Aivres shows 72-GPU Vera Rubin NVL72 and NVL8 systems at OCP

Marcus Bennett

Aivres put a 72-GPU NVIDIA Vera Rubin NVL72 rack on display at OCP Global Summit 2026 in San Jose. The KRS8000V4 also carries 36 Vera CPUs and more than 20TB of HBM. Most of the display consisted of dummy blanks, although one live node exposed the rack's working hardware for inspection by visitors at the booth.

The wider Aivres lineup ranged from a PCIe GPU server to the NVL72 rack. Between them sat Vera CPU hardware, an HGX Rubin NVL8 tray and an HGX B300 server.

Aivres had shown the rack at NVIDIA GTC 2026, giving visitors a closer look at the surrounding hardware. The enclosure remains cumbersome to transport, although this reporter has seen it at many previous events.

Rack configuration

The KRS8000V4 combines compute, switching, management and cooling in one enclosure. Its listed configuration includes:

• 18 1U compute trays
• Nine 1U NVLink switch trays in the middle
• Power shelves at the top and bottom

Those trays carry 72 Rubin GPUs and 36 Vera CPUs over sixth-generation NVLink. The assembly also includes management switches, an integrated NVLink spine, power delivery and liquid-cooling infrastructure. That combination creates one rack-wide NVLink GPU domain containing more than 20TB of HBM.

The rack combines compute, switching, management and cooling in one enclosure, so buyers must assess the complete infrastructure rather than treating the GPU count as a standalone metric.

Aivres says the system is 100% liquid-cooled and can operate at 45°C to improve cooling efficiency. For scale, a hot tub usually runs at 38–39°C, indicating how warm the rack coolant runs.

Eight-GPU compute tray

The Aivres 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. The exhibit demonstrates an eight-GPU HGX platform hosted by x86 CPUs in a fanless enclosure.

Aivres lists up to 400 PFLOPS for NVFP4 inference and 280 PFLOPS for NVFP4 training. The booth placard also records 2.3TB of HBM4 memory and 176TB/s of total HBM bandwidth.

Together, those figures describe a tray built to concentrate low-precision inference work and accelerator memory without conventional chassis fans, while the system remains an upcoming product.

Connectivity and expansion include:

• One UART port at the front
• Two front USB ports
• Two front RJ45 ports
• Eight front E1.S drive bays in one row
• Space for an NVIDIA BlueField DPU

The rear uses OCP rack quick disconnects and a busbar for power. The rear has no exhaust area for fans because the system is liquid-cooled.

Cooling and rear connections

Aivres calls the KL2288 “a modular, fanless, fully liquid-cooled design.” Both CPUs use copper cold plates. Banks of liquid-cooled DDR5 DIMMs flank those processors inside the tray.

Distribution blocks route coolant through the compute tray. Red fittings mark the warm line, while blue fittings mark the cooler line. Visible plumbing makes cooling part of the chassis architecture rather than an optional add-on.

Operators would need to account for coolant distribution and rack power when planning deployment, because both connections are built into the rear of the tray.

What does this change for AI-system buyers?

At OCP, Aivres showed two Rubin deployment scales: a 72-GPU rack and an eight-GPU 2U tray. The rack offers a shared NVLink domain and more than 20TB of HBM. The tray pairs x86 host processors with 2.3TB of HBM4 and 176TB/s bandwidth. The display gives buyers a clearer view of scale, interfaces and cooling requirements.

Deployment choices broaden from a 72-GPU rack to an eight-GPU tray, but each configuration requires coolant and power planning rather than a simple air-cooled server swap. That is the concrete change for buyers: the hardware form factors are clearer, while the KL2288 remains an upcoming system.