> ## Content Index > Fetch the complete content index at: https://nexi.fund/llms.txt > Use this file to discover other available public pages before exploring further. # 79.8% to 93.8%: How 800 Volts Rewrites AI Data Center Power - URL: https://nexi.fund/800vdc-ai-data-center-power/ - Published: 2026-10-13T15:30:28.000Z - Updated: 2026-10-13T15:30:28.000Z - Description: Rack power stopped being a silicon problem and became a wiring one. On 12 October NVIDIA put the Vera Rubin rack designs into OCP, turning 800 VDC into an open standard. Enverus models facility efficiency at 93.8%, up from 79.8%. - Author: Nexi.fund Labs - Tags: AI & Infrastructure, #mode-6, #hook-statistic, #track-D, #layout-radar-scan 79.8% is an odd number to lead with. It is the share of grid electricity that reaches the processors in a conventional AI data centre, after every conversion stage between the utility feed and the GPU has taken its cut. Enverus Intelligence Research modelled a different architecture and landed on 93.8%. Fourteen percentage points of stranded power sit between those two figures. The fix is electrical rather than computational: raise distribution inside the building from 415 volts alternating current (VAC) to 800 volts direct current (VDC), convert once at the perimeter, stop converting at every stage downstream. On 12 October 2026 that fix stopped being one vendor's white paper. 93.8% facility efficiency ↑ 14 pp vs 79.8% #### End-to-end modelled efficiency Grid power actually reaching compute under the modelled 800 VDC architecture, up from 79.8% on the legacy path · *Enverus EIR, 2026* \-13% electrical capex #### Distribution capex cut Modelled saving on electrical capital expenditure for 800 VDC distribution against 415 VAC, on the same campus · *Enverus EIR, 2026* 85% more power per conductor #### Busway throughput gain Additional power a row-level 800 VDC busway carries through the same conductor cross-section, paired with 45% less copper than 415 VAC · *NVIDIA 800 VDC specification* 250 kW max rack power #### NVIDIA MGX 1.2 rack ceiling Density the Vera Rubin rack specification supports, delivered on a liquid-cooled 5,000A busbar contributed to the Open Compute Project · *OCP Global Summit, October 2026* ## How 415 volts stopped scaling Power is voltage times current. Hold the watts fixed and double the voltage, and the current halves. That single relationship is the entire argument for 800 VDC, because resistive loss in a conductor goes as the square of current. Halve the current and you cut the loss fourfold. The consequence is arithmetic that engineers recognise instantly and procurement teams feel in the copper bill. Resistive heating in busway and cabling is not a defect to be engineered around. At these densities it is the dominant term, and the only levers are conductor mass, cross-section and voltage. That is why the legacy numbers stopped working. In-rack distribution settled on 54 VDC decades ago, because that is what semiconductor packaging could regulate cheaply at the rack edge. Rack loads stayed in the tens of kilowatts for most of the industry's life, and the voltage never had a reason to move. Enverus notes that AI rack density is already past 100 kW per rack and climbing. Renesas phrases the same pressure from the component side: racks climbing from roughly 120 kW toward megawatt scale have run out of board space and thermal headroom before they reach their power limits. As we wrote in our 13 October piece, GridCARE verified 5 gigawatts of spare grid capacity and attached a $350B figure to the rest of it. That is the supply side of the ledger. This is the delivery side, and it is where the copper and the conversion stages live. ## What 800 volts changes, physically Three things, in descending order of how often they get mentioned. Current falls, so copper falls with it. NVIDIA's own 800 VDC specification puts a row-level DC busway at 85% more power through the same conductor size and 45% less copper than 415 VAC distribution. A silicon-carbide and gallium-nitride supplier puts the same idea more bluntly: traditional enterprise data centres relying on legacy in-rack distribution can no longer meet the multi-megawatt densities accelerated computing platforms demand. Conversion stages disappear. The legacy path takes 13.8 kV AC from the utility and converts it repeatedly on the way down — transformers, switchgear, uninterruptible power supplies, AC distribution, then rack-level supplies feeding 54 VDC rails. 