People do not sell DRAM because they read a fifty-page proof. They sell because the headline promised six times less memory and their brain finished the sentence for them.
That reflex is not stupid. It is human. Markets run on partial information and tight deadlines. When the partial information sounds like it threatens the one thing that has been working—AI hardware scarcity—the body moves before the mind catches up.
In late March 2026, Google Research published TurboQuant. In ordinary language, it is a way to squeeze the scratchpad a large language model uses while it is answering you—the key-value cache—so the same chip can hold more conversation history without reaching for heroic tricks. It targets inference, not training. It is not a new NAND drive in your laptop. It is not a memo to DRAM procurement that says stop ordering.
The public write-up opened with big numbers: at least about six times less KV-cache memory, up to about eight times faster attention on H100-class silicon, plus demos that look like needle-in-a-haystack tests. The academic version is fussier—roughly three-and-a-half bits per channel described as quality-neutral for this job, about two-and-a-half bits with marginal loss. Your feed collapsed the nuance into “three bits, zero loss.” Both can be true at different levels of rounding; they are not the same sentence.
Here is a useful split. One number describes an engineering benchmark. The other describes how fear travels. Google, in its own tests, claims at least six times KV-cache savings. CNBC, in the same news cycle, reported SK Hynix down on the order of about six percent. Same digit. Different universes.
Once you see that split, the tape makes more sense. CNBC, Yahoo Finance, and the Seoul Economic Daily tied memory and storage weakness to the TurboQuant story. SK Hynix, Samsung, Kioxia, Micron, SanDisk, Western Digital—they all got mentioned in the wash. NVIDIA slipped a few percent too, in sessions where “AI efficiency” blended with legal headlines and positioning that had little to do with Google’s blog. Korean papers even stacked “Google shock” beside geopolitical noise. One story became a hook; the hook caught everything within reach.
There is a pattern here that shows up outside semiconductors. When a sector has sprinted on a simple story—in this case, AI needs memory, memory is tight—investors carry both profits and fragility. They are waiting for a reason to doubt. They do not need the doubt to be fair. They need it to be legible.
TurboQuant arrived legible.
Now the part that sounds like economics but is really psychology with a spreadsheet. When you make inference cheaper, you often do not get less inference. You get more of it—longer documents, bigger batches, more experiments, more products that were too expensive yesterday. Historians argue about labels; traders shorthand it as Jevons. The casual version is the same as cars: better gas mileage did not end driving. It expanded it.
Could this time be different? Of course. “Could be different” is always true. That is why investing is not physics. TurboQuant also has no stamped calendar for when it shows up inside the serving stacks you actually pay for in the cloud. Until it is ordinary infrastructure, it is closer to a ceiling on how wasteful you have to be about memory than a verdict on next quarter’s wafer starts.
The trade press quoted Cloudflare’s CEO placing the moment next to the DeepSeek lesson: inference still has fat to trim. That helps customers. It also rattles anyone who priced hardware as if fat were permanent.
If you own Micron, Korea-heavy ETFs, SanDisk, Western Digital, or a concentrated semiconductor sleeve, you are not really debating whether PolarQuant rotation is elegant. You are debating whether fear moved faster than facts—and how much room for error you want if it did.
If you own NVIDIA, the puzzle is second-order. Denser inference can mean fewer GPU-hours per question. Cheaper tokens can mean more questions. The net is opaque until utilization and pricing tell you who captured the savings.
The practical checklist is boring on purpose because boring survives headlines: hyperscaler roadmaps, supplier gross margins, guidance on high-bandwidth memory per accelerator, backlog language. Those items update a thesis. A viral thread updates your mood.
The honest close is anti-climactic. You should not rebuild a portfolio around a compression preprint. You should not confuse a violent week for a new law of nature. Headlines sell certainty because certainty sells clicks. Spreadsheets settle the account later—and they are usually less impressed by the story that felt undeniable on Tuesday.