A discussion on X about potential shifts in HBM (High Bandwidth Memory) demand for AI systems. The analysis suggests that rather than continuously increasing HBM capacity per GPU, future designs may use smaller HBM stacks combined with interconnect technologies like NVLink to distribute memory across larger clusters. This could pressure HBM wafer consumption by 2028 as new production capacity from SK Hynix, Micron, and Samsung comes online, though long-term contracts and pricing may protect memory manufacturers' margins.
A ByteDance group discussion analyzing AI industry trends covers programmer career prospects amid AI displacement, showing value comes from building abstraction layers rather than executing code. It also discusses Nvidia's shift toward direct optical engine procurement, disrupting traditional module suppliers, and how DeepSeek's reduced HBM requirements paradoxically increase storage value while expanding overall AI capex.
A post discusses how massive capital expenditures by US hyperscalers on AI infrastructure are flowing to South Korean companies, particularly SK Hailey, due to dependence on HBM (high-bandwidth memory) chips. The author argues that while the US captures about 50% of AI profits, the actual capital is being transferred to South Korea at an unprecedented rate as US tech giants must purchase components to make their expensive GPUs functional.
NVIDIA Research demonstrates accelerated video generation using MiniMax-H3 on a single DGX Spark, producing 768p video in one minute through a two-stage pipeline that generates a 384p draft and refines it to higher resolution, with detailed prompts for stop-motion and photorealistic video examples.