One algorithm. Six times less memory. Zero accuracy loss. And the entire investment thesis for the AI hardware boom just got compressed along with it.
Memory-chip stocks were supposed to be the safe way to play AI. The picks and shovels. The people selling jeans to the gold miners.
Now the gold miners have figured out how to pan without jeans.
The entire memory sector is under pressure. Not because of poor earnings. But because of a research paper that a team from Alphabet’s GOOG GOOGL Google published in April 2025.
That paper led to Google’s unveiling last month of TurboQuant — a compression algorithm that reduces the memory an AI model needs by 6x. Makes it 8x faster on the same GPUs. Loses zero accuracy.
Here’s how to explain this to your financial adviser: Every time an AI model has a conversation with you, it remembers what you’ve said. That memory lives on expensive chips. As conversations get longer, the memory required becomes enormous — hundreds of gigabytes per session.
That is why memory prices soared 7x in three months. Demand was insatiable. Supply was deliberately constrained. The memory producers were printing money the way the U.S. Treasury prints money — except with better margins and less apology.
TurboQuant compresses that memory by a factor of six. It’s the same conversation and same accuracy, but with ⅙ th of the chips. The analogy is shrinking a 4K image to the size of a postage stamp — except the picture is still 4K.
Cloudflare’s CEO called this “Google’s DeepSeek moment”— referring to China’s disruptive AI model. The internet called it “Pied Piper,” after the fictional compression algorithm in HBO’s “Silicon Valley” series.
Within 24 hours, developers had ported TurboQuant to Apple Silicon and every popular open-source AI library. A benchmark on the Qwen model showed perfect accuracy at every compression level. 100% exact match at much less memory. Its implementation requires no retraining. No fine-tuning. It drops into existing systems like a software update.
That is the part that should concern you if you own memory stocks. This is not a road map for 2028. This is a patch you can install before lunch.
Read: Why TurboQuant hammered memory stocks — and why “Jevons’ Paradox” means the market might be wrong
The memory bull case rested on one assumption: AI demand for DRAM and NAND would outstrip supply indefinitely.
South Korea’s Samsung KR:005930 and SK Hynix KR:000660 operated with what analysts politely call “supply discipline” and what everyone else calls a cartel. They refused to add capacity. Margins went vertical. Goldman Sachs told investors it was structural.
It was structural the way a dam is structural — until someone finds a way around it.
Google found a way around it. The algorithm is free. The paper is public. The code is already being replicated by developers in coffee shops on three continents.
If TurboQuant reduces memory demand by even half of what the paper claims, the fair value of the memory sector re-prices by 40% to 60%.
TurboQuant is the opening paragraph — not the conclusion — of the memory-stock correction. Micron Technology MU shares are up 28% so far this year — but pre-TurboQuant, the stock was up more than 60%. Rivals Sandisk SNDK, Western Digital WDC and Seagate Technology STX have also since come off their 2026 peaks.
TurboQuant compresses the KV cache — the temporary memory storing mid-conversation computations. It does not compress the model’s permanent weights. That is a harder problem. Google did not solve that. Not yet.
The research tested open-source models, such as Gemma, Mistral and Llama. Not Google’s Gemini running at production scale, serving billions of requests.
“Zero accuracy loss” is a lab result. The gap between the lab and the data center is where reputations go to be tested and frequently to die.
But the market is not pricing the fine print. The market is pricing the implication.
If Google achieved 6x compression today, someone will achieve 10x tomorrow. Software efficiency improvements compound. Hardware scarcity assumptions do not survive compounding software improvements.
They never have. Ask anyone who sold magnetic tape in 1985.
Read: Micron and Sandisk stock have been in a slump. Buy the pullback, analyst says.
1. AI capex deceleration: The $200 billion hyperscaler build-out just got cheaper. That sounds bullish for AI companies. It is bearish for everyone selling them hardware. The capex boom propping up GDP growth just got a deflationary haircut from a research paper.
AI is deflating its own supply chain. Meanwhile, oil is comfortably north of $100 a barrel, with an Iranian toll booth on the Strait of Hormuz.
The digital economy is getting cheaper to run — exactly as the physical economy is getting more expensive to operate.
The U.S. Federal Reserve is trapped between two forces moving in opposite directions. TurboQuant just made the deflationary force stronger and the rate-cut decision harder.
3. South Korea: Samsung and SK Hynix represented roughly 30% of the country’s Kospi Index. If the memory premium structurally compresses, Korea’s current account deteriorates and the won USDKRW weakens.
Moreover, a critical American ally becomes less economically stable at the worst possible geopolitical moment. SK Hynix just announced an $8 billion ASML equipment order and a U.S. stock listing to expand production of the exact memory type that Google’s algorithm just made 83% less necessary. Timing is everything. Sometimes the timing is terrible.
Avoid memory-chip stocks. Micron. Sandisk. Western Digital WDC. Seagate. And Samsung and SK Hynix through any Korean ETF. The broad semiconductor funds, such as VanEck Semiconductor ETF SMH and iShares Semiconductor ETF SOXX, carry embedded memory exposure that is now toxic.
What benefits: companies that deploy AI, not those supplying the hardware. Cheaper inference means faster adoption. Faster adoption means revenue for the companies using AI, not the companies selling the memory to run it.
More: There’s a new ETF for memory stocks. History suggests that might be an ominous sign.
What does not change: the physical economy. Oil. Copper. Gold. Defense. Critical minerals. Nuclear fuel. You cannot compress a barrel of oil with an algorithm. You cannot algorithmically reduce the amount of tungsten in a penetrator round. You cannot write software that replaces the 45 million barrels per day of new production the International Energy Agency says the world needs by 2050.
Google just published an algorithm that reduces AI memory requirements by 83%. The algorithm is free. The takeaway for investors: Own what cannot be compressed.
More from Charlie Garcia:
Big Tech’s AI fantasy hits a nuclear wall: No fuel, no welders — and no Plan B
Kevin Warsh isn’t who investors think he is — how you can profit from their mistake