ground and trips the breaker on leakage. The bus then enters the DC PDU cabinet, which splits one 800V feed into many branch outputs, one per rack, each with its own fuse or solid-state breaker, independent voltage and current sampling, and hot-swap protection. A master FPGA watches the bus in real time, dials the SST's output over fiber as load swings, and pulls protection triggers on branch faults.

 

Layer 3: the rack Sidecar (800V to 48V). The Sidecar is an external power cabinet bolted to the rack that steps 800V DC down to 48V. The 800V cable hits the rack's main DC breaker, which isolates rack faults and kills arcs, then a film capacitor filters cable noise and absorbs GPU peak shocks. The filtered 800V enters the Sidecar, whose heart is a SiC-plus-GaN LLC resonant converter: the control chip drives the primary SiC switches at high frequency, a resonant transformer provides isolation and steps the voltage down, and secondary GaN devices synchronous-rectify the output to a steady 48V. Closed-loop control tracks the GPU's idle-to-peak swings and holds 48V tight. The 48V outputs gather on a rack busbar that feeds each GPU tray through its own fuse.

Layer 4: server onboard supply (48V to chip voltage). Each tray takes 48V through a branch fuse and filter, then an Intermediate Bus Converter (IBC) steps it down to a 12V intermediate bus with isolation and full protection. A 12V filter bank buffers GPU peak dips, and multiple VRM multiphase buck converters step 12V down to the 0.8V to 1.2V the GPU core, CPU, and memory actually run on, each on its own closed loop so one VRM failure only drops one chip.

 

Pro Tip: The whole four-layer chain exists to do two things, move power at the highest practical voltage to keep current and heat low, and decouple every stage so a fault anywhere stops at the smallest possible boundary. For the component-level view of what lives inside each layer, TI's 800V DC data center power guide lists the exact parts: solid-state relays, 800V hot-swap controllers, high-accuracy battery monitors, isolated gate drivers, and isolated current and voltage sensors.

The make-or-break parts of an 800V HVDC system are wide-bandgap power semiconductors (SiC MOSFETs and GaN HEMTs), high-voltage film capacitors, nanocrystalline magnetic cores, and 800V solid-state DC breakers, plus the SiC gate-driver chips and ceramic substrates behind them. Silicon switches cannot hit the frequency and voltage at this density, so SiC and GaN are irreplaceable. Around those sit the support systems: liquid cooling, a BMS that watches bus voltage, insulation, and battery charge, and backup from diesel generators and 800V DC battery backup units (BBUs). The Solid State Transformer concentrates the hardest bottlenecks, which is why it is the most contested piece of the whole stack.

The supply chain splits into three tiers.

1700V and 3300V high-voltage SiC MOSFET modules are the core of the SST, costing 35% to 40% of the unit. Wolfspeed, Infineon, ON Semiconductor, and Rohm own this tier and are the default fit in NVIDIA's 800V SST. Domestic makers can ship 1200V SiC in volume, but their 1700V parts lag on yield, reliability, pulse current, and partial discharge for 24/7 AI loads. Upstream, the 8-inch and 12-inch SiC crystal growth and epitaxy equipment is locked down by U.S. and Japanese tool makers.

Fiber-optic isolated SiC gate drivers are mandatory because SiC switches at 50 to 100 volts per nanosecond, throwing fierce EMI that only optical isolation can keep off the control board. TI, ADI, ON Semi, and Power Integrations own the high end.

Premium ultrathin nanocrystalline ribbon (under 18 micrometers) and directionally annealed cores are the only material for the SST's high-frequency transformer, which decides its size and efficiency. Hitachi Metals and Toshiba hold the core recipes and continuous-casting equipment.

Large 800V and 1200V dry film DC-link capacitors must be low-ESR, last 100,000 hours, and stay low-loss at high frequency. Only Panasonic, TDK-EPCOS, and China's Faratronic have passed NVIDIA's GB200 and 800V certification, with Japanese makers holding about 70% of the high-end market.

Medium shortfalls (domestic volume exists, but a real performance gap at AI/SST grade): high-voltage GaN power ICs (Navitas is NVIDIA's named GaNFast supplier and the compute-power market leader), 10kV SST-grade FPGA control chips and voltage-balancing algorithm IP, 800V solid-state DC breakers (Schneider, Eaton, and Vertiv are mature; domestic makers are still on low-voltage solid switches), and SiC module silicon-nitride AMB ceramic substrates (Kyocera and Rogers dominate, since alumina cannot survive 1700V high-frequency thermal cycling).

