Orient: The reasoning weapon LPX is born, Agent AI and the space computing architecture welcome innovation.

Orient released a research report stating that in the short term, in the field of communication, a single-layer scale-up domain copper cable is still the preferred solution. If a scale-up domain adopts a dual-layer architecture, it will be a combination of copper and optical cables. For the high computing power density Kyber architecture, orthogonal backplanes are gradually making their way onto the stage. For scale-out scenarios, the CPO solution is gradually maturing. In terms of heat dissipation, all five different racks in the Vera Rubin POD use 100% liquid cooling. In terms of power supply, 800V HVDC has gone from "optional" to "AIDC standard". In terms of space computing power, with the addition of NVIDIA Corporation (NVDA.US), the ecosystem is expected to continue to mature. Orient's main points are as follows: Groq 3 LPX, a powerful inference tool, focuses on low-latency tasks with significant commercial benefits. On March 17, the 2026 GTC Conference was officially held, and NVIDIA Corporation released the LPX rack, which focuses on inference acceleration. The core of the LPX rack is composed of Groq 3 LP30 LPU, which is a flat, SRAM-based flat memory architecture designed to achieve fast and predictable token generation. A single Groq 3 LPX rack contains 32 liquid-cooled 1U compute trays, providing a total of 315 PFLOPS AI inference power, 128GB total SRAM capacity, and 640TB/s total scale-up bandwidth. When combined with the Rubin rack, LPX provides more than 35 times the token processing capacity per megawatt and more than 10 times the revenue opportunity for trillion-parameter models compared to Blackwell. The introduction of LPU solves the issue of high cost for single GPU systems in current high-end model inference scenarios. The deployment of Rubin+LPX is expected to indirectly improve the profitability of cloud operators, accelerate the landing of low-latency AI inference applications, gradually activate the vast market demand for inference, and further drive the demand for PCBs, liquid cooling, power supply, optical communication, and copper cables. Vera CPU rack: The cornerstone of large-scale AI agents and reinforcement learning. NVIDIA Corporation has released the Vera CPU rack, which includes 32 liquid-cooled Vera CPU trays. Each rack consists of 8 CPUs and 2 BlueField DPUs. The Vera CPU features the Olympus core, offering a 50% improvement in single-thread performance compared to traditional x86 CPUs. The rack contains 256 CPUs, providing 400TB of memory, 300TB/s memory bandwidth, and uses liquid cooling for heat dissipation and Ethernet connectivity. A single CPU rack can support over 22,500 concurrent reinforcement learning or Agentic sandbox environments, with performance that is twice as efficient and 50% faster than traditional rack-scale CPUs. It can test, execute, and validate the results of Vera Rubin NVL72 and LPX racks, helping to develop Agent AI. The deployment of CPUs for AI agents is crucial for inference, and Vera CPU's strong single-thread performance fully supports frequent tool calls in AI agent scenarios. The release of the Vera CPU rack is expected to kickstart the deployment of large-scale AI agents, assisting AI in transitioning from providing recommendations to task execution. Space computing module Space-1 makes its debut, Rubin Ultra uses orthogonal backplanes. NVIDIA Corporation is collaborating with partners to develop the Space-1 Vera Rubin Module, an AI computing module optimized for space, which is expected to accelerate the development of future space data centers. Additionally, NVIDIA Corporation has announced the Rubin Ultra Kyber architecture, which uses orthogonal backplanes for interconnection between computing and exchange boards. Kyber NVL144 supports scaling up to NVL1152 using optical interconnects. Related companies: PCB-related companies Wus Printed Circuit, Victory Giant Technology, Shenzhen Kinwong Electronic, Shennan Circuits, Founder Technology Group, SHENZHEN JOVE ENTERPRISE LIMITED, Avary Holding, Suzhou Dongshan Precision Manufacturing; CPO-related companies Yuanjie Semiconductor Technology, Suzhou Everbright Photonics, Henan Shijia Photons Technology, Accelink Technologies, Suzhou TFC Optical Communication, T&S Communications, Shenzhen Zesum Technology, EverProX Technologies; liquid cooling-related companies Shenzhen Envicool Technology, Lingyi Itech, Guangdong Shenling Environmental Systems, Guangzhou Goaland Energy Conservation Tech, Guangdong Suqun New Material; space computing-related companies Shanghai Fudan Microelectronics Group, Chengdu Sino-Microelectronics Tech., etc. Risk warning: AI development falling short of expectations, increased industry competition risks, political risks for GEO Group Inc.