Google’s TurboQuant Compresses AI Models by 6x, Triggers Memory Stock Selloff

Quick Facts

• TurboQuant compresses AI model memory by at least 6x with zero accuracy loss and up to 8x speed increases
• Memory stocks fell sharply: SanDisk down 5.7%, Micron down 3%, Samsung down 4.7%, SK Hynix down 6.2%
• Independent developers already built working implementations despite Google not releasing official code

Google developed TurboQuant, a compression technology that reduces AI model memory requirements by at least six times while maintaining perfect accuracy. The breakthrough triggered immediate stock market reactions as investors weighed the implications for memory chip demand.

TurboQuant compresses the key-value cache of large language models to 3-bit levels through a two-stage process. PolarQuant converts data vectors from Cartesian to polar coordinates, while QJL uses the Johnson-Lindenstrauss Transform to shrink high-dimensional data while preserving essential relationships.

In benchmarks on Nvidia H100 GPUs, 4-bit TurboQuant delivered up to eight times faster performance computing attention logits compared to unquantized 32-bit keys. The technology achieved perfect scores on needle-in-a-haystack retrieval tasks while compressing memory by at least six times.

Memory stocks fell sharply on the news. SanDisk Corporation dropped 5.7%, Micron Technology declined 3%, Western Digital fell 4.7%, and Seagate Technology slid 4%. Asian markets saw Samsung Electronics fall 4.7% and SK Hynix decline 6.2%. CPU makers Intel and AMD surged as investors anticipated increased demand for processing power.

Wells Fargo analyst Andrew Rocha called TurboQuant “directly attacking the cost curve” but noted concerns about memory demand. “If you’re lowering the specs needed for all this memory, it quickly calls into question how much memory capacity is needed,” Rocha said.

Cloudflare CEO Matthew Prince described the development as Google’s “DeepSeek moment,” representing a major breakthrough in AI efficiency. The technology particularly impacts inference workloads rather than training, which analysts say will have smaller effects on High Bandwidth Memory demand since HBM remains necessary for AI training.

Google has not released official code, but independent developers built working implementations in PyTorch, MLX for Apple Silicon, and C/CUDA for llama.cpp. Industry experts expect rapid adoption through ecosystems like vLLM and Hugging Face, with open-source integrations expected in Q2 and commercial products likely by Q4.

The findings will be presented at the ICLR 2026 conference next month. The paper was co-authored by research scientist Amir Zandieh and VP Vahab Mirrokni.

Analysts argue the technology follows the Jevons Paradox, where increased efficiency leads to higher total consumption. If Google enables models to run on 16GB instead of 96GB of VRAM, developers will likely use the saved capacity for models that are six times more complex.

Read more: Google develops TurboQuant compression technology for AI models