Publisher
Stories with reporting from SemiAnalysis, ranked by what is gaining ground now. Each story links to the original articles.
SemiAnalysisNvidia announces Rubin CPX, an accelerator designed to prioritize compute for the prefill phase of inference. The account highlights its emphasis on compute FLOPS over memory bandwidth and introduces a rack-scale system built around it.
SemiAnalysisAn analysis of xAI’s Colossus 2 discusses its plan to build a gigawatt-scale data center, alongside a distinct reinforcement-learning methodology and capital raise. It places the project in context with Colossus 1, which the account describes as a roughly 300 MW cluster with about 200,000 H100/H200 GPUs and around 30,000 GB200 NVL72 systems.
SemiAnalysisThe story concerns Huawei Ascend production, including die banks, continued TSMC production, and high-bandwidth memory as a supply bottleneck. It frames access to compute as a major constraint on AI production.
SemiAnalysisThe account discusses Meta's purchase of 49% of Scale AI at an approximately $30B valuation and Meta's position in foundation-model development. It frames Meta's spending capacity alongside concerns that its models are falling behind competitors.
SemiAnalysisThe analysis argues that AWS is struggling to translate its cloud-market strength into the GPU and XPU cloud era, while Microsoft Azure leads in that area. It describes AWS as Amazon's profit engine, generating about 60% of group profits, and examines a potential expansion involving AWS, Anthropic, and Trainium.
SemiAnalysisA report compares Nvidia H100 and GB200 NVL72 systems for frontier-model training, examining power, total cost of ownership, reliability, and performance. It also considers how software improvements over time affect the comparison.
SemiAnalysisAn analysis traces how advances in AI are changing the prospects for robot autonomy, arguing that modern models turn many longstanding robotics challenges into data problems. It examines the shift from robots designed for narrowly controlled settings toward broader capabilities.
SemiAnalysisA VLSI conference roundup examines semiconductor manufacturing and design topics, including fab digital twins, advanced logic transistors, interconnects, and DRAM. The coverage also discusses Intel 18A and the future of DRAM designs.
SemiAnalysis To many power users (Pro and Plus), GPT5 was a disappointing release. But with closer inspection, the real release is focused on the vast majority of ChatGPT’s users, which is the 700m+ free userbase that is growing rapidly. Power users should be disappointed; this release wasn’t…
SemiAnalysisA report examines high-bandwidth memory, including its manufacturing, suppliers, and uses in AI systems such as KVCache offload and disaggregated prefill and decode. It also explores HBM4 and the development of custom base dies.
