Aivres brought a liquid-cooled NVIDIA Vera Rubin rack and an HGX Rubin compute tray to OCP Global Summit in San Jose. The standout is a rack with 72 Rubin GPUs, 36 Vera CPUs and more than 20TB of HBM, a sizeable glimpse of the hardware being built for AI computing.
NVIDIA Watch analysis
What happened
ServeTheHome reports that the Aivres KRS8000V4 rack displayed at the event contained a mix of dummy blanks and a live node. Aivres lists 18 compute trays and nine NVLink switch trays, with sixth-generation NVLink connecting the system.
The company also lists a 2U, liquid-cooled HGX Rubin NVL8 tray with eight GPUs and dual AMD EPYC 9006 processors. Its stated figures are up to 400 PFLOPS for NVFP4 inference and 280 PFLOPS for NVFP4 training, with 2.3TB of HBM4 and 176TB/s of memory bandwidth. Those are specifications listed by Aivres, not independent benchmark results.
Why it matters
These systems show the scale and specialised cooling behind high-end AI compute: one rack brings together dozens of GPUs, CPUs, memory and networking, while the smaller tray puts eight GPUs into a 2U design. For operators, the practical story is not just raw compute but how that much hardware is powered, cooled and connected.
Aivres says the rack is fully liquid-cooled and can operate with coolant at 45°C. The display included dummy components, so the booth visit is a look at the systems and their listed specifications, not a demonstration that every component was running live.
Our read
This is a useful hardware story because it puts concrete configurations around the next wave of AI infrastructure. The 72-GPU rack makes for the headline, but the 2U tray may be the more revealing design: eight GPUs, substantial memory and liquid cooling packed into a much smaller footprint. The performance figures deserve interest, and the usual asterisk: they are vendor specifications, not proof of real-world throughput.
What to watch
- Whether Aivres publishes independent performance results for the Rubin systems.
- When the KRS8000V4 rack and KL2288 tray become available to customers.
- How operators handle power and liquid-cooling requirements at this scale.
Discussion spark: For AI infrastructure, which matters more to operators: maximum performance in a giant rack, or the ability to pack useful compute into smaller, easier-to-deploy systems?
Sources and evidence
- Source update (25 October 2026, 21:51 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.