Samsung Electronics shares reportedly fell 3.5% and SK Hynix 2.2% in South Korea on September 11 after reports that DeepSeek\u2019s AI models require less high-bandwidth memory \u2014 a real market jolt that does not prove global HBM demand has permanently cracked, NeoTeo reported.<\/p>\n
DeepSeek\u2019s Multi-Head Latent Attention (MLA) compresses key\/value attention state in the KV cache, and Engram describes conditional memory that can sit in an offloaded hierarchy. Secondary accounts attributed figures such as DeepSeek V4 using about 10% of DeepSeek V3.2\u2019s KV-cache requirement at a one-million-token context and 27% of its single-token inference FLOPs \u2014 specific workload metrics, not measurements of HBM shipments or chip revenue.<\/p>\n
Model weights, activations, multi-accelerator communication, and total request volume still drive memory needs. Cheaper inference can even expand total demand if providers run more queries. NeoTeo\u2019s takeaway: DeepSeek may ease accelerator-memory pressure in some inference cases, but the September 11 selloff is not evidence that HBM demand has permanently fallen.<\/p>\n