Google's TurboQuant 6x compression rattles memory chip stocks
Get the Finance newsletter
Daily finance — markets, central banks, M&A, the prints that move money. Free.
- Google unveiled TurboQuant, a compression algorithm that reduces AI memory needs by 6x and runs 8x faster on the same GPUs with zero accuracy loss, derived from a research paper the company published in April 2025.
- Memory prices had soared 7x in three months before TurboQuant, fueled by what the author calls cartel-like supply discipline from Samsung and SK Hynix, who refused to add capacity as margins went vertical.
- Cloudflare's CEO dubbed TurboQuant "Google's DeepSeek moment," and within 24 hours developers had ported it to Apple Silicon and every popular open-source AI library — requiring no retraining or fine-tuning.
- Micron shares remain up 28% year-to-date but were up more than 60% pre-TurboQuant, while Sandisk, Western Digital, and Seagate have retreated from their 2026 peaks.
- TurboQuant's limitations are explicit in the source: it compresses the temporary KV cache, not the model's permanent weights, and was tested on open-source models like Gemma, Mistral, and Llama — not Google Gemini at production scale.
- The $200 billion hyperscaler capex build-out just got a "deflationary haircut," the author argues, meaning cheaper AI economics for tech giants but a bearish signal for the hardware vendors selling into that build-out.
- Goldman Sachs had told investors the memory shortage was "structural" — the author frames TurboQuant as the moment that structural assumption broke, with the sector's fair value potentially re-pricing 40% to 60% if the compression holds even partially.
Why it matters: Memory stocks had traded as a 'picks and shovels' safe bet on AI, propped up by deliberate supply constraint from Samsung and SK Hynix that pushed prices 7x higher in three months; a free, public compression algorithm that needs no retraining potentially turns that scarcity thesis into a software-efficiency story, reshaping how investors value Micron, Sandisk, Western Digital, and Seagate — and, per the author, trimming the AI capex tailwind supporting GDP-adjacent hardware spending.
