Aivres unveils liquid-cooled NVIDIA Vera Rubin racks at OCP Summit

The Vera Rubin hardware lineup

At the OCP Global Summit 2026 in San Jose, Aivres displayed a comprehensive lineup of upcoming NVIDIA hardware. The exhibit spanned a wide range of configurations, from individual PCIe GPU servers to large-scale racks. This allowed attendees to see how different components fit into the broader ecosystem of the next-generation Rubin architecture.

The centerpiece of the display was the KRS8000V4, which is a liquid-cooled AI rack designed for NVIDIA Vera Rubin. This specific model was previously seen at the NVIDIA GTC 2026 event. By showing it alongside other hardware, Aivres demonstrated how the new Rubin chips integrate with various CPU and switch components.

The lineup also included the HGX Rubin NVL8 tray and the HGX B300 server. These pieces provide a look at the different ways companies can deploy high-performance computing. Whether for massive training clusters or smaller inference tasks, the hardware is designed to scale across different architectural needs.

Scaling the GPU domain

The KRS8000V4 rack is built to create a massive GPU domain over NVLink, which is the high-speed interconnect used by NVIDIA. The system includes 18 compute trays and nine NVLink switch trays. These components work together to link 72 Rubin GPUs and 36 Vera CPUs using sixth-generation NVLink technology.

This configuration creates a unified environment with over 20TB of HBM (High Bandwidth Memory). HBM is a type of memory designed to provide the high speeds necessary for AI workloads. By grouping these resources into a single domain, the system allows for much larger models to be processed more efficiently than if the chips were isolated.

The rack is organized with the power shelves positioned at the top and bottom, while the switch trays occupy the center. This layout integrates management switches, an integrated NVLink spine, and the liquid cooling infrastructure. It represents a highly dense way to pack significant computing power into a single physical footprint in a data center.

The move to liquid cooling

Aivres designed the KRS8000V4 to be 100% liquid-cooled to manage the extreme heat of the Rubin chips. The system can operate at temperatures as high as 45C. To put this in perspective, the article notes that a hot tub is typically between 38C and 39C, illustrating how warm the fluid in the rack actually is.

The liquid cooling approach is necessary because it provides much better efficiency than traditional air cooling. Because the system uses liquid to move heat away from the components, the hardware does not require traditional exhaust areas for fans. This allows for a more compact and modular design that can handle higher power densities without overheating.

The KL2288 compute tray serves as a specific example of this technology in a smaller form factor. It is a 2U tray that pairs eight Rubin GPUs with dual AMD EPYC 9006 processors. This tray is described as a modular, fanless, and fully liquid-cooled design, showing how x86 CPUs can host high-end GPU platforms without needing large cooling fans.

Internal hardware engineering

Inside the KL2288 compute tray, the engineering focuses on thermal management for every major component. The two CPUs are equipped with copper cold plates to facilitate heat transfer to the coolant. Additionally, the banks of DDR5 DIMMs, which are the memory modules, are also liquid-cooled to ensure they do not throttle under heavy use.

The coolant is routed through the tray using distribution blocks with specific fittings. These are color-coded with red for warm fluid and blue for cooler fluid to help with the routing of the liquid. This system ensures that the heat generated by the GPUs and memory is consistently moved away from the hardware to the cooling source.

The rear of the KL2288 tray features OCP rack quick disconnects and a busbar for power. These are standard features for modular data center designs, allowing for easier maintenance and power distribution. The lack of fan exhaust on the rear confirms that the liquid cooling system is doing all the work of keeping the components within safe operating temperatures.

Performance and connectivity

The hardware is designed to deliver high performance for both training and inference tasks. The KL2288 tray is capable of reaching 400 PFLOPS for NVFP4 inference and 280 PFLOPS for NVFP4 training. These numbers represent the amount of mathematical operations the system can perform per second using specific data formats.

The tray also provides 2.3TB of HBM4 memory with a total bandwidth of 176TB/s. This high bandwidth is critical for moving data quickly between the memory and the GPUs. Without this speed, the processors would spend too much time waiting for data, which would slow down the training of large AI models.

For connectivity, the front of the KL2288 includes a UART port, two USB ports, two RJ45 ports, and a row of eight E1.S drive bays. It also includes dedicated space for an NVIDIA BlueField DPU. These features ensure that the tray can be integrated into larger networks while still providing the necessary storage and management ports for the end user.