Xtreme-D, in cooperation with TES-AMM JAPAN and Cornelis Networks, has announced the launch of the AI-oriented computing infrastructure service "Raplase RCA". The new ecosystem relies on recycling and operating used and refurbished GPU servers to enable institutions and companies to establish private AI environments at the lowest cost and with standard delivery periods starting from a single graphics processor (1 GPU).
Addressing Supply Chain Crisis and Facing Excess Spending
The computing sector has witnessed a massive surge in demand for AI acceleration servers, accompanied by a sharp rise in the prices of new hardware and extended waiting and delivery periods reaching up to approximately ten months in some cases. Naoki Shibata, CEO of Xtreme-D, explained that many companies fall into the trap of purchasing superior technical specifications that exceed their actual needs and incur exorbitant costs to operate the latest generations of graphics processors, while most enterprise workloads can be executed with high efficiency on refurbished and software-tuned servers.
The new service aims to reduce the time required to obtain computing power and significantly cut capital expenditures, while providing a flexible subscription model that ensures the protection of sensitive data within the boundaries of the institution rather than exporting it to public cloud services.
Triple Engineering Integration to Bridge Memory and Network Gaps
Running large language models (LLMs) on older graphics processors involves the challenge of memory capacity and data transfer bottlenecks. To overcome these obstacles, the three parties integrated a technical ecosystem whose tasks are precisely distributed:
- Memory and Operations Solutions by Xtreme-D: The company utilized "FUJITSU ACB" technology and specialized software tuning layers to connect and expand the memory of graphics processors with host system memory, allowing modern smart models to run smoothly without a sharp drop in performance, alongside providing data center spaces equipped to host this hardware.
- Super-fast Networks from Cornelis Networks: The company provided "Cornelis CN5000 Omni-Path" network communication equipment supporting speeds up to 400 Gbps, to address the bottleneck in server interconnection and ensure maximum distributed computing efficiency for complex computational packages, AI processing, and High-Performance Computing (HPC).
- Supply Chains and Reliability from TES-AMM JAPAN: As the Japanese arm of the global "SK tes" group specializing in IT Asset Disposition (ITAD) in over 20 countries, the company secures used servers, tests them, ensures their reliability, and performs secure data sanitization, while providing a clear tracking path for the hardware's usage history to meet strict institutional standards.
What Practically Changes for Institutions?
This approach represents a practical outlet for companies that need to train and deploy their models without joining waiting lists for purchasing modern graphics processors, or those whose data governance regulations prevent them from resorting to shared cloud services. The supply model based on a single graphics processor also opens the door for small and medium-sized enterprises to test AI applications and expand gradually, with the ability to flexibly adjust the computing configuration according to actual workload requirements without long-term capital commitments.