OCP and "Diablo": the sidecar power-rack idea
OCP (Open Compute Project) is an open standardization community for data-center infrastructure led by hyperscalers such as Google, Meta, and Microsoft. One next-generation power concept it presents is the separate sidecar power rack, which moves power conversion out of the compute rack.
Conventionally, compute servers and power supplies have shared the same rack, with in-rack 48/54V DC distribution. In the separate approach, power conversion is consolidated in a dedicated "sidecar" rack and high-voltage DC is sent to the compute rack. The aim is to remove conversion stages from the compute space to raise density and efficiency, while lifting distribution voltage to avoid the high-current problem.
From 48V DC to ±400/800V DC
The key is raising the distribution voltage. A separate power rack raises power delivery from in-rack 48V DC to ±400V or 800V DC (HVDC), supporting 100kW to 1MW IT racks. Higher voltage lowers current at the same power level, reducing conductor loss and heat.
NVIDIA is also presenting an 800V ecosystem in the same direction (NVIDIA Technical Blog). The important point is that the shift to 400/800V is advancing as a hyperscaler-led standard, not as a single vendor's proprietary design.
Separate power from compute racks
Consolidate power conversion in a dedicated sidecar, improving compute-space density and maintainability.
High-voltage DC delivery
Move from in-rack 48V DC to ±400/800V DC. This avoids high current and reduces conductor loss and heat.
Support 100kW to 1MW
The target is a power-delivery method that can serve AI racks from above 100kW to the 1MW class.
Hyperscaler-led standard
The transition is advancing through OCP as a cross-industry axis, not as dependence on one vendor.
Relationship with NVIDIA's 800V ecosystem
The separate power-rack concept aligns with NVIDIA's 800V DC architecture. NVIDIA is developing a 660kW monopolar 800V reference design, with air-cooled samples and production planned for mid-2026 and a liquid-cooled derivative planned for the second half of 2026. In other words, "standardization" through OCP and "implementation" by NVIDIA and power or semiconductor vendors are converging toward the same 800V and separate-rack direction.
From a design and procurement standpoint, assuming a separate 800V architecture makes facility distribution, protection, and the lead times for 800V-compatible power and breaker equipment new planning variables. The choice of power architecture affects the whole facility, not just a compute rack. A related article covers the detailed architecture transition.