OCP Showcase Highlights 800VDC, 1.6T Networking and Hardware Security for AI Infrastructure

Morgan Stanley’s AI-infrastructure assessment, presented against the backdrop of the OCP Global Summit, points to a converging set of changes in networking, power delivery and platform control. The most immediate issue for data-center design is high-voltage DC distribution: Nvidia and Delta showed a reference architecture in which a single 660 kW 800VDC power rack supplies two Vera Rubin NVL72 compute racks.

The source estimates 800VDC conversion efficiency at 95%, compared with 93% for a conventional 480VAC architecture. For each 100 MW of IT load, it estimates potential annual savings of roughly 19.8 million kWh and US$1.6 million in operating costs. These figures are estimates rather than reported operating results, and the comparison conditions were not fully disclosed.

Small-scale deployment of high-voltage DC power is expected from late 2026 through 2027. This is a forecast, not evidence of commercial commissioning. The proposed architecture also incorporates liquid cooling for power shelves, reflecting the growing thermal and physical-density constraints associated with high-power AI racks.

Networking is advancing in parallel. Accton and Celestica demonstrated 1.6T switches based on Broadcom’s Tomahawk 6. Suppliers at the event also presented optical and cable interconnect technologies intended to address power efficiency at higher network speeds. Credo said its 1.6T ZeroFlap ALC cable consumes 12 W at each end over a 50 m transmission distance. TE Connectivity expects its 224G/448G-capable solutions to enter volume production in 2027.

Hardware trust, telemetry and management formed the third theme. As AI infrastructure becomes more complex across compute, power and cooling layers, the source identifies greater demand for platform observability and hardware security. Open Caliptra RoT was described as a widely adopted hardware root-of-trust standard.

Infineon also projected that solid-state transformers could raise power efficiency by 2–5 percentage points while reducing volume and weight by more than 30%. Those are vendor projections, as is its forecast of a market opportunity exceeding US$1 billion in small transformers by 2030.

Taken together, the demonstrations suggest that the AI data-center equipment stack is expanding beyond accelerators and servers to include high-voltage DC systems, liquid-cooled power nodes, high-speed optical or cable interconnects, and hardware-level security and telemetry. If the cited efficiency claims and deployment forecasts are validated, power delivery and cooling could account for a larger share of both the cost base and technical differentiation of AI campuses.