Alphabet’s memory‑cutting AI
Alphabet unveiled an AI algorithm that significantly reduces memory needs — analysts say that technical shift could pressure chip stocks even as it accelerates broader AI adoption. The change is already being discussed as a structural factor for long‑term hardware demand. (x.com)
Google Research published TurboQuant on March 24, 2026, authored by Amir Zandieh and Vahab Mirrokni, and said the work will be presented at ICLR 2026. (research.google/blog) The paper says TurboQuant pairs PolarQuant with a Quantized Johnson‑Lindenstrauss (QJL) step to compress high‑dimensional vectors in the key‑value (KV) cache and reports at least a 6x reduction in KV memory with no downstream accuracy loss. (research.google/blog) Independent reporting and benchmarks highlighted up to an 8x speedup on NVIDIA H100 accelerators and estimated inference cost reductions of roughly 50% in early tests. (venturebeat.com) The market reaction was immediate: SK Hynix and Samsung shares fell about 6% in Seoul, while U.S. memory names including Micron, Western Digital and SanDisk slid roughly 7% following the disclosure. (bloomberg.com) Analysts stressed the technical scope: TurboQuant focuses on inference KV‑cache compression rather than the high‑bandwidth memory used for large‑scale model training, suggesting training‑HBM demand may be less affected in the near term. (pcmag.com) At the same time, industry voices warned about second‑order effects — SemiAnalysis’s Ray Wang said removing bottlenecks can drive higher overall memory use as models scale, and Morgan Stanley analyst Shawn Kim flagged potential long‑term upside from wider adoption despite the initial sell‑off. (cnbc.com; bloomberg.com) Community uptake was rapid: developers began porting TurboQuant to local inference projects such as MLX and llama.cpp within 24 hours, and GitHub discussion threads show independent implementations and early reproduction benchmarks. (venturebeat.com; github.com)