800 VDC moves the AC-to-DC conversion to the data centre perimeter or into a row power station and hands the compute rack the voltage it wants. Voltage headroom extends. 800 VDC sits inside the IEC low-voltage DC classification of 1,500 VDC or below, which is the standard's stated anchor. Higher-voltage designs are being discussed up toward that ceiling. Raising the ceiling later is a distribution-layer change, not a rewiring of the building. | Parameter | Legacy 415 VAC path | 800 VDC path | | ------------------------------ | --------------------------- | --------------------------------------- | | **In-rack distribution** | 54 VDC | 800 VDC at the rack | | **AC-to-DC conversion** | Multiple stages on the path | One stage at the perimeter | | **Steps down to core voltage** | AC/DC then DC/DC | 800 V to 12 V or 6 V, then to under 1 V | | **Copper per MW** | Baseline | −45% to −60%, depending on the model | | **Facility efficiency** | 79.8% modelled | 93.8% modelled | | **Rack ceiling** | Several hundred kW targeted | 250 kW in MGX 1.2; 1 MW from 2027 | Compiled from NVIDIA's 800 VDC architecture specification, the Enverus EIR report of June 2026 and the MGX 1.2 rack specification contributed to OCP in October 2026 ## One conversion at the perimeter, two at the rack The conversion-stage collapse is where the efficiency numbers come from, and it is easier to count than to argue about. Legacy: medium-voltage AC arrives, a transformer steps it down, switchgear and uninterruptible power supplies condition it, AC distribution carries it, rack power supplies rectify it, and a final DC/DC stage regulates it to sub-1V core voltage. Every one of those stages has losses in the 97–99% range individually, and they multiply. 800 VDC collapses the middle. Convert 13.8 kV AC to 800 VDC once. Distribute DC across the hall. Convert twice more at the rack. Texas Instruments published a complete 800 VDC power architecture built on NVIDIA's reference design in March 2026 that requires exactly two conversion stages from 800 V to processor power, with the first stage an isolated bus converter integrating GaN power stages at 97.6% peak efficiency and above 2,000 W per cubic inch. Navitas Semiconductor brought the same idea to the Open Compute Project summit on 12 October: a 24 kW power delivery board, co-developed with NVIDIA, converting 800 VDC to 12 VDC in a single stage. That removes the intermediate low-voltage shelf entirely. ⚙ **The hybrid path is the underrated part** NVIDIA's MGX-compatible 800 VDC power rack is due in the second half of 2026 and is designed to slot into buildings that already run AC distribution, delivering 800 VDC within the row without touching the building's electrical system. Operators building dedicated campuses get a different unit — a row power station feeding an overhead 800 VDC busway at up to 2 MW per row, expected in 2027\. Most of the installed base will take the retrofit, not the rebuild. ## The supplier queue behind the standard A power architecture is a standard on paper and a bill of materials in silicon. The queue tells you which parts are real. Renesas shipped the clearest single component on 30 September: the TP65H020G4PLSGBD, a 650 V gallium-nitride device with 20 milliohms of on-resistance in a dual-side-cooled 8 by 8 millimetre package. The package is 57% smaller than the 10 by 15 millimetre top-side-cooled part it replaces, and dissipating heat from both faces drops top-side thermal impedance by 10%. A 6 kW reference design built on it reaches 2.6 kW per cubic inch, with measured full-load efficiency 0.21% above an equivalent board using the older package. Two dates matter more than the datasheet. The device is sampling now, with mass production planned for mid-2027\. That is the real constraint on the whole architecture, and it sits in a single supplier's qualification schedule. Around that anchor, the rest of the queue is filling in. Microchip and Navitas published an 800 VDC to 6 VDC reference design in early October. onsemi has worked with NVIDIA on 800 VDC distribution and solid-state transformers since 2025\. Texas Instruments covers the bus converter, a 30 kW 800 V AC/DC supply unit and capacitor-bank units. Vertiv's open rack for MGX v1.2 carries a rear liquid-cooled DC busbar rated to 5,000 A with up to four 110 kW power shelves. Visionbay unveiled the K-1 platform at the summit — a Vera Rubin-oriented AI factory built on NVIDIA's 800 VDC reference architecture with Vertiv power and thermal hardware, with cluster-level power stations centralising the AC-to-DC conversion. Scale on the supplier side is