Mild catch-up (domestic substitution largely done, only minor high-end gaps): 800V centralized HVDC rectifier cabinets, 800V high-voltage connectors and cable, liquid-cooling cold-plates, 1200V-and-below SiC modules, standard nanocrystalline cores, and low-voltage film capacitors and relays.

The two routes to 800V have very different bottleneck profiles. The current mainstream transitional scheme (line-frequency transformer plus 800V rectifier cabinet) is mainly bottlenecked on 1200V SiC, fiber drivers, and high-voltage film caps, with the rest fairly well localized. The next-generation SST end-state, NVIDIA's planned 2027 megawatt-rack standard, stacks the hard bottlenecks: 1700V SiC, fiber drivers, ultrathin nanocrystalline cores, large high-voltage film caps, industrial FPGA control chips, and silicon-nitride AMB substrates all at once. That makes the SST chain the single hardest part of the 800V industry to localize, and it is where NVIDIA's standard depends most on overseas parts.

Warning: The near-term risk is export control on 1700V SiC devices and the crystal and epitaxy tools to make them, which would directly throttle large-scale SST and 800V deployment. The cost risk is that high-end imported parts still top 40% of bill-of-materials cost, and the certification risk is that NVIDIA and Google's 800V ecosystem admission takes one to two years, with domestic parts running behind on validation.

AI data centers fear one thing more than inefficiency: an outage. In a traditional hall, the UPS and the backup battery are two separate cabinets that take hundreds of milliseconds to hand off, and a grid blip during that gap can reboot the whole GPU cluster and zero out hours of training. 800V HVDC kills that gap by hanging storage directly on the DC bus, with no inverter in the path.

The core idea is a three-level storage chain that rides through any outage with no break: busbar film capacitor, then supercapacitor, then high-voltage lithium battery. When the grid drops, the SST's master FPGA detects the input voltage collapse in microseconds, shuts down the grid-side rectifier, and the storage chain takes over, all in milliseconds, with no dead interval.

Here is the timing of a grid failure, step by step:

Fault detection (0 to a few hundred microseconds): the SST's sampling loop catches the AC input sag, and the FPGA immediately shuts off the rectifier-side switching so the rack stops pulling from the grid. The SST's internal bidirectional stage stays open, so bus-side storage holds the 800V rail.

Bus capacitor plus supercapacitor instant hold (0 to 20 milliseconds): the Cbus film capacitor dumps its stored energy first, covering the instant gap. Then the supercapacitor bank (CBU) discharges in under 2 milliseconds to absorb the GPU's shock load. Its job is to be fast and take the brutal pulse, which protects the lithium battery from huge current spikes. The goal: never let the 800V bus sag below the Sidecar's minimum input, so the GPU never sees a blip.

High-voltage battery sustained supply (after 20 milliseconds): the BMS sees the bus voltage deviation and the battery bank starts discharging 800V DC straight to the bus. A short backup (60 to 120 seconds) covers saving model parameters and a safe shutdown. A long backup (15 minutes plus) covers waiting for the diesel generator to restart.

There are two energy-flow modes. Normal (grid up): grid feeds the SST, the SST feeds the 800V bus, the bus feeds the load, and surplus trickle-charges the storage. Outage (grid down): supercapacitor and battery feed the 800V bus directly, through the DC PDU and Sidecar to the GPU, with no inverter and no rectifier in the path, so two conversion losses disappear.

The SST's bidirectional stage is the unsung hero. After a grid loss, its rectifier side is off, but the isolated bidirectional stage stays open as a controlled channel that lets bus energy flow only to the load, blocks any backfeed toward the 10kV switchgear, and closed-loop regulates the bus back to the plus and minus 400V band. The protection logic is deliberate: on a grid drop the system keeps the bus alive on storage rather than tripping the main breaker, and only orders an orderly shutdown if the bus voltage keeps falling past the storage floor. A downstream short just trips that branch's solid-state breaker, and storage keeps the rest of the hall running.

Yes, and on an 800V DC bus it is simpler than on AC. Solar panels and batteries are already DC, so they tie straight onto the 800V bus through DC-DC converters and skip the inverter entirely, which alone adds 3 to 5 efficiency points. Wind rectifies in. The data center stops being a passive load at the end of the grid and becomes a "source-grid-load-storage" DC microgrid node: solar drives by day, storage takes the night shift, wind fills gaps, and the grid is the backup.

The old AC way was clumsy: solar went through an inverter to the AC bus, then through the UPS to the server, and storage needed a bidirectional PCS with two AC-DC round trips. The 800V all-DC way puts solar and storage in parallel on the main 800V bus, using the SST's bidirectional power and bidirectional DC-DC converters to coordinate source, grid, load, and storage as one DC-coupled system.

