Sealand: The large-scale model training drives the growth of AI computing power demand, with continuous upgrades in computing, networking, and storage.

Sealand released a research report stating that the demand for AI computing power is driven by large-scale model training. New generation computing architectures such as GB300 and Vera Rubin will be introduced, and various segments of the computing industry chain including AI chips, server systems, copper connections, HBM, liquid cooling, optical modules, and IDC are expected to continue to benefit. The "recommendation" rating for the computer industry remains unchanged. Key points from Sealand include: GPU Core: Significant improvement in computing power for GB300, optimistic outlook for Rubin B300 is based on the Blackwell Ultra architecture and uses TSMC 4NP process and CoWoS-L packaging. The FP4 floating point computing power can reach 15PFLOPS, which is 1.5 times that of B200. GB300 is equipped with 288GB of HBM3E memory. The performance of Vera Rubin NVL144 is 3.3 times that of GB300 NVL72. The Rubin Ultra NVL576 has a performance increase of 14 times compared to GB300 NVL72, and the memory is 8 times that of GB300. In FY2026Q2, Nvidia's revenue was $46.7 billion, up 56% year-on-year. Server Details: Upgrade from HGX to MGX architecture for Rubin Ultra NVL576 The GB300 NVL72 system consists of 18 compute trays and 9 switch trays, equipped with 72 Blackwell Ultra GPUs and 36 Grace CPUs. Compared to Hopper, GB300 NVL72 has the potential to improve overall factory output performance by up to 50 times. GB300 NVL72 integrates ConnectX-8 network cards and BlueField-3 DPUs. Rubin Ultra NVL576 will be launched in 2027, using the Kyber architecture to significantly improve rack density. Networking: CPO replaces pluggable optical modules, achieving ultra-high-speed interconnection for Rubin CPO integrates silicon photonics devices with ASIC packaging, replacing traditional pluggable optical modules. Compared to traditional networks, it improves energy efficiency by 3.5 times and deployment speed by 1.3 times. The Quantum-X and Spectrum-X switches reduce dependence on traditional optical transceivers, providing up to 400Tbps throughput for ultra-large scale AI factories. Rubin will adopt NVLink 6.0 technology, doubling the speed to 3.6TB/s. NVLinkFusion opens up third-party interconnection ecosystems to support heterogeneous chip collaboration. The new generation NVSwitch 7.0 expands to connect 576 GPUs, achieving non-blocking communication for larger scale GPU connectivity. HBM: Mass production of HBM4 expected in 2026, SK Hynix leads the HBM market In March 2025, SK Hynix delivered the world's first 12-layer stacked HBM4 samples to major customers. Samsung is ready to provide HBM4 samples to AMD and Nvidia customers by the end of July 2025. Micron has showcased plans for the next generation HBM4 in its AI memory product portfolio, expected to be mass produced in 2026. SK Hynix has collaborated with NVIDIA, Microsoft, and Broadcom to design customer-specific customized HBM chips. Cold plate liquid cooling is more mature, GB300 NVL72 adopts full liquid cooling solution GB300 uses an independent liquid cooling plate design, with each chip equipped with a separate inlet and outlet liquid cooling plate, rather than a large area cold plate coverage. Liquid cooling technology has higher heat dissipation efficiency, including cold plate type and immersion type, where the cold plate type is indirect cooling, with moderate initial investment, low operating costs, relatively mature. Nvidia's GB300 NVL72 adopts a full liquid cooling design. Risk warning: downstream demand affected by macroeconomic factors, slow development in large model industries, slower development in AI chip industries, increased market competition, escalating U.S.-China tensions, company performance falling short of expectations, companies not directly comparable, data for reference only.