TurboQuant is a Google Research method that compresses the KV cache, a language model’s working memory for long context, at least sixfold with no loss of quality on Google’s tests. When Google described it in a blog post on 24 March 2026, memory makers’ shares lost 3–6% over two days, and flash companies such as SanDisk considerably more: investors feared that AI would need less of this expensive memory. The fear proved unfounded. An independent test in vLLM showed that the method saves memory at the cost of speed and accuracy, and by the summer memory makers had posted record revenue.
What TurboQuant is
The method compresses not the model itself but its KV cache, the data a model keeps in memory so that it can hold on to long context, and also the vectors used in vector search. It is not a March invention: the paper has been on arXiv since April 2025, and in April 2026 it was presented at the ICLR conference. The March blog post was a big publicity push ahead of the conference.
Google claimed at least sixfold KV cache compression with no losses on long-context tests, and up to eight times faster attention computation. Two caveats. “Up to eight times” is measured against an uncompressed 32-bit baseline that has long since dropped out of real-world use, and it applies to a single operation, not to the model’s whole response. The method was tested on models with 7–8 billion parameters.
How the market reacted
On 25 March in the US, SanDisk fell almost 6%, Western Digital about 5%, Seagate 4% and Micron 3%. The next day in Asia, SK hynix and Kioxia lost about 6% and Samsung almost 5%. CNBC explicitly linked the drop to TurboQuant, although some analysts called it profit-taking after a long rally. According to Bloomberg, as early as 27 March Samsung and SK hynix, which make HBM memory for AI, won back most of their losses during the trading day.
What the tests showed
- vLLM, a popular engine for running models, added TurboQuant in April and May, and in May Red Hat AI engineers published a test on the vLLM blog covering four models, from 30 to over 200 billion parameters. The verdict: rather than rising, throughput fell to 66–80% of the standard setup, and the 3-bit variants lost quality noticeably on long context and hard tasks. The authors of the test kept plain 8-bit compression as the best default, with TurboQuant as an option for when memory is very tight.
- llama.cpp, an engine for running models on your own computer, never took TurboQuant on: the pull requests were not merged, the feature request was closed automatically for inactivity in August, and the engine makes do with a simpler rotation and 4-bit cache compression.
- Competitors, the authors of the RaBitQ and EDEN methods, published papers in April claiming that in a fair comparison TurboQuant falls short of their methods and that some of its results cannot be reproduced.
Google has not said whether it uses TurboQuant in Gemini.
What happened to the memory market
Instead of a slump, records. In the second quarter of 2026 SK hynix earned 60.5 trillion won in operating profit on revenue of 79.3 trillion. Micron posted record revenue of $41.5 billion in its fiscal third quarter, against $9.3 billion a year earlier. The next reports were stronger still: Micron took a record $54.2 billion in its fiscal fourth quarter, and SK hynix passed 100 trillion won in quarterly revenue for the first time in the third quarter (104.6 trillion). Contract prices for conventional DRAM were expected, according to a TrendForce forecast, to rise by another 13–18% in the third quarter. The shares more than recovered from the March fall: by the summer SanDisk was up several times over since the start of the year. From July memory stocks went through a marked correction, on worries about AI spending and competition from China’s CXMT, but at the end of October they were still far above their March levels.
The bottom line
TurboQuant is good research, but not a death sentence for the memory market. Saving memory has not yet reduced demand for hardware: the cheaper it is to run a model, the more models get run. The main lesson of March is that the market reacts to a headline faster than to test results, and those who sold their shares in the March panic lost more than the summer and autumn correction would have cost them.
On the other side of the same market is xAI’s gigawatt cluster, where the bottleneck is no longer chips but electricity.
