NVIDIA 800V vs OCP Mt Diablo

COMPARISON · AI POWER REFERENCE ARCHITECTURE

NVIDIA 800V vs OCP Mt Diablo: two references, not two architectures

NVIDIA's 800V HVDC architecture and OCP Mt. Diablo v0.7.0 are frequently discussed as competing "architectures" for AI data centre power. They are not. NVIDIA describes what its next-generation reference designs assume at the FACILITY feed (Layer 2-3 upstream); Mt. Diablo specifies how the RACK-LEVEL (Layer 4) is built to accept that feed. Both are needed together.

01Bottom-line verdict

They are complementary, not competing. NVIDIA's 800V HVDC guidance addresses facility-side architecture assumptions (what NVIDIA reference AI systems expect at the building distribution level). OCP Mt. Diablo v0.7.0 addresses rack-side architecture (how the Layer 4 sidecar rack is built + connected). A hyperscaler-scale AI facility for NVIDIA reference designs typically uses both together. The confusion arises because both use "800V DC" in the description but refer to different layers.

02Comparison table

NVIDIA 800V HVDCOCP Mt. Diablo v0.7.0
Author / forumNVIDIA reference architecture guidanceOCP Rack + Power working group
Layer addressedLayer 2-3 (facility distribution)Layer 4 (rack-level power conversion)
Voltage scopeMV-to-800V DC facility feed800V DC to 48V DC rack conversion
What is specifiedReference architecture pattern + design assumptionsPhysical + electrical + mechanical rack spec
GovernanceNVIDIA-published guidance (not open standard)Open Compute Project, 80+ ecosystem partners
First publishedNVIDIA blog + reference designs 2025-2026Mt. Diablo v0.7.0 March 2026
Compatible with each otherYes; the facility feed pattern matches Mt. Diablo rack inputYes; rack accepts 800V DC facility feed
Vendor ecosystemFlex + Delta + Vertiv + others building to NVIDIA referenceWiwynn + Foxconn + Inventec + Quanta + others build Mt Diablo racks

03Decision framework

Choose NVIDIA 800V HVDC when

  • Building an AI facility for NVIDIA next-gen accelerators. The facility-side reference architecture is NVIDIA-led
  • Working with NVIDIA reference-design integrator. Their design assumes NVIDIA facility-feed guidance
  • Evaluating specific product compatibility. NVIDIA guidance covers what accelerator systems expect

Choose OCP Mt. Diablo v0.7.0 when

  • Specifying the Layer 4 rack + power shelf. Mt. Diablo is the open spec for this layer
  • Sourcing from OCP-ecosystem rack OEMs (Wiwynn, Foxconn, etc.). Their products are Mt. Diablo-compliant
  • Building multi-vendor procurement strategy. Open spec enables cross-vendor compatibility
  • Accelerator-agnostic infrastructure planning. Mt. Diablo works with non-NVIDIA silicon too

04Deep-dive research

05Primary sources

06Frequently asked

Do I have to choose one?

No. If you're building NVIDIA-based AI infrastructure at hyperscaler scale you typically use both together. NVIDIA guides Layer 2-3; Mt. Diablo specifies Layer 4.

What if I'm not building for NVIDIA accelerators?

Mt. Diablo still applies as the Layer 4 open standard. The facility-side guidance may come from OCP or the specific accelerator vendor (AMD, Google TPU, etc.).

Is there tension between NVIDIA-led guidance and OCP open spec?

Practical answer: not much in the current 2026 window. Both are converging on 800V DC + sidecar rack pattern. The tension is more governance-shape (NVIDIA-published vs open-standard) than technical.

What is the future beyond Mt. Diablo v0.7.0?

1500V DC (facility feed) + higher rack density are in scope for future OCP work. See Supplement C on the voltage ceiling.

Definition The comparison between NVIDIA's 800V DC data centre power specification and the Open Compute Project Mount Diablo v0.5 specification. NVIDIA specifies a unipolar 800V rail for its Blackwell and Rubin generation platforms, engineered around greenfield hyperscaler facility design. OCP Mount Diablo v0.5 specifies a bipolar ±400V rail sponsored by Microsoft, Meta and Google in the March 2026 draft, engineered for retrofit into existing 415V AC facility distribution. The two architectures diverge on grounding scheme, fault-clearing behaviour and qualified vendor stack.