LITEON and QCT showcase 110 kW Power Shelf for Vera Rubin NVL72 infrastructure
Source report: 2026-10-16 · Editorial analysis published: 2026-10-16
LITEON’s October 16 release describes a joint OCP demonstration and platform co-design. The 110 kW shelf rating is distinct from rack consumption, deployed capacity and verified mining-hardware compatibility.
Analysis and practical implications
This section is our analysis and illustrative calculations, separate from the source report.
An October 16 report on the OCP demonstration
LITEON and Quanta Cloud Technology published a joint showcase update on October 16 for infrastructure supporting NVIDIA Vera Rubin NVL72. The announcement highlights LITEON's 110 kW Power Shelf within a rack-level power architecture and cooperation on system design and validation.
The event information places OCP Global Summit in San Jose on October 12–15. The publication date is therefore not a claim that a new commercial installation opened on October 16. Nor does a demonstration confirm a customer acceptance test or a volume-shipment date.
Shelf rating versus rack demand
The 110 kW figure identifies the showcased Power Shelf. It should not be read as measured consumption for every NVL72 rack, a guaranteed site connection, or 110 kW available separately for every computing node. The release does not provide those allocations.
Capacity and utilization are different quantities. A power shelf can be selected to cover a design envelope while the connected workload changes over time. To assess a real installation, the relevant records would include the operating voltage, accepted configuration, load profile and redundancy arrangement.
Integration is the main technical theme
LITEON describes high-efficiency conversion, intelligent load balancing and digital power management designed for rapid load fluctuations. QCT contributes AI platform, system design and data-center integration expertise. The companies describe platform-level co-design and validation rather than an isolated commodity supply.
The key practical issue is how computing, networking, cooling and power behave together. A rating on one subsystem cannot prove the performance of the entire rack. Transient response, protection coordination and recovery behavior require measurements under a defined load and operating configuration.
Efficiency needs a measurement boundary
The release argues that conversion losses add heat and that power-system footprint can reduce room for compute. It does not disclose a measured efficiency curve, a numerical whole-site energy saving or an independently verified performance-per-watt result for this demonstration.
For illustration only, delivering 100 kW at 96% conversion efficiency requires about 104.17 kW input, with about 4.17 kW lost in that conversion stage. These chosen inputs are not LITEON specifications. Cooling, distribution and other stages must be accounted for separately before comparing complete facilities.
A power rating is not an energy bill
If a hypothetical load continuously consumed 110 kW for 24 hours, it would use 2,640 kWh. The example illustrates the difference between power and energy; it is not a claim that the showcased shelf or an NVL72 system draws 110 kW continuously.
Real operating costs require time-resolved demand, electricity terms and auxiliary consumption. Nameplate power multiplied by a full year can be a scenario, but cannot establish an observed bill or annual energy consumption without evidence about utilization and the measurement boundary.
Resilience is an architecture question
The supplier describes the solution as intended to improve system resilience and accommodate dynamic AI workloads. The announcement does not specify a complete redundancy topology, reserve-battery duration or tested ability to keep every workload running through a utility outage.
Digital balancing and management are useful capabilities, but they are not synonyms for energy storage or a second independent supply. An operating design needs to state which components can fail, what capacity remains and how the computing load behaves during a transition.
Implications for ASIC, GPU and CPU miners
AI infrastructure and mining share concerns about power delivery and heat, but use different electrical and computing platforms. Support for NVIDIA MGX and Vera Rubin NVL72 does not establish compatibility with an ASIC power supply or a mining algorithm.
The release announces no mining hashrate, validated mining profile, replacement PSU for a named miner or firmware package. Calculator and catalogue data should therefore retain their existing device-specific sources. Transferring an AI shelf rating into a miner's specifications would mix unrelated hardware boundaries.
Commercial checkpoints after the showcase
The next useful disclosures would identify production configurations, voltage and connector requirements, measured conversion and transient behavior, and customer delivery or deployment evidence. These distinguish a demonstrated architecture from a product available for a particular installation.
The verified news is LITEON and QCT's joint 110 kW shelf and rack-integration showcase, reported after OCP. Its significance is the attention to system-level power design; actual energy savings, deployed capacity and mining compatibility remain questions for separate technical evidence.
Source: LITEON ↗
Mining calculator ↗
