机构：AMD Advancing AI 2026大会或成股价催化剂 MI550X有望领先英伟达Rubin Ultra
Related
SemiAnalysis: NVIDIA Delays or Adjustments to Multiple AI Rack-Scale Architectures, Limiting Rubin Ultra's Expansion Path
According to Mars Finance, on July 6th, SemiAnalysis reported that NVIDIA's latest rack-mount interconnect architecture, Kyber NVL144, has undergone a major adjustment just three months after its release, being delayed by more than 12 months to 2028. The main reason is the ongoing challenges in PCB planar design regarding manufacturing feasibility. Simultaneously, the NVL72x2 back-to-back rack architecture has been cancelled. This solution was originally intended to enhance the scalability of pure copper NVLink by deploying two Oberon racks back-to-back, but its complex structure and high operational burden on hyperscale cloud providers (CSPs) led to strong market skepticism and its eventual abandonment. Furthermore, due to the immaturity of CPO (Co-packaged Optics) technology, NVIDIA's larger-scale scaling solutions based on CPO NVSwitch (such as NVL576) may also continue to be delayed, or limited to small-batch trial production. This means that NVIDIA lacks a stable large-scale scale-up solution until CPO matures. Furthermore, the Rubin Ultra product roadmap has also changed: the originally planned "four-computing-chip" version has been cancelled, with only a "dual-computing-chip" version remaining, and the overall system-level performance is expected to drop to about half of the original solution. These adjustments mean that Nvidia's scale-up capabilities in the Rubin Ultra generation are limited. With the CPO NVSwitch unable to be deployed before the Feynman architecture, competitors such as the AMD MI500X or Google TPU v8i may have a relative window of opportunity in terms of scaling capabilities for large-scale training clusters. Meanwhile, Nvidia is expected to fill market demand during the product transition period and maintain overall supply chain pacing by shipping large quantities of Oberon Rubin racks and their "Ultra" versions.
Research: Nvidia's 800V Power Rack becomes the selected solution for Vera Rubin, with wider adoption expected in the Rubin Ultra generation.
According to a report by Mars Finance on June 25th, based on TrendForce's latest AI Server power architecture research, NVIDIA is actively developing its own 800V HVDC Power Rack solution, aiming to complete inventory preparation by the third quarter of 2026 and provide it to Vera Rubin customers who require it, as a non-standard configuration. Judging from the power consumption design of each rack, NVIDIA's 800V Power Rack should gradually see wider adoption after the Rubin Ultra series in the second half of 2027, with large-scale adoption currently estimated to occur in 2028.
Citrini vs. SemiAnalysis? SemiAnalysis founder challenges Citrini researcher, claiming he revealed information about the Rubin Ultra 12 Hi to clients back in March.
According to Odaily Odaily, a new debate has erupted between two well-known investment research firms, Citrini and SemiAnalysis. Citrini researcher Zephyr cited sources claiming that "Rubin Ultra HBM performance has actually dropped to the 12 Hi level, and they haven't even started using HB technology yet." They will not be using 16-Hi chips in 2027. Currently, hybrid bonding technology is very expensive for HBM (its yield is extremely low). BESI's biggest advantage in the short term lies in TSMC's use of hybrid bonding technology in its chip manufacturing and packaging processes, enabling the integration of EICs and PICs, thereby improving CPO production efficiency. In response, Citrini researcher Jukan commented with a related tweet, "I told you this back in April." In response to this, Dylan Patel, founder of SemiAnalysis, forwarded a screenshot from Discord and said: "In March, customers using the SemiAnalysis Memory and Accelerator models already knew that the Rubin Ultra 16 Hi would be replaced by a 12 Hi version." Previously, Citrini researcher Jukan cited reports that "Nvidia plans to drastically cut its next-generation rack memory configurations," causing a sharp drop in US and South Korean memory stocks. SemiAnalysis founder Dylan Patel clarified that the report was taken out of context and was sensationalist.
SemiAnalysis: Bullish on Nvidia's second-half performance; Rubin ramp-up will drive data center business growth.
BlockBeats reported on June 30th that SemiAnalysis, an independent research firm specializing in semiconductors and AI, predicts significant growth for Nvidia in the second half of this year. According to SemiAnalysis' Accelerator Model, Nvidia's data center computing (DC) revenue in the second half of fiscal year 2027 (2H FY2027) is expected to be 20% higher than market consensus. The Rubin platform, previously delayed due to HBM4-related issues, is now expected to see large-scale production in the second half of this year as these issues are resolved and front-end wafer supply capabilities improve. Unlike some sell-side analysts who tend to set conservative earnings forecasts that allow for "exceeding expectations," SemiAnalysis' forecast is based on in-depth research covering the entire supply chain, including but not limited to materials, wafer manufacturing, components, server manufacturers, as well as the latest developments from hyperscalers and cutting-edge AI labs. Insights about NVIDIA, as well as AI chip companies such as Broadcom, AMD, MediaTek, and Marvell, are all integrated into their Accelerator Model and HBM Model.
Citrini: AMD and Apple are simultaneously pushing forward with flash memory solutions to replace DRAM in AI products.
According to Mars Finance, on June 16th, Citrini Research, the organization behind the "AI Doomsday Report," published an article stating that AMD and Apple are simultaneously advancing solutions to partially replace DRAM with flash memory in their AI products. AMD acquired MEXT to optimize flash memory, bringing its performance close to DRAM, thereby reducing memory costs in AI data centers; Apple, on the other hand, achieves similar optimizations on the device side through its "LLM in a flash" technology. Citrini's latest research report highlights the high memory demands of key-value caching in AI inference, and the "memory tax" pressure from HBM (Hardware Bus) already accounting for 25% of DRAM capacity. Flash memory, costing only 1/55th that of DRAM, can provide a capacity and bandwidth alternative for edge AI through controller optimization, NAND stacking, and cell mode adjustments. This report provides theoretical support for the recent surge in memory stocks, particularly SanDisk.
"Photon Stock God": Nvidia's Rubin architecture upgrade warrants attention from power semiconductor suppliers.
According to Mars Finance, on June 8th, newly emerging star US stock analyst KawzInvests (@KawzInvests) published an article stating that NVIDIA's Rubin Ultra AI rack power architecture will be upgraded in 2027, transitioning from 48V/54V to an 800V DC system. The value of power semiconductors per rack will surge from approximately $20,000 to $123,000, primarily due to the increase in power from 225kW to 600kW, requiring advanced materials such as silicon carbide and gallium nitride. KawzInvests analysis suggests that this upgrade aims to address the weight and efficiency issues of copper buses under high power conditions, with the increased value primarily flowing to power semiconductor suppliers rather than NVIDIA or rack manufacturers. Potential beneficiaries include: IFX, MPWR, VICR, NVTS, STM, ON Semiconductor, TXN, Analog Devices, POWI, DIOD, WOLF, and AOSL.