Electrical equipment implications of Nvidia 800VDC data-center architecture Report Interpretation

Overview

The report summarizes an expert webcast on how Nvidia's proposed 800VDC data-center architecture could affect electrical-equipment demand and competitive positioning. Its central conclusion is that adoption is progressing more slowly than initially expected, while conventional data-center power systems should remain a substantial and potentially higher-margin market for many years.

Core views

The expert expects 800VDC power architecture to develop more slowly than anticipated earlier in 2026 because AI inference is growing faster than expected and has materially lower power-density requirements than AI training. Goldman Sachs now cites expectations for 800VDC to account for 21% of US new-build data centers in 2030, revised down from 25%, and 17% in Europe, down from 20%. Inference requires up to 150kW per rack at maximum, allowing traditional architecture to remain viable; the expert therefore considers 800VDC highly unlikely for most inference workloads, where the more likely upgrade is liquid cooling rather than a wholesale conversion of power architecture. The transition is also expected to be incremental rather than immediate. Full 800VDC installations are not anticipated for at least five years. Near-term deployments are more likely to use 800VDC for servers while retaining conventional systems for networking, storage and lighting. Nvidia's Kyber NVL144 racks, designed for Rubin Ultra GPUs, have reportedly been pushed to 2028 from the original 2027 target, reinforcing the slower timing. The report further notes that enterprise customers run most inference workloads and are becoming more focused on cost efficiency rather than maximizing token output; outside specialized compute-intensive uses such as drug discovery, typical enterprise applications may not need very dense AI clusters. The delayed transition does not weaken the underlying data-center equipment opportunity in the report's view. The expert expects non-AI data centers to remain dominant in 2032. Microsoft is cited as an example, with plans targeting as much as 38GW of installed power by 2032, equivalent to roughly 20% CAGR over the next six years. Non-AI facilities are expected to comprise two-thirds of that installed base in 2032, versus about 83% currently, still implying 17% CAGR; AI data centers are expected to rise from roughly 17% to one-third of the installed base, implying 35% CAGR. This leaves a large addressable market for conventional power architecture, including suppliers such as Legrand that have not announced solid-state transformer development. The expert also expects traditional equipment areas to retain higher margins than 800VDC architecture. Manufacturers are pursuing different routes toward 800VDC. Delta Electronics is conducting SST pilots in Asia and targeting a 5MW design, which the expert regards as a favorable balance of scale and flexibility. Eaton plans to sample a 2MW SST with customers by end-2026, drawing on its 2025 acquisition of Resilient Power. Schneider Electric is offering transformer rectifier units and DC rectifiers, while Vertiv and ABB are promoting medium-voltage UPS products that remove some conversion steps but still use conventional technology. GE Vernova has announced a 175MW MV UPS, compared with ABB's 1-2MW offering, while continuing to target SST introduction in 2027. Heron Power is developing a 5MW SST and expects its new production facility to be ready by end-2027. The report identifies meaningful technical and commercial hurdles to an SST ramp. Solar entrants Enphase and SolarEdge may draw on DC-distribution expertise, but the expert questions whether they can meet mission-critical data-center requirements for redundancy and field service. Hyperscalers are issuing large SST requests for proposals primarily to test pricing, volumes and competing technologies. Current SST prices remain about $400,000-$500,000 per MW because manufacturing scale is limited; after product selection, evaluation and ordering could take a further 12-18 months, making 2029 the earliest likely window for broad deployment. Attention is shifting toward other components required by 800VDC, particularly solid-state breakers. Conventional circuit breakers operate in milliseconds, whereas 800VDC requires microsecond operation. The expert expects this to be technically difficult and costly, at around $600,000 per MW, roughly five times current breaker pricing and above SST pricing. Higher technology content is expected to lift total powertrain content per MW by 2030 to about $4 million for AI training, from $1.7-$1.8 million in 2025, and to $2.6 million for AI inference, from $1.4 million in 2025.

Analysis framework

The report uses an expert-call assessment to link AI training and inference power-density needs with likely data-center architecture choices. It then compares adoption forecasts, installed-power growth, product roadmaps, component costs and vendor approaches to assess timing, competitive implications and equipment content per MW.

