The power architecture of data centers stands at a historic inflection point. As GPU rack power density surges towards the 600kW level, the 800VDC direct current power distribution revolution, driven by the laws of physics, has evolved from internal experiments by hyperscale cloud providers into an unavoidable systemic restructuring for the entire industry. This transformation is set to profoundly reshape the landscape of the multi-billion dollar power equipment market over the coming years.
According to a recent in-depth report from industry research firm SemiAnalysis, the core logic for 800VDC lies in physical constraints. At a rack power of 600kW, raising the distribution voltage from 54V to 800V reduces the current by approximately 15 times and cuts conductor resistive heat loss by about 219 times. This significantly reduces copper usage, lowers thermal load, and decreases conversion losses. For a 1GW IT load scale, a full migration to an 800VDC architecture could yield roughly 69MW of continuous grid power savings, translating to tens of millions of dollars in annual electricity cost reduction or an equivalent increase in new computing capacity.
SemiAnalysis forecasts that by 2030, incremental data center capacity covered by 800VDC will reach approximately 39GW. The supporting power equipment market is expected to form two core categories: the Power Rack/Sidecar market is projected to peak at around $11 billion in 2028, and the Solid-State Transformer (SST) market is estimated to reach about $13 billion by 2030. This progression will advance in four distinct stages, each corresponding to different equipment changes and supplier realignment dynamics.
**Why 800VDC is Inevitable: An Architecture Revolution Driven by Physics** Current data centers commonly use 415V or 480V three-phase alternating current (AC) distribution, with 48-54V direct current (DC) power supplied within the rack. This architecture functions adequately at today's rack power levels. However, as next-generation GPU clusters like NVIDIA's Kyber Ultra push single-rack power towards 660kW, the existing low-voltage distribution system faces three physical limits.
First is the uncontrolled weight of copper. At 48-54V, a 1MW rack requires about 200kg of copper busbars; at a 1GW scale, the total copper mass would reach hundreds of tons, exceeding feasible engineering boundaries for cost, weight, and installation complexity. Second is the encroachment of rack space by power equipment. Existing NVL72 racks already occupy up to 8 power shelves. If low-voltage solutions are maintained, the power hardware required for Kyber-level racks would fill the entire rack, leaving no space for compute units. Third, current itself becomes a bottleneck. At 600kW, 54V distribution would need to carry about 12,500A. Switching to 800V reduces the current to approximately 750A, significantly decreasing conductor cross-section and thermal stress.
SemiAnalysis notes that larger-scale compute domains mean denser racks, and denser racks necessitate a 600kW power envelope. 800VDC is the physical enabling technology that makes this power envelope possible.
**Four-Stage Migration Path: From Rack-Side Power Shelves to Solid-State Transformers** SemiAnalysis divides the 800VDC migration into four stages, spanning from 2026 to 2029 and beyond.
**Stage 1 (2026/2027): White Space Refit.** Migration is being led by Google and Meta, who have been advancing the 800VDC architecture through the OCP working group for over 18 months and, together with Microsoft, have co-developed the Diablo 400 open specification. The core equipment of this stage is the row-level HVDC power sidecar: a 42U independent cabinet that receives AC power from overhead busways, outputs 800VDC to adjacent IT racks, and integrates rectification, BBU battery backup modules, and optional supercapacitor buffering. The data center's original transformers, UPS, and switchgear remain unchanged. SemiAnalysis estimates the ASP for this power sidecar at approximately $400,000-$500,000 per unit, or about $500,000 per MW, roughly 10 times higher than the current ~$40,000 per unit for existing AC power equipment.
**Stage 2 (2027/2028): Arrival of 800VDC-Native Compute.** With the volume shipment of 800VDC-native chip systems (like Kyber racks), 800VDC shifts from a voluntary early deployment to a physical necessity. The architecture is similar to Stage 1, with a key difference: the voltage step-down point moves from the power shelf inside the IT rack to the onboard power module on the compute blade. Concurrently, centralized low-voltage UPS systems begin to phase out, with their backup functions taken over by rack-level BBUs and supercapacitors. Google and Meta have already adopted this "distributed UPS" architecture.
**Stage 3 (2028/2029): Overall Rewriting of the Electrical Architecture.** 800VDC distribution moves up from the rack row level to the facility level. Dedicated rectifiers deployed in grey space directly convert 415V AC to 800VDC, distributing it throughout the data hall via DC busways. AC distribution panels and AC floor PDUs subsequently exit the main power path. In white space, the power sidecar is replaced by a "battery rack" – which no longer performs AC-to-DC rectification but directly receives 800VDC from the grey space, retaining only the DC distribution unit, BBU shelf, and supercapacitors. SemiAnalysis estimates the content of a battery rack at about $200,000 per MW.
**Stage 4 (2029 and beyond): The Solid-State Transformer End-State.** The Solid-State Transformer (SST) directly converts medium-voltage AC to 800VDC, replacing two conversion stages (the low-voltage transformer and rectifier) with a single device. In theory, this can increase system efficiency from the current ~82% to over 87%, while achieving approximately a 40x reduction in weight and a 14x reduction in volume. SemiAnalysis projects the SST market to reach around $13 billion by 2030.
**Power Sidecar and SST: Market Opportunities for Two Core Equipment Categories** At the equipment market level, SemiAnalysis's industrial model breaks down the equipment content per MW for each 800VDC stage.
