GOOGLE
Google Shrinks AI Memory With Zero Accuracy Loss
Google Research introduced TurboQuant, an algorithm that compresses AI model memory over 6x without any retraining — while delivering up to 8x speed gains on Nvidia H100 chips and losing almost zero accuracy.
- AI models keep a running log of each conversation, and as chats get longer, that storage balloons, which slows responses and drives up costs.
- TurboQuant shrinks that storage by over 6x with zero accuracy loss, scoring perfectly on tests that bury a key detail in a large amount of text.
- On Nvidia's top server chips, it also sped up response processing up to 8x compared to standard methods, without adding any extra cost to run.
- The paper, set to be presented at ICLR 2026 in April, also topped rival methods in vector search — the tech search engines use to match similar results quickly.
Miért fontos?
Despite being first published in April 2025, top AI memory companies felt the heat of the official release, with stocks dropping 3-5%. One compression paper won't crater memory demand overnight, but the selloff shows Wall Street is pricing in a world where smarter software cuts into the premium AI memory commands.