DDR5 prices plunged
DDR5 RAM street prices dropped roughly $40–$100 for common 32GB kits after reports that Google’s TurboQuant KV cache tech can cut AI memory needs by about 6x — a potential cost shift for ML rigs and systems projects. (x.com)
Google Research published the TurboQuant compression method on March 24, 2026, crediting lead authors Amir Zandieh and Vahab Mirrokni and describing KV-cache quantization down to about 3 bits per channel. (research.google) Google’s blog and independent benchmarks report TurboQuant reduces LLM KV-cache memory by roughly 6× and can deliver up to an 8× attention-speed boost on Nvidia H100 accelerators with no measured accuracy loss in their tests. (tomshardware.com) Community ports appeared within days: a PyTorch implementation (tonbistudio/turboquant-pytorch) validated compression claims on Qwen2.5-3B and an RTX 3060, showing the technique is already reproducible on consumer GPUs. (github.com) Equity markets reacted immediately—shares of Samsung, SK hynix and Micron fell across global markets after the announcement, with Korean disclosures showing Samsung down about 4–5% and SK hynix around 6% on the sell-off days. (cnbc.com) Retail DDR5 listings tracked by hardware sites show specific 32 GB kits sliding from roughly $440 to ~$370 for Corsair Vengeance RGB (2×16 GB) and similar markdowns observed on Amazon and Newegg this week. (notebookcheck.net) Price-trackers and industry outlets flagged a wider, first notable weekly drop in DDR5 after months of gains, with several outlets documenting reductions across multiple vendors rather than isolated promos. (wccftech.com) Market analysts and several reports cautioned the reaction may be overdone, noting TurboQuant compresses working-memory use but does not erase long-term DRAM demand from larger future models or data-center expansion plans. (cnbc.com) TurboQuant is scheduled for formal presentation at ICLR 2026 (April 23–27, Rio de Janeiro), giving researchers and infrastructure teams a venue to inspect method details and implementation trade-offs next month. (iclr.cc)
