Data Center Power: The Transition to 800 VDC

Walking through the show floor at the recent Data Center World event in Washington, DC, felt a bit like walking through the plumbing and electrical aisles at Home Depot, a reaction to the ever-escalating power needs of today’s AI accelerators.

For the past two decades, power architecture has evolved incrementally around a familiar model: utility AC enters the facility, passes through switchgear, transformers, UPS systems, power distribution units, and rack power shelves, and is ultimately converted to low-voltage DC for CPUs, GPUs, memory, networking, and storage.

That model worked for conventional enterprise and cloud computing. It was good enough when racks consumed 10kW, 20kW, or even 50kW, but it becomes far more difficult as AI infrastructure pushes rack densities toward hundreds of kilowatts and, in some designs, beyond 1MW.

To address this, NVIDIA established 800VDC as the reference architecture for its forthcoming Vera Rubin Ultra Kyber rack system. NVIDIA drives the agenda for AI, so it’s no surprise to see the broader ecosystem of infrastructure vendors, power semiconductor suppliers, hyperscalers, and OEMs aligning behind this standard.

The benefits of transitioning to 800VDC are very real. At 1 MW per rack, for example, a 54VDC distribution system requires up to 200 kilograms of copper busbar per rack and consumes rack space. 800VDC eliminates both constraints, reduces conversion losses, and simplifies the power chain.

The change is already happening:

• CoreWeave, Lambda, Nebius , Oracle Cloud Infrastructure, and Together AI are designing facilities for 800VDC distribution.
• Foxconn‘s 40MW Kaohsiung-1 data center in Taiwan is operational on 800 VDC.
• Vertiv, Schneider Electric, Eaton, and Delta Power Group have announced commercial products that will be available in H2 2026, aligned with the Kyber rack launch timeline.
• Power semiconductor vendors, including Infineon Technologies, STMicroelectronics, Texas Instruments, and Navitas Semiconductor, have validated components and reference designs for the NVIDIA MGX ecosystem.

Why 800VDC?

The simple answer is that higher voltage allows data centers to deliver more power with less current.

This matters because electrical losses increase with current. When power demand rises sharply, low-voltage distribution requires very high current, which in turn requires larger conductors, heavier copper busbars, greater heat dissipation, larger conversion hardware, and more physical space occupied by power delivery rather than revenue-generating compute.

Moving to 800 VDC reduces the current at a given power level. That can lower resistive losses, reduce copper requirements, shrink power-distribution hardware, and improve the economics of very high-density racks.

To put some numbers to it, traditional data center racks may draw around 10kW, while AI racks are beginning to approach 1MW, making conversion losses, current levels, and copper requirements much harder to justify.

Switching from 415VAC to 800VDC also delivers significantly more power through the same conductor size, while reducing copper use and improving efficiency at scale.

NVIDIA says that 800VDC transmits over 150 percent more power through the same copper cross-section than 54VDC configurations. That’s significant.

The other reason is conversion efficiency. Most IT equipment ultimately consumes DC power, even if the facility distributes AC through much of the chain. Traditional architectures often convert AC to DC for UPS battery storage, back to AC for distribution, and then back again to DC near the server.

Each conversion adds loss, heat, equipment, and complexity. An 800VDC architecture can remove some of those intermediate steps and bring high-voltage DC closer to the compute load.

800VDC Architecture and Conversion Chain

800VDC restructures the data center power chain from the grid to the GPU:

• Grid connection: Medium-voltage AC (typically 13.8 kV) is converted to 800 VDC at the facility perimeter using a solid-state transformer (SST) or a transformer-rectifier unit (TRU), thereby eliminating most intermediate conversion stages.
• Distribution: 800VDC is delivered via liquid-cooled busbars and conductors to each rack. The copper wire cross-section is reduced by up to 45 percent compared with low-voltage DC configurations.
• Rack-level conversion: Server boards accept 800VDC directly and perform local DC-to-DC conversion to supply GPU core voltages. Power shelves are eliminated from the compute rack, freeing 64U or more of usable rack space.
• Sidecar model (near-term): For initial deployments and existing facilities, a dedicated power rack adjacent to each compute rack houses the AC-to-800VDC conversion equipment. This approach preserves the existing upstream AC infrastructure while delivering 800VDC at the rack.
• Energy storage: The Vera Rubin NVL144 rack provides 20 times more energy storage than prior generations, stabilizing power delivery across variable GPU load profiles.

NVIDIA claims the architecture improves end-to-end power efficiency by up to 5 percent compared to 54VDC systems and reduces maintenance costs by 70 percent by eliminating rack-level PSU complexity.

The Convergence of Power & Cooling

The 800VDC transition is also happening alongside the move to direct liquid cooling, rear-door heat exchangers, coolant distribution units, and more-integrated thermal systems.

While 800VDC itself does not require a fundamental change to cooling infrastructure beyond what high-density GPU deployments have already imposed, a high-density AI rack requires coordinated delivery of power, heat removal, telemetry, protection, serviceability, and redundancy.

