GO
| HSI1 | 24,917.60 | +111.97 | 192.63B |
| HSCEI1 | 8,284.58 | +38.25 | 51.08B |
| Back Zoom + Zoom - Block Traded | |
|
2026-09-14 11:42:49 CLSA published a research report stating that DeepSeek's latest model, DeepSeek-V4.1-Flash, was launched last Thursday (10th). The model reduces HBM-based KV cache demand by 75% (to 890 bytes per token) and lowers eSSD cache demand by 87.5%, seemingly putting pressure on memory stocks. The broker emphasized that the improvement only applies to KV cache efficiency and does not imply a reduction in overall GPU memory or storage demand, while maintaining its positive view on memory stocks. The broker noted that AI models have long driven demand through improvements in memory efficiency. DeepSeek improved per-token memory efficiency by about 110x from V1 to V4, yet memory demand continued to grow as lower costs encouraged larger usage volumes. Token demand doubled in August, and model efficiency improvements together with enhanced memory/computing performance are expected to continue driving overall memory demand. The 75% reduction only targets KV cache during the inference stage, rather than model weights, training memory, or broader HBM usage. If inference costs decline, developers are likely to run more agents, longer context content and more inference workloads, meaning token processing volume could rise much faster than the decline in per-token memory demand. Even if cache compression lowers capacity demand, AI models still require massive bandwidth to transfer data between GPUs and memory, meaning lower capacity intensity may not proportionally reduce HBM demand. The broker believes software innovation will continue improving memory efficiency, and investors should expect recurring headlines about "reduced memory usage". However, exponential growth in computing demand, expanding context length and AI agents significantly increasing inference volume mean the net result is still expected to be strong growth in total memory demand even as per-token memory consumption declines. The broker rated SK Hynix (000660.KS) as "High Conviction Outperform" with a TP of KRW3.7 million; Samsung Electronics (005930.KS) as "Outperform" with a TP of KRW400,000; and Micron Technology, Inc. (MU.US) as "Outperform" with a TP of USD1,700. The broker expects global HBM TAM to achieve a 40% CAGR through 2028 and remains optimistic about capital expenditure visibility among hyperscalers. (da/ad)~ AASTOCKS Financial News Website: www.aastocks.com This article was automatically translated by AI, the Chinese version should be considered the authoritative version. AASTOCKS.com Limited does not guarantee its accuracy or completeness and accepts no liability for any damages or losses arising from the use of this translation. | |