The hardware ties in at the second distribution layer, right after the SST output and before the DC PDU:

Solar (and wind, similarly): the solar array's DC output runs through a maximum-power-point-tracking (MPPT) boost DC-DC converter onto the 800V bus, matching the bus's 720V to 880V operating band, with its own DC breaker, surge protection, and anti-islanding. Dropping the inverter lifts system efficiency 3% to 5%.

Storage (lithium iron phosphate banks): because battery voltage is not a fixed 800V, it needs a bidirectional isolated DC-DC converter. Charge mode sends 800V bus to battery; discharge mode sends battery to 800V bus. The converter isolates the battery, controls current, limits cell string count, and balances voltage.

Supercapacitor: also through a bidirectional DC-DC, it handles millisecond-scale smoothing for solar sags and GPU load shocks, shielding the lithium battery from big current pulses.

SST bidirectional role: excess solar can flow back through the SST, inverted to 10kV AC and sold to the grid. At off-peak prices, grid power comes in through the SST to charge storage. That is two-way grid tie and peak-valley arbitrage in one box.

An energy management system (EMS) runs four modes: day solar surplus (solar feeds GPUs, excess charges storage or sells to grid), day solar short (solar feeds load, storage and grid cover the gap), night (storage discharges to cut grid draw, or holds the bus on outage), and off-peak (grid charges storage through the SST). The same seamless outage logic applies: on a 10kV loss, the SST blocks the grid-side rectifier, the Cbus capacitor stabilizes instantly, the supercapacitor holds the bus, and the storage DC-DC sustains discharge, so the GPUs never feel it.

Safety is built around the 800V DC arc risk. Every solar and storage branch has its own DC breaker, the bus IM insulation monitor watches ground leakage around the clock, every bidirectional DC-DC carries over-voltage, over-current, short-circuit, and reverse-charge protection, and the EMS enforces state-of-charge limits. This architecture fits national "East-Data-West-Compute" and "green power plus compute" strategies cleanly, and gives the site a strong carbon-accounting position.

You cool a megawatt-per-rack HVDC system in four separated layers: liquid cold-plates on the SST's SiC modules, forced air in the 800V distribution room, liquid or air on the Sidecar depending on density, and direct liquid cold-plates or immersion on the GPU and VRM tray, all under one closed-loop thermal controller. The guiding principle is to keep each layer's heat sources fully separated so power cooling and compute cooling never fight each other.

Layer 1, the SST (power room): the SiC power modules, nanocrystalline high-frequency transformer, and film capacitors produce very high heat flux that air cannot handle at megawatt scale. The fix is microchannel cold-plate liquid cooling, with the SiC baseplate bonded to the cold plate through a thermally conductive insulating pad, running deionized water plus glycol. An aluminum-nitride insulating baseplate separates the high voltage from the coolant plumbing, and the control FPGA gets isolated air cooling.

Layer 2, the 800V main distribution (DC PDU, storage, busbar): the heat here is mostly I²R loss from the laminated busbar, plus breakers, fuses, and the battery bank. There is almost no high-frequency conversion loss, so this layer rarely needs water. Precision CRAC units force air across the busbar and breaker terminals to stop oxidation hot spots that could seed a DC arc, and the high-voltage battery bank gets its own air or liquid cabin for temperature uniformity.

Layer 3, the rack Sidecar (800V to 48V): the SiC and GaN LLC resonant devices are the heat. The high-density route uses Sidecar cold-plate liquid cooling tied to the rack's common CDU loop: quiet, dense, and good past 200 kilowatts per rack. The mid-low-density route uses forced air with a big heatsink: cheaper but louder and capped lower. The key trick is physical separation of the Sidecar from the compute trays, so GPU hot exhaust never bakes the power modules.

Layer 4, the server tray (IBC, VRM, GPU): the hottest zone. The GPU core gets a direct microchannel liquid cold-plate. The VRMs and IBC either bond to a heatspreader sharing the GPU liquid loop or get their own small cold-plates. For the next generation of ultra-high density, full-tray immersion in dielectric fluorinated fluid is the emerging option, putting the GPU, VRM, and IBC in one isothermal bath that kills hot spots.

The whole system runs on one closed thermal loop. Sensors read SST, Sidecar, DC PDU, GPU junction, and battery temperatures, then act in tiers: a mild rise speeds up the CDU pump and fans; a mid-level warning dials back SST and Sidecar output and limits GPU power; a severe over-temperature triggers staged shutdown that drops the faulted branch first. HVDC has a built-in cooling advantage over the old 48V architecture: the high-voltage bus carries far less current so busbar copper loss drops sharply, the Sidecar sits outside the tray so power heat is isolated from the GPU, SiC switches have lower switching loss than silicon, and with storage on the DC bus the UPS's conversion heat disappears.