Delta, for example, has announced 800VDC-aligned power, cooling, and microgrid solutions for NVIDIA AI factory architectures, including a 2.4MW liquid-to-liquid CDU for 800VDC environments.

Schneider Electric has also framed 800VDC adoption around both power and cooling readiness for future NVIDIA platform evolution.

This convergence is also reshaping vendor strategy. Power vendors, cooling vendors, silicon providers, rack integrators, and data center operators now must coordinate earlier in the design cycle. The traditional model of independently selecting servers, racks, power distribution, and cooling equipment is harder to sustain at an AI-factory scale.

Energy Storage

One of the less obvious drivers behind 800VDC is the associated energy storage architecture.

Traditional data centers often place battery systems within or alongside UPS architectures tied to AC distribution. In an 800VDC model, energy storage can potentially connect more directly to the DC bus, enabling more flexible placement at the facility, pod, row, or rack-adjacent level.

NVIDIA has argued that 800VDC makes it easier to place storage where it best fits the workload and facility design.

Schneider Electric has also highlighted energy storage placement as a design variable in 800VDC architectures, with options ranging from centralized facility storage to pod-level or rack-adjacent storage.

This matters for AI data centers because the workload profile can be extremely dynamic, with training, inference, checkpointing, networking bursts, and cooling demand creating highly variable power behavior.

As AI facilities scale toward hundreds of megawatts or more, operators will increasingly need power architectures that support buffering, fast response, grid interaction, and potentially microgrid integration.

800VDC does not solve grid scarcity. But it may give operators more tools to manage the power they can obtain.

Impact to DC Operators

For operators, the 800VDC transition introduces both opportunity and complexity.

The opportunity is higher density, better electrical efficiency, and potentially lower material cost per megawatt delivered. That is highly attractive for AI campuses, hyperscale facilities, sovereign AI infrastructure, and specialized GPU cloud providers, where power delivery is now a primary business constraint.

The complexity is operational. High-voltage DC requires new protection schemes, safety practices, maintenance procedures, and workforce training. DC arcs behave differently from AC arcs, making serviceability a major design consideration.

Path to Deployment

Data center operators face a two-phase transition with distinct planning and investment requirements in each phase:

The near-term phase, enabled by the sidecar model, allows operators to deploy 800VDC compute infrastructure without replacing upstream AC distribution, switchgear, or UPS systems. This path mirrors the approach used when liquid cooling was introduced into air-cooled facilities and provides a manageable on-ramp for operators whose existing facilities were not designed for 800VDC.

The longer-term phase, which converts medium-voltage grid power directly to 800VDC at the facility perimeter, requires facility-level redesign and has a multi-year lead time.

Key considerations for operators include:

• Procurement timeline: Vertiv’s 800VDC MGX ecosystem targets commercial availability in H2 2026, aligned with Kyber deployments. Operators need to initiate vendor qualification and reference architecture reviews immediately to avoid procurement delays against the 2027 Rubin Ultra Kyber production ramp.
• Electrical safety: High-voltage DC exhibits arc-flash and fault characteristics distinct from those of AC systems. Operators require updated electrical certifications, revised safety procedures, and staff training before commissioning 800VDC infrastructure.
• Hot-swap capability: At a 30x cost premium over traditional server hardware, minimizing maintenance downtime is a TCO requirement. Infineon’s CoolSiC JFET-based hot-swap solution enables board exchange while adjacent servers in the same rack remain operational; operators should qualify this capability as part of vendor selection.
• Stranded asset risk: Operators who commit to large-scale traditional UPS and PDU infrastructure over the next 12 to 18 months face stranded-cost exposure as the industry transitions to DC UPS and SST topologies. Modular sidecar deployment mitigates near-term risk while preserving optionality for the full DC transition.
• Facility power capacity: 800VDC, and the associated compute density increase the total power draw per square foot. Operators must assess facility power capacity, transformer ratings, and grid connection agreements against projected rack densities before committing to Kyber deployments.

Vendor Ecosystem Impact

The move to 800VDC creates new competitive openings across the infrastructure stack:

• Power infrastructure vendors gain strategic relevance because power delivery is now a gating factor for AI deployment.
• Component vendors gain a larger role as high-voltage controllers, converters, GaN and SiC devices, protection systems, and point-of-load designs become central to rack-scale performance.
• Cooling vendors become more tightly linked to power design because heat removal must scale with power density.
• Server and rack vendors must integrate more deeply with facility infrastructure.

This also changes the buying conversation. For AI infrastructure, buyers will increasingly evaluate vendors not only on accelerator performance but also on power-path efficiency, rack-level serviceability, cooling integration, deployment repeatability, and ecosystem maturity.

That is a significant shift. In the AI data center, power architecture becomes part of the compute architecture.

Power Semiconductors

800VDC requires changes to the power semiconductor stack across the conversion chain. Silicon MOSFETs used in traditional 54VDC power supplies can’t handle 800V bus voltages.

