TurboQuant Shockwave: Google Algorithm Triggers 5-6% Memory Chip Selloff

Samsung fell 5%, Micron 3%, SK Hynix 4% and SanDisk 6% in two days after Google published TurboQuant on Tuesday. Investors dumped the names because the algorithm compresses the key-value cache in transformer models to just 3 bits, cutting DRAM demand by up to 6× for the same context window. The knee-jerk reflected fears that every H100-class GPU would need fewer or smaller memory modules right as hyperscalers are still scrambling for supply.

Google researchers stripped the KV cache—the working memory that stores attention states—to one-twentieth of its usual footprint without measurable accuracy loss. On an Nvidia H100 the method yields an 8× speed-up on long-context inference. The trick is training-free quantization plus a learned lookup table that rehydrates compressed vectors on the fly. Crucially, it only works on the cache, not the model weights, so the 100 GB-plus parameter blob still needs to live somewhere.

Wall Street views the rout as premature. Bloomberg Intelligence notes that TurboQuant is still a lab demo; no cloud provider has announced adoption timelines. Memory demand is driven by training clusters where weights dominate capacity, not by inference caches. Analysts point out that cheaper inference usually expands usage, so the net effect on DRAM could be neutral or even positive. The selloff may simply be an excuse to lock in gains after a 200% memory-chip rally since early 2024.

Taiwan’s Nanya and Winbond also slid, while logic-heavy Nvidia and AMD barely moved. The split shows investors are discriminating between what TurboQuant touches (cache DRAM) and what it does not (HBM and high-bandwidth logic). Contract manufacturers TSMC and GlobalFoundries could benefit if lower memory pressure frees up reticle space for more compute tiles.

Google says TurboQuant is “ready for integration” into TPU and GPU runtimes. Cloud providers will test it first on cost-sensitive services like long-document summarization. If adoption spreads, expect new tiers of high-memory GPUs to be re-balanced, but not eliminated. Memory makers are already downplaying risk: Micron cites continued AI model growth as the primary driver. Investors will watch Google I/O in May for any TurboQuant rollout and hyperscaler capex plans due in July.