Methodology notes

• Industry AnalysisSupply-demand framework

  Data-center power demand and architecture adoption analysis

  The report compares demand from AI training, AI inference and non-AI data centers with the technical suitability, timing and cost of conventional and 800VDC power systems.

• Industry AnalysisVolume-price decomposition

  Installed-power growth and equipment content per MW

  The analysis combines projected installed power capacity with rising dollar content per MW to explain the revenue opportunity for electrical-equipment suppliers.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Nvidia (NVDA)

  Its proposed 800VDC architecture and Kyber NVL144 rack timing are key drivers of the industry transition.

  Strengths

  Its platform roadmap influences data-center power-architecture development.

  Weaknesses

  Kyber NVL144 is reported to have moved to 2028 from the original 2027 target.

  Risks

  A slower architecture transition could delay associated equipment adoption.

• Legrand

  A potential beneficiary of sustained demand for traditional data-center power architecture.

  Strengths

  Can address the large conventional-architecture market.

  Weaknesses

  Has not announced solid-state transformer development.

  Comparison

  Traditional equipment is expected to remain more profitable than 800VDC architecture.

• Eaton

  Developing an SST offering for the emerging 800VDC opportunity.

  Strengths

  Plans to sample a 2MW SST with customers by end-2026 and can leverage its 2025 Resilient Power acquisition.

  Weaknesses

  Commercial adoption remains dependent on customer testing and the broader timing of SST deployment.

  Comparison

  Its planned 2MW SST is smaller than the 5MW designs being pursued by Delta Electronics and Heron Power.

• GE Vernova (GEV)

  Developing medium-voltage UPS products while remaining vocal on SST development.

  Strengths

  Has announced a 175MW MV UPS and expects SST introduction in 2027.

  Weaknesses

  The broader SST market is not expected to see mass deployment before 2029.

  Comparison

  Its announced MV UPS scale is larger than ABB's cited 1-2MW offering.

Key data

• US 800VDC share of new-build data centers in 203021%Revised down from 25%.
• European 800VDC share of new-build data centers in 203017%Revised down from 20%.
• AI inference rack power requirementUp to 150kW per rackSupports continued use of traditional architecture for inference.
• Microsoft targeted installed power by 2032Up to 38GWEquivalent to about 20% CAGR over the next six years.
• Non-AI data-center share of 2032 installed baseTwo-thirdsDown from approximately 83%, but still implying 17% CAGR.
• AI data-center share of 2032 installed baseOne-thirdUp from approximately 17%, implying 35% CAGR.
• Current SST pricingApproximately $400,000-$500,000 per MWHigh pricing reflects limited manufacturing scale.
• Solid-state breaker pricingApproximately $600,000 per MWAbout five times the price of current circuit breakers.
• AI-training powertrain content per MW in 2030Approximately $4 millionUp from $1.7-$1.8 million per MW in 2025.
• AI-inference powertrain content per MW in 2030Approximately $2.6 millionUp from $1.4 million per MW in 2025.

Impact & implications

The report argues that slower 800VDC deployment extends the opportunity for conventional power equipment while creating a longer-dated upgrade cycle for SSTs, MV UPS, rectification equipment and solid-state breakers. AI training may drive faster capacity growth and higher content per MW, but non-AI and inference workloads should continue to support conventional architecture at scale.

Risks

• SST deployment may be delayed by high costs, limited manufacturing scale, vendor testing and a further 12-18 month evaluation-and-ordering cycle after product selection.
• Solar entrants face barriers in meeting mission-critical data-center requirements for redundancy and field-service availability.
• Solid-state breakers must achieve microsecond operation and are expected to be expensive, creating a challenging development hurdle.

What to watch

• The OCP Summit on 12-15 October.
• Nvidia GTC on 20-22 October.
• Hyperscaler SST pilot outcomes, preferred-product decisions and resulting order timing.
• Development of second-generation SSTs, which the expert expects to become a greater source of product differentiation in four to five years.