AI racks are moving toward 1 MW and beyond. At this level, traditional 54 VDC power shelves can take up to 64U of rack space in an NVIDIA Kyber rack. That leaves less room for the compute hardware that matters most.
Two different answers are now emerging. NVIDIA is pushing a monopolar 800 VDC design. The Open Compute Project (OCP) has also explored the bipolar ±400 VDC approach, known as Mt Diablo, with support from major hyperscalers. As AI rack power rises both approaches aim to make power delivery easier.
So, which one has the stronger future? This article looks at how both designs work, where they fit, and why 800 VDC may have the edge. It also explains how 800V DC could change power delivery in an AI data center.
What Are 800 VDC and ±400 VDC?
800 VDC: NVIDIA's Monopolar Approach
The 800 VDC design uses one 800 V conductor and a return. NVIDIA's reference architecture converts medium-voltage AC from the grid to 800 VDC at the data center perimeter. The DC power can then be distributed to AI racks with fewer conversion stages.
NVIDIA expects its MGX-compatible 800 VDC power rack to arrive in the second half of 2026. Its Kyber rack architecture is expected to follow in 2027. More than 80 companies are now working on products that support the wider 800 VDC ecosystem.
±400 VDC: The OCP Mt Diablo Bipolar Approach
The ±400 VDC design uses two poles. One sits at +400 V and the other at ?400 V. This gives 800 V between the two poles, while each pole is only 400 V from ground.
The approach grew from OCP's Diablo 400 work, involving companies such as Google, Meta, and Microsoft. One advantage is that it can make use of parts already developed for the 400 V EV supply chain. This could help with component availability and cost.
800 VDC vs ±400 VDC: Side-by-Side Comparison
Both designs are built to handle the rising power needs of AI racks. However, they take different approaches to voltage, grounding, components, and deployment.
| Factor | 800 VDC | ±400 VDC |
| Topology | Monopolar | Bipolar |
| Voltage to ground | 800 V | 400 V |
| Main backers | NVIDIA, Vertiv, Eaton, ABB, Delta | Google, Meta, Microsoft (OCP) |
| Component ecosystem | New data center parts, including SST, SiC and GaN | Draws on EV-derived parts |
| Best fit | Greenfield AI factories, 1 MW racks | Hyperscaler in-house designs |
| Timeline | MGX H2 2026, Kyber 2027 | OCP specifications in progress |
Both approaches carry the same 800 V of potential difference. The main difference is how that voltage is referenced to ground.
The Case for 800 VDC
The biggest argument for 800 VDC is efficiency. Enverus Intelligence Research estimates that an 800 VDC architecture can raise end-to-end facility efficiency from 79.8% to 93.8% compared with 415 VAC. It also estimates a 13% reduction in electrical capital expenditure.
There is also a major copper benefit. Enverus estimates that copper use per MW can fall by up to 60%. Navitas estimates that copper conductor thickness can be reduced by up to 45%. Higher voltage means lower current for the same power, which helps reduce the amount of copper needed.
The path to the GPU can also be simpler. NVIDIA's Kyber design uses a 64:1 LLC converter to take 800 VDC down to about 12 VDC close to the GPU. This reduces conversion stages and saves valuable rack space.
The supplier ecosystem is also growing quickly. Companies are developing products such as ABB Infinitus, Eaton MVSST 2.0 and Flex power rack solutions. Solid-state transformers are another important part of this shift.
The Case for ±400 VDC
The strongest argument for ±400 VDC is safety. Each pole is only 400 V from ground. That can make personnel protection easier than a system where the full 800 V sits relative to ground.
However, DC safety standards are still developing. Existing rules were largely built around AC systems. Schneider Electric notes that NFPA 70E and IEEE 1584 do not fully capture the behavior of converter-based 800 VDC systems.
Cost is another advantage. A ±400 V system can draw from a large ecosystem of 400 V-class components already used in electric vehicles and other applications. These parts are produced at scale.
There is also a strategic benefit for hyperscalers. Companies such as Google, Meta and Microsoft have the resources to develop and control their own power designs.
The trade-off is that two poles can add some complexity to wiring, protection and system coordination.
Is the Industry Converging on One Standard?
The industry now appears to be moving toward a common 800 VDC framework. In August 2026, Google, Microsoft and NVIDIA announced work through OCP to establish common 800 VDC requirements. More than 80 partners are already developing compatible infrastructure.
This does not mean every data center will use exactly the same design. OCP's approach allows different deployment paths based on the needs of each facility.
Safety is another important step. Schneider Electric's 2026 study found that 800 VDC arc-flash risks can be managed with suitable protection and modeling. In some cases, the results were comparable to or lower than typical AC systems.
UL Solutions, along with further safety standards and solid-state transformer certification, will still play an important role before wider adoption. These developments will be especially important for any high-voltage DC data center moving toward large-scale 800 V power distribution.
Verdict: Which Data Center Power Standard Will Win?
The short answer is 800 VDC.
It is increasingly becoming the reference architecture for next-generation AI factories. NVIDIA, Google and Microsoft are now working through OCP around common 800 VDC requirements. This gives the approach strong industry support.
That does not mean ±400 VDC will disappear. It may continue in some hyperscaler-built facilities where existing designs and component choices make it attractive.
The transition is also unlikely to happen overnight. The industry will move through several stages. AC rack supply will come first. Sidecar power systems can then support existing sites. Hybrid designs will follow. Fully native 800 VDC facilities are expected to become more common from around 2029.
The market is already reflecting this shift. The global 800 VDC data center power architecture market was valued at USD 16.24 billion in 2025. It is projected to reach USD 59.76 billion by 2034, growing at a 15.6% CAGR.
Final Takeaway
AI rack power is rising too quickly for older power architectures to remain the only option. Operators should prepare for 800 V-class DC, whether they choose a monopolar or bipolar design. Brownfield sites can start with sidecar systems and move toward more native designs over time.
For businesses planning their next AI facility, understanding data center power architecture is becoming just as important as choosing GPUs, cooling systems, and networking equipment.
Download a free sample of the 800 VDC Data Center Power Architecture Market report.