In brief
- Aivres’s KRS8000V4 rack houses 72 Rubin GPUs and 36 Vera CPUs.
- According to the company, the NVL72 rack is fully liquid-cooled and can operate at 45 °C.
- KL2288 combines eight Rubin GPUs with dual AMD EPYC 9006 “Venice” processors.
- Aivres reports up to 400 PFLOPS for NVFP4 inference and 280 PFLOPS for training on KL2288.
At OCP Global Summit 2026 in San Jose, Aivres showcased its Vera Rubin NVL72 AI rack and the KL2288 compute tray it has developed for the HGX Rubin NVL8 platform. According to ServeTheHome’s report from the show floor, the booth also featured Vera CPU hardware, an HGX B300 server and a PCIe GPU server. The source describes KL2288 as an upcoming system; no release date or pricing was provided.
NVL72 rack houses 72 Rubin GPUs
Aivres’s KRS8000V4 rack system is a liquid-cooled AI infrastructure platform designed for NVIDIA Vera Rubin NVL72. The company’s configuration details show that the system contains eighteen 1U compute trays and nine 1U NVLink switch trays. Together, these house 72 Rubin GPUs and 36 Vera CPUs.
Aivres says the GPUs are connected via sixth-generation NVLink and that the system has more than 20 TB of HBM memory. According to the company, the entire rack is liquid-cooled and can operate at temperatures of up to 45 °C. In the example shown at the event, some sections were running, while other trays had representative dummy modules in place. Along with compute and switch trays, the rack includes management switches, power components and a liquid-cooling system.
KL2288 GPU, processor and memory specifications
KL2288 is a 2U compute tray built for HGX Rubin NVL8. It pairs eight Rubin GPUs with dual AMD EPYC 9006 “Venice” processors. Aivres describes the design as modular, fanless and fully liquid-cooled. The hardware reviewed by ServeTheHome uses copper cold plates for the processors, with the DDR5 memory modules also incorporated into the cooling design.
The manufacturer reports peak performance of up to 400 PFLOPS for NVFP4 inference and up to 280 PFLOPS for NVFP4 training. The memory configuration listed for KL2288 includes 2.3 TB of HBM4 and total HBM bandwidth of 176 TB/s. These are figures provided by the manufacturer; the source does not include independent performance testing.
The front of the tray has one UART connection, two USB ports, two RJ45 ports and eight E1.S drive bays. Space is also provided for an NVIDIA BlueField DPU. At the rear, OCP rack quick-connects and busbars are used for power delivery. Because of the liquid-cooled, fanless design, the system has no fan-exhaust area at the back.
What this means for architecture and visualization workflows
These systems are infrastructure for intensive AI computing, rather than desktop modeling or rendering workstations. For architecture, interior design and archviz teams, the showcase points to server hardware capable of running AI workloads, rather than a new modeling feature. However, the source provides no information on compatibility with specific architecture software, integration with existing rendering tools or the performance these systems could deliver in office workflows.
Organizations evaluating the hardware should therefore consider the rack requirements and liquid-cooling setup alongside the quoted performance figures. Since no price or release date has been announced, cost and availability remain uncertain for investment planning. The source also provides no information on licensing terms or software support, so these should be verified separately during procurement.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
The hardware Aivres showcased illustrates that AI computing is about more than choosing CPUs and GPUs: rack and cooling systems are also part of the design. For most architecture and visualization studios in Turkey, this is a development in the AI infrastructure they use, rather than a new desktop workstation.
Before making an investment decision, compare the manufacturer’s performance figures with independent tests and wait for pricing and availability details. Liquid-cooling requirements and rack compatibility should also be factored into cost and installation assessments. Since the source does not specify software compatibility or licensing terms, don’t assume the system will suit your workflow until those details are clarified.
Frequently asked questions
Which GPUs and processors does Aivres KL2288 use?
KL2288 pairs eight Rubin GPUs with dual AMD EPYC 9006 “Venice” processors.
What performance and memory specifications have been announced for KL2288?
Aivres reports up to 400 PFLOPS for NVFP4 inference and up to 280 PFLOPS for training. The system is specified with 2.3 TB of HBM4 memory and total HBM bandwidth of 176 TB/s.
When will Aivres’s Vera Rubin NVL72 system go on sale?
The source does not announce a release date or price.