real. NVIDIA says more than 80 companies are building to its 800 VDC specifications, which is what turns a reference design into an interoperable one. Its MGX-compatible power rack, the row power station and the facility-scale DC power block are captured as reference designs in the company's DSX blueprint. Chinese vendors are working the same seam from the other direction. At its APSARA conference on 29 September, Sungrow presented ten trends at the clean-energy and AI computing convergence, with 800 VDC named as an emerging supply standard, systems targeting more than 94% end-to-end efficiency, and silicon carbide and gallium-nitride devices named as the beneficiary opportunity. The same forecast puts grid-forming converter penetration above 40% by 2028 as tenders start specifying the capability for AI computing facilities. > Next-generation AI chips are forcing a rethink of data center power architecture. At the rack densities now coming into view, 800 VDC is less about chasing incremental savings and more about making the next phase of AI infrastructure physically and electrically workable.— Carson Kearl, senior analyst and report author, Enverus Intelligence Research ## Copper, and why two credible models disagree The cleanest way to see that this is still a modelling exercise is to put the two serious numbers next to each other. Enverus EIR puts the copper saving at up to 60% per megawatt against 415 VAC, lifts end-to-end facility efficiency from 79.8% to 93.8%, and cuts electrical capital expenditure by 13%. Its wider model, pairing 800 VDC with behind-the-meter solar, storage and fuel cells, produces a ten-year value stack of $1.63B per 1 GW campus, of which $440M comes from eliminating AC-path conversion losses. NVIDIA's specification puts the copper saving at 45% against the same baseline, the end-to-end efficiency gain at up to 5%, maintenance cost down up to 70% and total cost of ownership down up to 30%. Bloom Energy, selling into the same transition, claims 6–8% overall efficiency improvement, $3.6B in non-compute capital savings per 1 GW and $5.5B off five-year total cost of ownership. A fourteen-percentage-point efficiency gain and a five-percentage-point gain are not the same claim. They differ in where the boundary sits — facility input to processor pin, or utility input to rack busbar — and in whether the modelled system includes the liquid cooling loop. Read as contradictions, they discredit the transition. Read as different boundaries, they describe the same physics from two vantage points. The copper gap is narrower and harder to explain. Sixty percent and forty-five percent are both plausible for different conductor mixes and busway designs, and neither publisher has published the assumptions. For a procurement team sizing a campus, that is the number worth demanding. ## Three dates decide whether this lands The direction is settled. The schedule is not. The MGX-compatible 800 VDC power rack is due in the second half of 2026, which is the first hardware that can be ordered into an existing AC building. Renesas mass production follows in mid-2027\. NVIDIA's row power station, rated up to 2 MW per row, is expected in 2027\. The 1 MW IT rack and the Kyber rack-scale systems that carry the full architecture are 2027 as well. Three dates, one year, and every one of them is a supplier or vendor commitment rather than a measured result. 📊 **Signals worth tracking** Renesas TP65H020G4PLSGBD moving from sampling to production, and whether that mid-2027 date holds First installed base of the MGX-compatible 800 VDC power rack in the second half of 2026, and any published measured efficiency rather than modelled Whether the OCP Open Data Center Specification v0.7 and Grid Resilience Specification get written into procurement language or stay voluntary Copper intensity per megawatt as reported by two independent operators, which would settle the 45-versus-60 gap Grid-forming inverter share in AI facility tenders, which Sungrow forecasts above 40% by 2028 ## What the numbers do not settle Every efficiency figure above is either a vendor specification, an analyst model, or a design target. There is no large operating 800 VDC data centre publishing measured numbers yet, because there is no large operating 800 VDC data centre yet. The Open Compute Project grid resilience specification targets restoration to near-full load in under two seconds after a disturbance. That is a specification target, and the same applies to the 1.3 times power-efficiency claim attached to the proposed 