Memory markets reacted sharply to TurboQuant, Google’s algorithm that reduces KV-cache size by at least six times and speeds up specific attention-logit operations by as much as eight times. The important qualification is that this is not a sixfold reduction in all memory used by a model; it targets a particular inference bottleneck.
Shares of memory and storage manufacturers fell noticeably. Depending on the period used for comparison, Micron lost around 20%, SanDisk roughly 10%, and several Asian manufacturers declined by single digits. Headlines quickly converted that reaction into a simple conclusion: demand from the AI industry is about to collapse.
Now let us examine how the manipulation works. Doing that requires digging into the facts — difficult work when you also have a personal life.
- The TurboQuant research first appeared in spring 2025; the broad public presentation followed in March 2026 ahead of ICLR.
- Many independent frontier companies and individual AI divisions operate under enormous cost pressure. But saying that every major player is unprofitable would be wrong: Google and Microsoft remain profitable corporations. Optimization is about efficiency and scale, not a simple instruction to buy less hardware.
- Algorithmic optimization did not begin yesterday. KV-cache and other memory-saving techniques became standard industry tools long before TurboQuant.
- TurboQuant primarily targets inference and vector representations. It does not reduce the cost of model training by the same factor of six.
The conclusion
Google built a powerful mechanism that reduces costs at a specific point in the system. But concluding that hardware will automatically become cheaper or less important is far too linear.
Freed capacity will most likely be redirected toward longer contexts, heavier models, and new workloads. This is a familiar effect: higher efficiency often expands the use of a resource instead of reducing total consumption.
The build-out will not stop. The market is still in an aggressive land-grab phase, and there will be few winners. My bet is on Gemini, Grok, and China.
Large players also enjoy the old game of “buy low, sell high,” which is why stories like this should always be read beyond the headline.
Think.