Two wide-bandgap semiconductor materials address this requirement:

• Gallium Nitride (GaN): Enables high switching frequencies at 800V, reducing passive component size and enabling compact power delivery boards. STMicroelectronics’ NVIDIA-validated 12-kW GaN power-delivery board achieves a power density of 2,500 W/in3. GaN is the primary technology for the AC-to-800VDC conversion stage and for server board power delivery.
• Silicon Carbide (SiC): Preferred for hot-swap controllers and medium-voltage conversion due to its high breakdown voltage and thermal characteristics. Infineon’s CoolSiC JFET technology enables board-level hot-swap on an active 800VDC bus, a critical operational requirement given the 30x cost premium of AI server hardware over traditional servers.
• Texas Instruments unveiled a complete 800VDC reference architecture at NVIDIA GTC 2026 covering solid-state transformers, sidecar racks, server IT racks, and cooling distribution units.
• STMicroelectronics‘ 12 kW board entered production testing after NVIDIA validation, with a 20-kW variant targeting low-power device rails also announced.

Vendor Strategies

Infrastructure vendors are adopting distinct strategies:

• Vertiv and Schneider Electric are full-ecosystem providers, offering integrated power and cooling architectures rather than point products. Both vendors have existing relationships with hyperscale operators and can leverage those accounts for initial 800VDC deployments.
• Eaton‘s focus on the medium-voltage solid-state transformer positions it at the facility perimeter rather than at the rack, targeting the longer-term full-DC conversion phase. This is a differentiated but higher-latency opportunity.
• Delta‘s embedded battery backup in its power rack addresses energy storage and power continuity requirements in a single product, potentially reducing the total number of infrastructure components operators need to qualify for.

Power semiconductor vendors face the most direct competitive pressure, as the wide-bandgap transition opens the market to new entrants.

Established analog semiconductor vendors such as Texas Instruments and Infineon have responded by publishing complete reference designs validated within the NVIDIA ecosystem, thereby raising the barrier to entry for less-resourced competitors.

Competitive Dynamics

Some of the competitive dynamics worth watching include:

• Vertiv vs. Schneider Electric: Both are pursuing integrated power-and-cooling reference architectures and have announced 800VDC products aligned with Kyber timelines. Differentiation will come from deployment experience, service capabilities, and hyperscale relationships rather than product specifications at this stage.
• Infineon, STMicroelectronics, and Texas Instruments: All three have NVIDIA-validated components or reference designs. Infineon’s CoolSiC JFET hot-swap solution addresses a capability gap that neither ST nor TI has prominently featured, giving Infineon a differentiated position in the operational reliability segment.
• GaN vs. SiC: The two semiconductor materials serve different segments of the power chain and are largely complementary rather than competitive. GaN excels in high-frequency, high-density conversion stages; SiC excels in high-voltage, high-reliability applications such as hot-swap and medium-voltage conversion. Vendors with both GaN and SiC portfolios, including Infineon and Navitas, have a broader addressable opportunity.
• China market: Higher-voltage DC data centers have already emerged in China, showing that some Chinese hyperscalers are ahead of Western operators in adopting DC architectures. This creates a favorable reference base for Chinese component vendors competing against established Western suppliers in global markets.
• Incumbent AC infrastructure vendors: Traditional AC UPS and PDU vendors without 800VDC roadmaps face displacement risk as greenfield AI factory construction standardizes on DC distribution. The transition timeline gives incumbent vendors 18 to 24 months to introduce competitive products before the AI factory build cycle accelerates.

Final Thoughts

For years, the industry optimized around PUE, virtualization efficiency, cloud scale, and incremental improvements in server density. AI changes the unit of design. The relevant unit is now a systems-level consideration, in which accelerators, memory, interconnect, power, cooling, storage, networking, software, and operations function as a tightly coupled platform.

Despite the momentum, 800VDC will not become universal immediately.

Conventional enterprise data centers do not require 800VDC for typical virtualized workloads, SaaS infrastructure, storage arrays, or general-purpose private cloud. Many facilities will continue to use existing AC distribution and lower-voltage rack architectures because the economics do not justify a major redesign.

The strongest adoption will come from environments where rack density and energy economics dominate the business case:

• AI training clusters
• Large-scale inference platforms
• Neoclouds
• Hyperscale AI campuses
• National or sovereign AI infrastructure
• Facilities designed around liquid-cooled accelerator platforms
• New-build data centers where power and cooling can be engineered together from the start

Retrofits will be more selective. Operators will seek out modular approaches, such as rack-sidecars or pod-level power systems, that enable them to introduce 800VDC without rebuilding the entire electrical chain.

From an ecosystem perspective, Vertiv, Schneider Electric, Eaton, Delta, Infineon, STMicroelectronics, and Texas Instruments all have commercial products or validated reference designs either in the market or in development. The question for operators is how quickly to deploy and via which deployment path.

For hyperscalers, GPU clouds, and AI factory operators, 800VDC is likely to become part of the next high-density design baseline. For the broader enterprise market, it is a preview of where infrastructure is headed as AI workloads continue to move from pilot projects to production-scale systems.

The practical message for enterprises is not that every data center needs to convert to 800VDC. Most do not. The message is that AI infrastructure planning can no longer treat power as a facility afterthought. Power architecture is becoming a first-order determinant of AI capacity, cost, reliability, and time-to-deployment.