As AI racks scale, the optics inside them scale too, and the DAC vs AOC choice in AI data centers becomes part of the same density and thermal conversation.

HVDC is one of the most certain growth bets in power electronics, riding two trillion-dollar tracks at once: electric vehicles and AI compute. The 800V grade is in a fast penetration climb now, with mass deployment expected from 2027, and a long-term path upward toward 1000V and 1500V as SiC and GaN mature.

The market signals are loud. The 800V HVDC data center market was about $3 billion in 2025 and is growing at over 40% a year. NVIDIA has set 2027 for its full switch to 800V HVDC, and Google, Meta, and Oracle are building new AI compute centers on it at 100% adoption, with OCP naming it the next-generation power standard. SemiAnalysis's 800VDC industry analysis tracks a four-phase rollout starting in late 2026 and early 2027, where Phases 1 and 2 retrofit existing AC distribution to 800VDC at the rack level via the power rack, led by hyperscalers, with Meta running 600 to 800 kilowatts per rack and Amazon landing at 800 kilowatts on plus and minus 400V.

Demand on the semiconductor side is just as clear. Vicor reported record AI data center demand and signals over $600 million in 2026 revenue on the way to a $2.5 billion long-term target, and Infineon confirms AI racks will need one megawatt or more before the decade is out.

The electric vehicle track runs in parallel and validates the same components. In the first quarter of 2026, 800V pure-EV penetration in China hit 22%, double 2024, with near-100% fitment above the premium tier. By 2028, over 60% of new pure EVs are expected to run 800V. The 800V fast-charger market is forecast at 31.1 billion yuan in 2026, growing over 42% a year, with 90% of new highway chargers 800V-compatible, delivering 250 to 400 kilometers of range in 10 minutes.

The roadmap points one way: voltage climbs 800V to 1000V to 1500V, SiC goes universal, GaN spreads through high-frequency spots, and the architecture reaches from cars and data centers into storage, solar, rail, and marine. As that compute base grows, the optical fabric feeding it grows too, which is why 400G and 800G optical transceivers and high-speed DAC and AOC cables ship right alongside the HVDC power build-out.

The blockers are real but specific: DC arc safety, a still-maturing Solid State Transformer, 800V-rated protection parts, liquid-cooling insulation, and unfinished standards.

DC arc risk. A DC arc, unlike an AC arc, has no zero-crossing to put it out, so it needs dedicated arc-extinguishing design in every breaker and contactor. That is why solid-state DC breakers, not mechanical fuses, do the protection work at 800V.

Personnel safety. 800V DC contact voltage is far above the safe threshold, so the site needs stricter isolation and lockout/tagout (LOTO) procedures than a 48V or 208V AC hall. Worker training is a real part of the timeline, and NVIDIA's own announcement flags safety, standards, and workforce readiness as joint challenges.

The Solid State Transformer. The SST is the core of the whole scheme, with high technical barriers, hard development, and a need for long on-grid proving and certification. Eaton's 800VDC deployment whitepaper frames the practical roadmap around medium-voltage solid-state transformers and direct-to-rack DC feeds, which tells you the SST is the gating item.

Protection devices. The hall needs 800V-and-up rated DC breakers, fuses, and contactors, and the high-end solid-state versions are still dominated by Schneider, Eaton, and Vertiv.

Liquid-cooling insulation. Bringing 800V DC into a liquid-cooled rack adds insulation and leak-detection requirements that a 48V rack never worried about, since a coolant leak near 800V is a serious arc and corrosion hazard.

Standards. IEC, NEC, and OCP are still updating their specs. OCP's Diablo 400 spec and the EMerge Alliance are pushing 380V to 800V DC standardization, and an independent 800V HVDC research note tracks the IEEE C57.16 path that is extending the traditional transformer standard to cover the SST, with early participants like Delta, Eaton, and Siemens. Full global safety certification and compliance will still take time.

The 800V HVDC supply chain spans power semiconductors, power-supply OEMs, and an EV-grade component ecosystem that overlaps heavily with the data center side.

Power semiconductors. Wolfspeed leads global SiC and supplies 800V inverter core devices. Infineon is the world's largest power semi maker and NVIDIA's named 800V HVDC partner on SiC and GaN. ON Semiconductor and Rohm round out the high-voltage SiC field. Texas Instruments supplies the analog and isolated gate-driver, sensing, and hot-swap silicon. Navitas is NVIDIA's official GaNFast supplier and the compute-power market leader in GaN.