The Power Sidecar/Power Rack is the core incremental equipment for Stages 1 and 2. SemiAnalysis expects its market size to peak at approximately $11 billion in 2028 before declining as facility-level 800VDC distribution becomes widespread in Stage 3. The Diablo 400 specification, co-developed by Google, Meta, and Microsoft, has established a multi-vendor interoperability standard with participants including Delta, Advanced Energy, and TE Connectivity. Notably, NVIDIA is not adopting the Diablo 400 specification but has independently developed a 660kW single-polarity 800V reference design. The air-cooled version entered volume production in mid-2026, with the liquid-cooled version expected to begin sampling by the end of 2026.
The Solid-State Transformer (SST) is the end-state equipment for Stage 4 and is currently the most active segment for funding. According to SemiAnalysis, SST startups raised over $320 million in the 12 months leading up to March 2026. Key players include: DG Matrix (with strategic backing from ABB, a SiC supply agreement with Infineon, and the only SST product included in NVIDIA's MGX reference architecture, targeting UL certification by the end of Q2 2026); Amperesand (targeting 30MW commercial deployment by 2026); Heron Power (building a 40GW annual capacity US manufacturing base, focusing on 4.2MW direct medium-voltage input products); and Novos Power (focusing on direct medium-voltage to 800VDC conversion, claiming a 50% footprint reduction). Among traditional giants, Eaton acquired Resilient Power Systems in August 2025 to gain SST technology.
Regarding efficiency benchmarks, the best publicly available test results from ETH Zurich at INTELEC 2025 showed a 13.2kVAC to 800VDC prototype achieving 98% efficiency at 400kW. Major players like DG Matrix, Amperesand, and Heron Power claim an upper efficiency limit of 98.5%. However, real-world data center deployment requires 3-6MW class equipment to maintain over 99% efficiency under continuous load, a target no vendor has yet achieved.
**Four Key Challenges: Variables Determining 800VDC Adoption Speed** SemiAnalysis identifies four core obstacles that will directly influence the pace of 800VDC diffusion from hyperscale cloud pilot projects to the broader market.
**Regulation and Safety.** Full support for 800VDC in the US National Electrical Code (NEC) is targeted for the NEC 2029 edition. Deployments before 2029 will require custom approval from local Authorities Having Jurisdiction (AHJs) on a site-by-site basis. SemiAnalysis expects NEC 2029 to provide partial coverage, with full code maturity potentially not arriving until NEC 2032 or 2035. On safety, IEEE 1584 does not cover DC systems, and NFPA 70E lacks a Personal Protective Equipment (PPE) rating table for 600-1000VDC. UL Solutions has launched a DC Safety Research Consortium to address this gap.
**Cooling and Auxiliary AC Loads.** The cooling system is the largest AC load in an 800VDC data center. Currently, no vendor offers a complete DC-native cooling ecosystem. At GTC 2026, Delta launched a 2.4MW in-row cooling distribution unit supporting 800VDC, marking the first major cooling component designed for DC-native operation. However, the complete cooling stack (chillers, compressors, pumps, building controls) still relies on AC power. At the OCP Global Summit 2025, NVIDIA explicitly stated that its 800VDC reference architecture will retain an AC auxiliary bus.
**Supply Chain and Lagging Standards.** Technological innovation in DC distribution is outpacing standard development. For example, the UL 857 standard for busways only raised its coverage limit from 600V to 1000VDC in its 14th edition, published in 2025; the 15th edition is still under development. As of May 2026, no SST vendor has completed UL certification for data center deployment. The OCP working group is coordinating with regulators to establish initial standards by the end of 2026.
**Grid Interconnection and Regulatory Pressure.** In May 2026, NERC issued a highest-level (Level 3) Essential Action Alert covering large computing loads, with a mandatory response deadline of August 3, 2026, and proposed a "Computing Load Entity" registration system for data centers consuming over 1MW. ERCOT's NOGRR282 introduced new voltage and frequency ride-through requirements and mandated that large loads submit PSS/E and PSCAD electromagnetic transient models. SemiAnalysis points out that the grid response behavior of 800VDC facilities, determined by SST control algorithms, energy storage state of charge, and GPU load curves, is far more complex than that of traditional AC data centers. This is giving rise to new AI-native EPC service providers like Aran Industries.
**Total Electrical Cost Remains Largely Stable, Content Structure Undergoes Deep Reshuffle** SemiAnalysis's model indicates that across all four stages, the total electrical equipment cost per MW for a data center remains within the $3.6 to $4.8 million range, with overall magnitude change being limited. What truly changes is the structural migration of equipment content: the value of grey space equipment contracts as centralized UPS systems (approx. $1.2 million per MW) phase out. The value of white space equipment peaks in Stage 1 with the introduction of HVDC power sidecars and climbs again in Stage 4 as SSTs (approx. $1.0-$1.5 million per MW) replace low-voltage transformers and rectifiers.
The path for efficiency improvement is also clear: the baseline AC architecture has an end-to-end efficiency of ~82%, jumping to 86.5% in Stage 2 (eliminating UPS double conversion), reaching 86.9% in Stage 3, and achieving 87.4% in Stage 4. For a 1GW IT load, the ~5 percentage point efficiency gain in Stage 4 relative to baseline aligns with data publicly cited by NVIDIA.
For investors, the core logic of this transformation is that the total market size is not expanding dramatically. However, the reallocation of equipment content will profoundly alter the revenue trajectories of existing suppliers—some traditional equipment categories will be compressed or even eliminated, while market space for emerging categories is rapidly opening up.
Comments