Did Alphabet Just End the AI Memory Boom?

Memory stocks got hammered this week after Google dropped a research paper that has investors questioning the entire thesis for the AI-driven memory bull run. Alphabet’s (GOOGL) Google Research group published details on Tuesday of a new compression algorithm called TurboQuant, and the fallout was swift. Sandisk Corporation (SNDK) plunged as much as 11%, Micron…


Did Alphabet Just End the AI Memory Boom?

Memory stocks got hammered this week after Google dropped a research paper that has investors questioning the entire thesis for the AI-driven memory bull run. Alphabet’s (GOOGL) Google Research group published details on Tuesday of a new compression algorithm called TurboQuant, and the fallout was swift. Sandisk Corporation (SNDK) plunged as much as 11%, Micron Technology (MU) dropped 7% on Thursday as the selling accelerated, and Western Digital and Seagate each fell over 7%. The damage extended overseas, with Samsung and SK Hynix both sliding more than 5% in Seoul. As of this morning, the group is rebounding.

Both Micron and SanDisk currently carry a Zacks Rank #1 (Strong Buy). The supply constrained dynamic and exceptional demand for DRAM has pushed these stocks to incredible performances in the last six months, with SNDK up nearly 5x and MU more than doubling over that period.

So is Alphabet building a technology that destroys the bull case for this niche AI infrastructure boom? Or is this a case of headline-driven profit-taking in an overextended sector?

Zacks Investment Research
Zacks Investment Research


Image Source: Zacks Investment Research

At its core, TurboQuant addresses one of the most expensive bottlenecks in running large language models: the key-value (KV) cache. This is the high-speed data store that retains context so a model doesn’t have to recompute everything with each new token it generates. As models process longer inputs, the KV cache balloons, consuming GPU memory that could otherwise serve more users or run larger models.

Google’s algorithm compresses the KV cache significantly, reducing its memory footprint by nearly 6x without sacrificing accuracy or requiring model retraining. Tested across five standard AI models, TurboQuant achieved perfect scores on retrieval tasks. Testing showed that it delivered up to an 8x acceleration in computing attention on Nvidia H100 GPUs.

Beyond LLMs, Alphabet noted that TurboQuant also improves vector search, the technology underpinning everything from Google Search to YouTube recommendations to ad targeting.

The sell-off in Micron, Sandisk, and their memory peers echoed the DeepSeek-driven panic from early 2025, when a Chinese AI lab demonstrated that competitive models could be trained with far less compute than assumed. That episode triggered a one-day massacre in semiconductor names before the market ultimately concluded that efficiency gains accelerate adoption, and adoption drives more hardware demand, not less.

Source link