Led by Blackwell Ultra, the Rubin platform is beginning to take shape, and AI hardware is undergoing a comprehensive upgrade... What surprises will NVIDIA unveil at GTC 2025? On March 18, NVIDIA is set to hold the 2025 Global AI Conference for developers. Yesterday, JP Morgan analysts Gokul Hariharan, Albert Hung, and others released a report for preview. JP Morgan expects that NVIDIA will launch the Blackwell Ultra chip (GB300) at the conference and may disclose some details about the Rubin platform. This conference will also focus on the comprehensive upgrade of AI hardware, including higher performance GPUs, HBM memory, enhanced thermal and power management, and the CPO (Co-Packaged Optics) technology roadmap. Additionally, JP Morgan anticipates that humanoid robots and physical AI will also be highlights of the conference—for investors, this means NVIDIA will continue to lead AI hardware innovation, and related industry chain companies are likely to benefit. However, investors should also be aware of the risks of slowing growth in data center AI spending. Currently, market sentiment towards AI remains pessimistic, with major concerns including the peak of data center AI spending in 2025, competition between GPUs and ASICs, and the reduction of TSMC's CoWoS orders. However, JP Morgan believes that the GTC conference should help revive positive sentiment in the market for AI stocks. In the long term, global AI capital expenditure is expected to have room for growth in 2026, primarily driven by increased spending from U.S. cloud providers, a rebound in capital expenditure from Chinese CSPs, and rising enterprise-level AI demand. Blackwell Ultra: A Performance Monster on the Horizon JP Morgan expects that Blackwell Ultra (GB300) will be the highlight of this GTC conference. This chip will adopt a logic chip similar to the B200 but with significant improvements in HBM memory capacity and power consumption. JP Morgan forecasts: HBM capacity skyrocketing: 288GB, utilizing HBM3e 12-high stacking technology; Power consumption rising: thermal design power (TDP) reaching 1.4kW; Significant performance improvement: expected to be 50% higher than B200 in FP4 computing performance; Shipping time: expected to start shipping in the third quarter of 2025. JP Morgan also mentioned that changes in the Blackwell Ultra system will benefit suppliers in areas such as power, batteries, thermal management, connectors, ODM, and HBM. Rubin Platform: The New Engine of AI Computing Power in 2026 Although the Rubin platform is not expected to enter mass production until 2026, JP Morgan believes that NVIDIA may share some details about the Rubin platform at this GTC conference. The Rubin GPU is expected to feature the following configuration: Dual logic chip structure: similar to Blackwell, but equipped with two TSMC N3 process chips; HBM4 memory: 8 HBM4 cubes, total capacity of 384GB, a 33% increase compared to Blackwell Ultra; Power consumption continues to rise: expected to reach approximately 1.8kW TDP; Vera ARM CPU upgrade: expected to migrate to N3 process and may adopt a 2.5D packaging structure; 1.6T network: Rubin platform may adopt a 1.6T network architecture, equipped with two Connect X9 network cards; NVL144/NVL288: possible launch of NVL144 and NVL288 rack structures to increase GPU density; Mass production timeline: Rubin platform may have an earlier mass production timeline, with the earliest mass production starting at the end of 2025 or early 2026, and large-scale shipments expected in the second quarter of 2026. CPO Technology: The Future of Data Center Interconnect JP Morgan believes that NVIDIA's CPO (Co-Packaged Optics) technology roadmap will be another highlight of this GTC conference. CPO helps to increase bandwidth, reduce latency, and lower power consumption, but currently, GPU-level CPO still faces significant technical challenges, such as heat dissipation issues (high heat generation of optical engines), reliability, and the risk of packaging substrate deformation. JP Morgan expects that CPO will initially be applied to switches, namely the Quantum (InfiniBand) and Spectrum (Ethernet) series, as an optional solution for the Blackwell Ultra platform. The widespread application of CPO on the GPU side is expected to be realized no earlier than the 2027 Rubin Ultra era. Physical AI and Humanoid Robots: Growing Attention JP Morgan stated that NVIDIA showcased advancements in physical AI at the previous GTC conference, and given the accelerating development of humanoid robots and physical AI, market attention in this area may significantly increase at this GTC conference. Currently, NVIDIA has released Cosmos (AI foundational model platform) and GR00T (humanoid robot development platform), and more announcements regarding multimodal AI, robotics, and digital twin fields are expected at this GTC conference. JP Morgan also mentioned that the rising sentiment in the robotics field will benefit supply chain manufacturers such as BizLink and Sinbon