Micron hit an all-time high on March 18. Six days later, it’s down 22%. SanDisk is down 20%. SK Hynix fell 6% today. Samsung dropped 5%. Western Digital lost 4.7%.

What happened? Google unveiled an AI breakthrough that could make memory chips a lot less necessary.

Welcome to the efficiency era.

On Tuesday, Google unveiled TurboQuant, a new compression algorithm that reduces the amount of memory required to run large language models by six times—with zero accuracy loss.

That’s not a typo. Six times less memory. Zero performance hit.

The algorithm targets something called the “key-value cache,” which is essentially an AI model’s working memory. As AI models process longer inputs, this cache grows rapidly, eating up GPU memory. TurboQuant compresses the cache from 16 bits per value down to just 3 bits, shrinking the memory footprint by at least 6X.

On NVIDIA H100 GPUs, TurboQuant delivered an 8X speedup in computing attention logs, the critical operations that make AI models work.

The implications are immediate: if you can run the same AI workload with 6X less memory, you need to buy a lot fewer memory chips.

The market got the message instantly. Memory chip stocks crashed across the board on Wednesday and continued falling Thursday. Analysts at BTIG noted that Micron hasn’t fallen 20% in six days after hitting a 52-week high since 1999. “When good news gets sold, pay attention,” the firm said.

This isn’t just a one-day selloff. It’s a fundamental reassessment of how much memory the AI industry actually needs.

Here’s how TurboQuant works, in simple terms: AI models store past calculations so they don’t have to recompute everything from scratch. This stored information—the key-value cache—lives in high-speed memory and grows as the model processes more data.

The problem is that this cache can become massive, consuming GPU memory that could otherwise be used to serve more users or run larger models.

Traditional compression methods exist, but they usually require storing additional constants and normalization values that add 1-2 extra bits per number, partially undoing the compression. TurboQuant eliminates this overhead through a two-stage process.

The first stage, called PolarQuant, converts data from standard Cartesian coordinates into polar coordinates—separating each vector into a magnitude and a set of angles. This makes the data easier to compress without needing extra normalization steps.

The second stage, called QJL (Quantized Johnson-Lindenstrauss), handles the tiny errors left over from the first stage. It reduces each remaining value to a single sign bit—positive or negative—with zero memory overhead. The result is 3-bit compression with no accuracy loss.

Google tested TurboQuant across standard benchmarks including question answering, code generation, and summarization tasks using open-source models like Gemma and Mistral. The results showed 100% exact match rates with zero performance degradation.

The internet immediately started calling it “Pied Piper”—a reference to the fictional compression algorithm from the HBO show Silicon Valley that achieved impossible compression ratios. The comparison isn’t perfect, but the sentiment is real: this feels like a shift.

Some are calling it Google’s “DeepSeek moment”—a reference to the Chinese AI model that achieved competitive results while being trained at a fraction of the cost on inferior chips. DeepSeek proved that efficiency could compete with brute force. TurboQuant is the same bet applied to memory.

Google isn’t alone. NVIDIA has a rival compression method called KVTC that achieves 20X compression, compared to TurboQuant’s 6X. The catch is that KVTC requires a one-time calibration step per model, while TurboQuant works immediately without any tuning. Both methods will be presented at the ICLR 2026 conference in late April.

The stock market reaction was swift. Wells Fargo analyst Andrew Rocha noted that TurboQuant directly attacks the cost curve for memory in AI systems. If broadly adopted, it raises the question: how much memory capacity does the industry actually need?

But analysts also cautioned that the demand picture for AI memory remains strong. Compression algorithms have existed for years without fundamentally altering procurement volumes. Meta alone committed $27 billion recently for dedicated compute capacity. Google, Microsoft, and Amazon are collectively planning hundreds of billions in data center spending through 2026.

A technology that reduces memory requirements by 6X doesn’t reduce spending by 6X, because memory is only one component of a data center. But it does change the calculus.

Here’s what investors need to understand: this isn’t about memory chips becoming obsolete. It’s about the growth rate slowing. AI companies were building infrastructure assuming they’d need massive amounts of high-bandwidth memory to handle growing context windows and larger models. TurboQuant suggests they might get away with less.

The timing matters too. Big tech has been spending unprecedented amounts on AI infrastructure—$650-700 billion estimated for 2026. Memory chips were a major beneficiary. Companies like Micron, SK Hynix, and Samsung were riding the AI wave to record valuations.

Now the market is pricing in the possibility that software efficiency could slow hardware demand. It’s not that AI is slowing down—it’s that AI might not need as much hardware as everyone thought.

For Google, TurboQuant has direct commercial applications beyond language models. The algorithm improves vector search—the technology that powers semantic similarity across billions of items. That’s the foundation of Google Search, YouTube recommendations, and advertising targeting. In other words, it underpins Google’s revenue.

The research paper is real. The compression results are validated. Independent developers have already built working implementations in PyTorch, even without official code from Google. One developer tested it on an RTX 4090 and reported character-identical output to the uncompressed baseline at 2-bit precision.

So is this the end of the memory chip boom? Probably not. AI infrastructure spending is growing at extraordinary rates, and memory remains essential. But the growth trajectory just got a lot less certain.

The lesson for investors: efficiency is the new scale. The AI industry spent the past two years building bigger models with more parameters on more chips. Now the focus is shifting to doing more with less. Companies that can compress, optimize, and extract more performance from existing hardware have an edge.

Memory chip stocks aren’t dead. But the easy bull case—AI needs infinite memory, so buy memory stocks—just got a lot more complicated.

Guardian Financial Publishing, 3571 Far West Blvd · Texas · Austin · 78731 · United StatesNo longer want to receive these emails. Unsubscribe