NAND Flash’s Reversal of Fortune Amid the AI Boom

//php echo do_shortcode('[responsivevoice_button voice="US English Male" buttontext="Listen to Post"]') ?>

The memory industry, long known for its hyper-cyclical business, is once more at a crossroads. It all began with an AI boom a few years ago, which made high-bandwidth memory (HBM) a device of choice alongside AI accelerators for training models. HBM, a specialized form of DRAM, offered far better profit margins than NAND flash, whose price drops and shrinking margins made large memory suppliers like Samsung and SK Hynix cautious about expanding NAND manufacturing capacity.

Moreover, on the technology side, as NAND surpassed the 200-layer mark, each new generation required advanced manufacturing equipment and massive capital investment. The most advanced NAND devices currently have more than 330 layers. So, while large memory vendors began redirecting capital investments toward HBM production, NAND capacity was no longer a strategic focus by 2023.

Both Samsung and SK Hynix turned their focus toward DRAM-centric HBM devices, offering much higher profitability, rather than building new NAND production lines. According to Omdia, Samsung lowered its NAND wafer output from 4.9 million in 2024 to 4.68 million in 2025. Likewise, SK Hynix’s NAND wafer output fell from around 1.9 million in 2024 to 1.7 million in 2025.

Samsung and SK Hynix haven’t yet announced any NAND flash capacity expansions. And these two companies command more than 60% of the NAND flash market.

However, in retrospect, NAND flash was also crucial in the AI scheme of things. For instance, shifting the strategic focus from training to inference workloads led to steady demand for NAND flash. Consequently, demand for NAND flash is surging as solid-state disks (SSDs) are now widely used for inference workloads in AI data centers.

Take the case of Kioxia, currently in partnership with Nvidia; the Japanese NAND flash maker plans to launch SSDs nearly 100 times faster in 2027. These SSDs, tailored for AI servers, are intended to replace HBM as a memory pair to Nvidia’s GPUs. According to Nikkei, Kioxia anticipates that by 2029, nearly half of its NAND flash will be used in AI applications.

Where Kioxia and Sandisk stand

While Samsung and SK Hynix are stuck in the HBM conundrum, what about Kioxia and Sandisk, which are not into the DRAM business and solely produce flash memory chips? Well, that’s another story, but not entirely immune to NAND supply challenges.

Start with Kioxia, which spun off as Toshiba’s memory business in 2017. It’s now past its turbulent days, thanks to a robust demand for NAND flash in AI inference workloads. Since its inception, Kioxia has been struggling due to corporate hiccups and brutal commoditization of NAND flash devices. But then an unprecedented demand for SSDs built around NAND flash led to a reversal of fortune.

The managing director of Kioxia’s memory business unit, Shunsuke Nakato, startled the market by disclosing that his company has sold out its entire NAND flash production volume for 2026. Another surprise: The company is starting the production of BICS10 NAND memory ahead of schedule in 2026. BICS10—a 332-layer 3D NAND array aiming to facilitate high-capacity storage solutions for AI and hyperscalers—was originally planned for production in the second half of 2027.

Sandisk, which spun out Western Digital in 2025, is also riding high due to continued NAND shortages and sharp price hikes of high-capacity SSDs. Like Kioxia, a prolonged NAND shortage could encourage Sandisk to increase its NAND output eventually.

Reversal of fortune

Memory super cycles are a routine affair in the semiconductor industry, but this time around, AI is rewriting the rules of the memory playbook. Nvidia chief Jensen Huang echoed this sentiment at CES 2026 when he called storage a “completely unserved market.”

The NAND flash market experienced brutal commoditization during the early 2020s, and in 2024, prices fell to a record low of 50%. Then came the AI boom, and NAND makers began moving away from consumer markets for smartphones, tablets, and laptops to a new class of AI servers that offered much healthier margins.

However, by then, the NAND flash industry had already entered a conservative production mode, which eventually thrust NAND underdogs such as Kioxia and Sandisk to the fore while leaving memory heavyweights such as Samsung and SK Hynix in a bind.

AI, in the true spirit of disruption, has turned the memory industry upside down on both the DRAM and flash frontiers. But while DRAM-based HBM devices have remained in the limelight for the past couple of years, the surging demand for NAND-based SSDs in AI clusters is now showing ripple effects.

There is a broad consensus that the structural imbalance in storage—encompassing both DRAM and NAND flash—will likely persist through 2026 and possibly into 2027. How NAND flash makers respond to this scarcity in AI and consumer markets will be a story to watch in 2026.

See also: