Aivres shows liquid-cooled NVIDIA Rubin rack and fanless HGX tray at OCP Summit

OCP Global Summit display pairs a Vera Rubin NVL72 rack with a 2U HGX Rubin NVL8 compute tray

By LineZotpaper AI staff

Published25 October 2026

Read Time2 min

Illustration: Aivres shows liquid-cooled NVIDIA Rubin rack and fanless HGX tray at OCP Summit

Aivres brought a range of NVIDIA-based systems to the OCP Global Summit 2026 in San Jose, headlined by the KRS8000V4, a fully liquid-cooled NVIDIA Vera Rubin NVL72 AI rack, alongside a new fanless 2U compute tray for HGX Rubin NVL8.

At the OCP Global Summit 2026 in San Jose, the Aivres booth spanned a Vera Rubin NVL72 rack down to a PCIe GPU server, with an HGX Rubin NVL8 tray, Vera CPU hardware, and an HGX B300 server in between. The company also showed the Vera Rubin rack at NVIDIA GTC 2026, making this stop a chance to see the surrounding hardware.

The centerpiece was the KRS8000V4, Aivres's liquid-cooled NVIDIA Vera Rubin NVL72 AI rack. The display mixed dummy blanks with a live node. According to Aivres, the rack holds 18 1U compute trays and nine 1U NVLink switch trays, with the switch trays in the middle and power shelves at the top and bottom. That adds up to 72 Rubin GPUs and 36 Vera CPUs on sixth-generation NVLink, forming a single GPU domain with over 20TB of HBM. Aivres says the rack is 100% liquid-cooled and can operate at 45C to aid cooling efficiency.

The second system, the KL2288, is a 2U liquid-cooled compute tray for NVIDIA HGX Rubin NVL8. It pairs eight Rubin GPUs with dual AMD EPYC 9006 "Venice" processors and is rated at up to 400 PFLOPS for NVFP4 inference and 280 PFLOPS for NVFP4 training, with 2.3TB of HBM4 memory and 176TB/s of total HBM bandwidth. Aivres calls it a modular, fanless, fully liquid-cooled design. Inside, the two CPUs sit under copper cold plates, flanked by banks of liquid-cooled DDR5 DIMMs, while distribution blocks with red and blue fittings route coolant through the tray. The front carries a UART port, two USB ports, two RJ45 ports, and a row of eight E1.S drive bays, with space for an NVIDIA BlueField DPU. The rear uses OCP rack quick disconnects and a busbar for power, with no fan exhaust area because the system is liquid-cooled.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Analysis

Why This Matters

Rack-scale liquid-cooled AI systems are becoming the norm for large AI workloads, and these previews show how NVIDIA's next-generation Rubin hardware will be packaged and shipped to data centers.
The 45C coolant temperature and fanless tray design show how much emphasis vendors now place on data center cooling efficiency as system power grows.
The Vera Rubin NVL72 rack is positioned as a major platform for large AI clusters, so early previews like this one signal the platform's rollout is approaching.

Background

NVIDIA's Rubin platform is the company's next generation of AI accelerators, following the Blackwell family. The NVL72 configuration scales the GPUs into a rack-sized, liquid-cooled system designed for very large training and inference deployments. Aivres builds servers and racks around NVIDIA platforms and regularly shows upcoming systems at industry events such as the OCP Global Summit, a gathering focused on open data center hardware.

What to Watch

Whether the KRS8000V4 and KL2288 move from trade-show displays into production systems and customer deployments.
How widely NVIDIA's Vera Rubin NVL72 racks are adopted by cloud providers and AI companies as the platform ramps up.
The role of x86 hosts in Rubin systems, since the KL2288 pairs Rubin GPUs with AMD EPYC 9006 "Venice" processors while NVIDIA also promotes its own Vera CPUs.

Sources

Aivres NVIDIA Vera Rubin and HGX B300 at OCP Summit · ServeTheHome

Checked

12 claims in this story checked against 1 source: 12 supported.

Show each claim

SupportedThe KRS8000V4 is a liquid-cooled NVIDIA Vera Rubin NVL72 AI rack.ServeTheHome: “the Aivres KRS8000V4, the company’s NVIDIA Vera Rubin NVL72 liquid-cooled AI rack.”
SupportedThe HGX Rubin NVL8 system is a fanless 2U compute tray.ServeTheHome: “The Aivres KL2288 is the 2U liquid-cooled compute tray for NVIDIA HGX Rubin NVL8”
SupportedThe OCP Summit display pairs a Vera Rubin NVL72 rack with an HGX Rubin NVL8 tray.ServeTheHome: “The lineup featured a lot of NVIDIA, spanning a Vera Rubin NVL72 rack down to a PCIe GPU server, with an HGX Rubin NVL8 tray”
SupportedAivres’s OCP Global Summit 2026 booth in San Jose featured a range of NVIDIA hardware.ServeTheHome: “At OCP Global Summit 2026 in San Jose, we stopped by the Aivres booth. The lineup featured a lot of NVIDIA”
SupportedThe rack display included dummy blanks and a live node.ServeTheHome: “At the show, this was a mix of dummy blanks and a live node.”
SupportedThe rack has 18 1U compute trays and nine 1U NVLink switch trays, with the switch trays in the middle and power shelves at the top and bottom.ServeTheHome: “Aivres lists 18 1U compute trays and nine 1U NVLink switch trays, as shown above, with the switch trays in the middle and power shelves at the top and bottom.”
SupportedThe rack configuration provides 72 Rubin GPUs and 36 Vera CPUs on sixth-generation NVLink.ServeTheHome: “That is 72 Rubin GPUs and 36 Vera CPUs on sixth-generation NVLink.”
SupportedThe rack forms a single GPU domain with over 20TB of HBM.ServeTheHome: “That creates a giant GPU domain over NVLink with over 20TB of HBM.”
SupportedAivres says the rack can operate at 45C to help cooling efficiency.ServeTheHome: “Aivres says this is 100% liquid-cooled and can operate at 45C to help cooling efficiency.”
SupportedThe KL2288 pairs eight Rubin GPUs with dual AMD EPYC 9006 “Venice” processors.ServeTheHome: “The Aivres KL2288 is the 2U liquid-cooled compute tray for NVIDIA HGX Rubin NVL8, pairing eight Rubin GPUs with dual AMD EPYC 9006 “Venice” processors.”
SupportedThe KL2288 is rated at up to 400 PFLOPS for NVFP4 inference and 280 PFLOPS for NVFP4 training.ServeTheHome: “Aivres lists the KL2288 at up to 400 PFLOPS for NVFP4 inference and 280 PFLOPS for NVFP4 training.”
SupportedThe KL2288 has 2.3TB of HBM4 memory and 176TB/s of total HBM bandwidth.ServeTheHome: “The booth placard also lists 2.3TB of HBM4 memory with 176TB/s of total HBM bandwidth.”