OCP rack standards now stretch across at least three separate specifications – power delivery, mechanical form factor, and firmware recovery – and none of them alone makes two vendors’ AI racks interchangeable. The clearest evidence: AMD and Schneider Electric needed a dedicated reference design, published July 23, 2026, just to map the Helios rack onto real facility power and cooling, splitting the work into facility power, facility cooling, IT space, and lifecycle software as four technical areas.
That fragmentation tracks a broader shift TrendForce describes: AI computing’s core efficiency metric is moving from individual chip performance to rack-level performance per watt and performance per total cost of ownership, pushing chip vendors into rack-and-system design rather than pure silicon.
Quick take
OCP now covers rack power (ORV3, Sidecar), mechanical form factor (ORW), and firmware recovery as separate specifications, so an “OCP-compliant” claim does not by itself say which of these a given rack actually follows.
AMD’s Helios rack follows Meta’s ORW open rack-wide design but still shipped with a dedicated vendor reference blueprint for facility power and cooling on July 23, 2026.
Cabinet vendors like Legrand still bolt 19-inch EIA/ECA-310 rails into OCP-based structures because AI hardware routinely exceeds standard OCP dimensions.
Google, AWS, and Microsoft are building proprietary scale-up interconnects instead of adopting OCP racks wholesale, so OCP coverage applies mainly to Meta-aligned and merchant-silicon deployments.
OCP rack standards cover more ground than they used to
The Open Rack V3 (ORV3) standard replaces per-server AC cabling with a centralized power shelf and a 48V DC busbar running down the back of the rack, letting server sleds blind-mate into shared power instead of routing copper to each node.
ORV3’s busbar approach becomes necessary once a rack needs to sustain loads above 40 kW, because AC cabling at that density gets thick enough to block airflow and adds transmission losses.
Separately, OCP’s Sidecar architecture targets higher voltages for disaggregated power racks, and ABB’s July 17, 2026 whitepaper compares 800V DC against ±400V DC topologies specifically for grounding, protection, and UL/NEC compliance in AI-driven facilities.
A third OCP track covers firmware rather than power: NVIDIA’s code-management stack implements the OCP Recovery Protocol to query GPU health and force firmware recovery over I2C or USB-to-I2C bridges when a device becomes unresponsive.
None of these three tracks – ORV3 power, Sidecar high-voltage DC, and OCP Recovery firmware – guarantees compatibility with the others inside the same deployment; each solves a different problem at a different layer of the rack.
Why “OCP-compliant” still needs a vendor blueprint
Meta submitted the open rack-wide (ORW) form factor to OCP at the OCP Global Summit in San Jose, and AMD’s earlier Helios unveiling described MI450 series GPUs built on CDNA architecture with up to 432 GB of HBM4 memory per GPU on that blueprint.
The later Helios rack that the July 23, 2026 Schneider Electric reference design actually covers is integrated with Zen 6 Eypc CPUs, MI400 GPUs, and Vulcano NICs in a double-wide format, a different configuration than the earlier announcement described.
That reference design supports AI clusters up to 10.4MW at new-build sites and individual rack densities up to 246kW, using Schneider’s Motivair division for CDU-based and hybrid air-liquid cooling.
Legrand’s MGX/NVL cabinet shows the same pattern from the enclosure side: it follows OCP principles but still supports mounting 19-inch EIA/ECA-310 rails inside an OCP-based structure, using blanking panels to fit legacy 19-inch gear alongside OCP hardware in the same rack.
The vendor framing is explicit that OCP standards handle consistency at a general level, while GPU-driven power density, thermal profiles, and non-standard cabinet widths still require case-by-case customization.
Where interoperability actually breaks
Lining up three OCP-adjacent rack approaches side by side shows where the shared standard ends and vendor-specific engineering begins.
| Rack / cabinet | Governing OCP-related spec | Power handling | Mechanical / cabling note |
|---|---|---|---|
| AMD Helios (2026 reference design) | ORW open rack-wide form factor | Clusters up to 10.4MW, racks up to 246kW | Double-wide, Zen 6 Eypc CPUs, MI400 GPUs, Vulcano NICs |
| Legrand MGX/NVL cabinet | OCP-aligned, not one single OCP document | Not specified in the available evidence | Mounts 19-inch EIA/ECA-310 rails inside OCP structure via blanking panels |
| ORV3 baseline rack | OCP Open Rack V3 | Sustains loads above 40 kW on 48V DC busbars | Centralized power shelf converts AC to DC once per rack |
The pattern across all three rows is the same: OCP supplies shared vocabulary – rack width, power busbar, mechanical envelope – but power capacity, cooling method, and cabling detail still get finished by the vendor building the actual deployment.
The hyperscaler holdouts change the interoperability math
OCP coverage also depends on who is buying: Meta is the hyperscaler most committed to the open OCP ecosystem specifically to standardize hardware and reduce cost, while Google, AWS, and Microsoft are pursuing proprietary scale-up designs instead.
Google has used its own ICI interconnect since 2017 and added Optical Circuit Switching in 2021, and plans its in-house Axion CPU and Boardfly network architecture for the TPU v8 series in 2026; AWS is deploying proprietary NeuronLink and Graviton CPUs in 2026 Trainium3 UltraServers; Microsoft is building an Ethernet scale-up network with Cobalt 200 CPUs for its Maia 200 POD.
That split means a rack spec built around OCP – ORV3 power, ORW mechanical, or Sidecar high-voltage DC – mainly governs Meta-aligned deployments and merchant-silicon racks like Helios, not the proprietary racks running inside Google, AWS, or Microsoft’s own fleets.
What to verify before you spec an AI rack around OCP
Before committing budget to an “OCP-compliant” AI rack, get the vendor’s answer in writing on three separate questions rather than treating one compliance claim as covering all of them.
- Which specific OCP document the rack claims to follow – ORV3, Sidecar, or ORW – and whether that covers power, mechanical form factor, or both
- Whether a separate facility-integration reference design exists, the way AMD and Schneider Electric published one for Helios, or whether that integration work is still unfinished
- Whether the cabinet supports mixed 19-inch and OCP mounting if the deployment carries any legacy EIA/ECA-310 equipment
The rack density numbers in vendor material keep climbing, but the standards question is narrower and more mechanical than density: confirm which OCP document a rack actually implements, and treat every remaining gap as vendor-specific integration work rather than something the OCP label already resolved.