400G-per-lane optical interface. Design intent is not measurement. DIGITIMES Intelligence framed its October report on the transition as taking shape with future hurdles requiring several factors to be balanced, which is the careful way of saying the conversion is not settled engineering. Copper costs, protection gear that exists for AC networks and not DC ones, arc-fault behaviour in DC distribution, and the service skill base for high-voltage DC equipment are all still open. A liquid-cooled 5,000 A busbar carrying the whole rack is also a new failure mode in a building designed around redundant strings. One thing the transition will not fix is the grid queue. Interconnection timetables, not busway voltage, set the schedule for most of the announced capacity. As we wrote in September about TAR's $120M Series A at a $1B valuation for off-grid power at AI data centres, the supply question and the delivery question are being solved by different people on different clocks. 800 VDC addresses the second one well. 🎯 Rack power stopped being a silicon problem some time ago and became a wiring problem. 800 VDC is the industry's answer, and as of 12 October 2026 it is an open standard with more than 80 companies building to it. The economics are modelled, not measured. Enverus puts the facility-efficiency gain at fourteen percentage points and the copper saving at up to 60% per megawatt; NVIDIA claims 45% and up to 5%. Different boundaries, same physics. The binding constraint is a sampling schedule. Renesas' GaN device is the component the architecture leans on hardest, and mass production is mid-2027\. Every headline number above arrives with that date attached. [ NVIDIA pushes 800V DC, 400G optics and open rack designs at OCP Roundup of the announcement set from OCP Global Summit: MGX 1.2 at up to 250 kW per rack, the DSX AI Factory components, Open Data Center Specification v0.7, and the Visionbay K-1 800V DC factory with Vertiv. Converge Digest ](https://convergedigest.com/nvidia-ocp-2026-vera-rubin-mgx-spectrum6-usfp/?ref=nexi.fund) The single most complete inventory of what was actually announced on 12 October, and explicit that the grid-resilience and optics figures are design targets rather than verified results. [ NVIDIA opens Vera Rubin rack designs to OCP as blueprint for AI factories Same-day confirmation that NVIDIA contributed its Vera Rubin rack and compute-tray designs under MGX 1.2, with blind-plug liquid cooling, a liquid-cooled 5000A busbar and native 19-inch device support. Particle ](https://particle.news/story/nvidia-opens-vera-rubin-rack-designs-to-ocp-as-blueprint-for-ai-factories?ref=nexi.fund) Establishes the moment the architecture stopped being proprietary. Common rack-to-grid power, cooling and networking specs are what let operators mix vendors instead of buying a single stack. [ NVIDIA 800 VDC architecture will power the next generation of AI factories The primary technical specification behind the transition: 13.8 kV AC converted directly to 800 VDC at the perimeter, row-level busway at 85% more power through the same conductor, 1 MW IT racks from 2027. NVIDIA Developer Blog ](https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories?ref=nexi.fund) First-party source, so treat every figure in it as a design claim rather than an independent result. Useful precisely because it states the boundaries of the model. [ 800 VDC rewrites AI data center power economics Enverus Intelligence Research models the transition outside the vendor ecosystem: electrical capex down 13%, facility efficiency from 79.8% to 93.8%, copper mass down up to 60% per MW, and a $1.63B ten-year value stack per 1 GW campus. Enverus ](http://www.enverus.com/newsroom/800-vdc-rewrites-ai-data-center-power-economics?ref=nexi.fund) The most independent economic model available on this question, and the source that disagrees with NVIDIA on copper. That disagreement is the most useful number in the whole set. [ Renesas unveils the first 650V GaN in a dual-side-cooled package for megawatt-scale AI data centers The TP65H020G4PLSGBD at 20 milliohms in an 8x8mm package, 57% smaller than the TOLT part it replaces, validated in a 6kW 800V-to-48V reference design at 2.6 kW per cubic inch. Renesas Electronics ](https://www.renesas.com/en/about/newsroom/renesas-unveils-industrys-first-650v-gan-dual-side-cooled-8x8mm-package-megawatt-scale-ai-data?ref=nexi.fund) The component the schedule actually depends on. Sampling now, mass production mid-2027 — that date is the real constraint on the whole architecture.