Power-supply and infrastructure OEMs. Delta is a leading power-supply OEM and named 800V partner. Vertiv, Schneider, and Eaton are the mature high-voltage DC breaker players, with Eaton publishing an MVSST direct-to-rack roadmap. Vicor leads modular factorized power conversion and is central to the 800V DC-DC stage. Siemens is an early SST participant, and ABB is a long-standing HVDC grid player.

Chinese players. On SiC and substrates: StarPower, Innoscience (GaN, an NVIDIA alliance member), Sanan Optoelectronics, and Tianyue Advanced. On power equipment: Zhongheng Electric, Kehua Data, and Megmeet lead domestic 800V rectifier cabinets. Faratronic broke through on compute-grade high-voltage film capacitors. On liquid cooling: Envicool and Gaolan. On high-voltage connectors: AVIC Optoelectronics, Yonggui, and Reach.

EV 800V names (overlapping component base). BYD, Xpeng, and Li Auto lead Chinese 800V platforms. Lucid built the first production 900V car. BorgWarner, Bosch, and Continental supply high-voltage inverters and thermal management. Tesla's 4680-plus-high-voltage architecture is the North American benchmark. For 800V DC-DC and onboard chargers, Inovance and Xinrui are domestic leaders.

Pro Tip: If you are mapping a supply chain for an 800V project, start from the SST and work outward, because the SST is where foreign dominance and certification delay both concentrate. Everything downstream of the Sidecar has a healthier domestic substitute pool.

Three things to take away. First, HVDC is fundamentally about moving power as direct current at high voltage, whether that means 800 kV across a continent or 800 volts across a data center hall. In both cases DC beats AC because it carries only real power, with no reactive loss and no skin effect. Second, the classical transmission application is the established, evergreen meaning of "what is HVDC": converter stations, long-distance lines, submarine cables, and grid interconnection, all proven over 70 years. Third, the emerging 800V data center application is where the concept is now moving fastest, driven by AI racks that will hit one megawatt by 2027 and physically cannot be fed at 48 volts.

At COBTEL, we pair that power transformation with the infrastructure it runs on, from MPO patch cords and 400G/800G transceivers to the rack cabinets and cable management that hold an AI hall together. If you are planning an AI data center build, fill the inquiry form at the bottom of this page and our team will help you spec the cabling, optics, and rack infrastructure that fit your power architecture. The future of compute runs on HVDC, and the fiber and copper that carries its data has to keep up.

 

NVIDIA 800V HVDC is NVIDIA's power architecture for AI data centers that converts 13.8 kV AC grid power directly to 800V DC and distributes it to megawatt-class AI racks. Announced in 2025 for full deployment in 2027, it cuts copper use by about 45%, transmits up to 85% more power through the same conductor, and lowers total cost of ownership by up to 30% compared to the old 48V AC architecture. It uses a Solid State Transformer and SiC and GaN power semiconductors to collapse the old five-stage AC chain into two core conversions.

At 48V, feeding a one-megawatt rack means 20,800 amps and roughly 200 kilograms of copper busbar, which is physically unbuildable. At 800V, the current drops to 1,250 amps, about one sixteenth, and heat loss falls by about 278 times. Going above 1000V triggers extra-high-voltage DC rules that make every component far more expensive, so 800V sits in the sweet spot where 1200V SiC devices are mature and affordable.

Yes. Instead of a separate UPS cabinet, 800V HVDC hangs storage directly on the DC bus as a three-level chain: a busbar film capacitor, a supercapacitor bank, and a high-voltage lithium battery. When the grid drops, this chain takes over in milliseconds with no inverter in the path, so there is no conversion gap and no reboot.

It can be, but it demands stricter discipline than a 48V or 208V AC hall. Because 800V DC contact voltage is far above the safe threshold, sites need solid-state DC breakers that can extinguish arcs (which do not self-extinguish in DC), strong insulation, leak detection in liquid-cooled racks, and rigorous lockout/tagout procedures. Worker training is a real part of any deployment timeline.

Mass deployment is expected from 2027, timed to NVIDIA's Kyber rack-scale systems. SemiAnalysis tracks a four-phase rollout starting in late 2026 and early 2027, led by hyperscalers like Meta, Google, and Amazon. The 800V HVDC data center market was about $3 billion in 2025 and is growing at over 40% a year, with full global safety certification still catching up through IEC, NEC, and OCP.