TLDR
- Google’s TurboQuant algorithm compresses AI memory requirements by a minimum of 6x, triggering sharp declines in Micron and Sandisk stock prices
- Market participants worried that reduced memory needs would translate to weaker chip demand, sending both equities down more than 15% from peak levels
- Mizuho’s Vijay Rakesh maintains Outperform ratings on both companies, calling the decline excessive
- Rakesh contends that efficiency gains like TurboQuant typically drive higher AI adoption rates, increasing total memory demand through Jevons’ Paradox
- AI server NAND content has grown twofold over twelve months, with spot prices climbing consistently quarter after quarter
Shares of Micron and Sandisk experienced significant selling pressure following Google’s publication of research detailing TurboQuant, an innovative compression algorithm that reduces AI model memory requirements by a factor of six or more. The technology simultaneously delivers inference speed improvements of up to eight times while maintaining model accuracy.
Market participants interpreted this development negatively for memory semiconductor manufacturers. The logic was straightforward: if AI models require less memory per deployment, demand for chips from producers like Micron and Sandisk would decline accordingly.
Both equities have retreated at least 15% from peak valuations reached in late April. During Thursday’s trading session, Sandisk shares declined 5.9% to $652, while Micron fell 5.5% to $347.78.
Market sentiment was further pressured by President Trump’s Wednesday evening remarks, which failed to provide clarity on when the Iran conflict might conclude, creating additional uncertainty as Thursday’s trading began.
Initial TurboQuant research emerged in 2025, with Google scientists publishing updated findings on AI inference performance gains in recent weeks.
Mizuho’s Case Against the Market Reaction
Vijay Rakesh, an analyst at Mizuho Securities, challenged the market’s interpretation in his client communication. He maintained Outperform ratings on both Micron and Sandisk while keeping price targets of $530 and $710 intact.
Rakesh advised clients to “buy the TurboQuant memory pullback,” dismissing concerns about peak memory demand as “overblown.”
His central thesis rests on historical precedent showing that AI efficiency improvements typically increase total spending rather than reducing it. This phenomenon, known as Jevons’ Paradox, describes how making a resource more efficient or affordable generally increases aggregate demand.
Rakesh cited three historical examples supporting his position. Server virtualization was anticipated to reduce hardware demand but actually expanded it. DeepSeek’s 2025 introduction raised concerns about GPU market saturation, yet AI infrastructure investment continued accelerating. The transition from copper to optical networking, despite offering 10x bandwidth improvements, drove higher AI server capital spending instead of cost reductions.
NAND Market Fundamentals Stay Robust
Rakesh highlighted that NAND memory content per AI server has increased 100% in the past twelve months. Spot market pricing has maintained an upward trajectory each quarter.
He contended that compression technologies like TurboQuant would “enable larger large-language models, faster inference and better tokenomics, spurring more spending” throughout the AI ecosystem.
Given the combination of price strength and solid underlying demand trends, Mizuho expects Micron and Sandisk to surpass current consensus earnings projections.
Sandisk currently trades near $652, compared to Mizuho’s $710 price objective. Micron sits around $347, against the firm’s $530 target price.
