Google’s TurboQuant dents memory-chip stocks, but analysts say ‘buy the dip’

Market shock, brief and sharp

Shares in global memory-chip makers slid after Google said an AI algorithm called TurboQuant dramatically cuts the memory demands of key-value (KV) caches — a core part of how large models serve users. Samsung and SK Hynix shares tumbled, and Chinese memory firms including GigaDevice Semiconductor (兆易创新) and Montage Technology (蒙特科技) saw declines on Shanghai trading as investors feared that extreme compression would reduce future memory demand.

What TurboQuant does — and what’s confirmed

Google said TurboQuant achieves “extreme compression,” reportedly reducing KV cache memory needs by roughly sixfold, a claim detailed in a research paper that, it has been reported, was published in April. That kind of efficiency matters: KV caches sit at the centre of inference and serving costs for generative AI. So a smaller footprint on paper looks like fewer chips needed per deployment. But is that the whole story?

Analysts argue the opposite — meet the Jevons Paradox

Analysts from Morgan Stanley and elsewhere pushed back, calling the sell-off a buying opportunity. Shawn Kim, head of Asia technology research at Morgan Stanley, said TurboQuant should lower inference costs and therefore increase throughput per chip, making AI cheaper to run — and, crucially, spurring more applications and users. They invoke the Jevons Paradox: efficiency gains can raise overall demand because the service becomes cheaper and more widely used. Markets are grappling with this demand-side uncertainty even as valuations for memory and semiconductor stocks remain elevated.

Bigger picture: geopolitics and the AI infrastructure race

The episode comes as the world’s chip supply chain is already strained by trade policy and export controls that have pushed China to accelerate domestic memory capacity. For Western investors, the immediate takeaway is tactical: a short-term price hit, but analysts say buying the dip if you believe AI infrastructure buildout continues. For policymakers and industry chiefs, TurboQuant is a reminder that technical advances can reshuffle demand in unexpected ways — and that efficiency does not always mean less strategic urgency in the semiconductor race.