Nvidia’s $279 Billion Memory Demand: Micron or Sandisk Poised to Win?

Key Points

• Nvidia is currently budgeting to spend $279 billion on memory through 2032.

• Micron and Sandisk are top suppliers of high bandwidth memory, DRAM, NAND, and enterprise solid-state drives.

• Micron is better positioned to capture infrastructure spend from the hyperscalers given its place in AI chip stacks.

The artificial intelligence (AI) memory market is currently experiencing a highly profitable yet constrained environment. As training models grow larger and inference moves into full production, every new rack of accelerators demands greater bandwidth and storage capacity closer to the GPU. This dynamic is driving a massive memory supercycle.

This shortage stems from fundamental supply limitations. Expanding wafer capacity and converting manufacturing lines to produce high-bandwidth memory (HBM) is a complex and expensive process. While NAND supply is also constrained, the driver differs: hyperscale storage demands are surging due to the massive volume of tokens generated during inference that require persistent storage.

Micron Technology (NASDAQ: MU) is capitalizing on both aspects of this demand. The company manufactures the HBM integrated directly onto GPU packages, alongside the server DRAM that supports the broader system. Sandisk (NASDAQ: SNDK) is also benefiting, though primarily from the storage side, supplying enterprise flash drives to the same data centers.

While both companies are seeing unprecedented profitability, Nvidia’s recent commentary on the memory market raises an important question: which of these two stocks is more closely tied to the true memory bottleneck?

Image source: Getty Images.

What did Nvidia say about the memory trade?

During its fiscal second-quarter earnings call, Nvidia emphasized that memory has transitioned from a background expense to a strategic constraint. CFO Colette Kress noted that the company is experiencing “extreme pricing conditions in memory,” with price increases already surpassing internal forecasts and expected to climb further next year.

Consequently, Kress guided Nvidia’s gross margin to decline and bottom out in the low-70 percentage range before pricing adjustments take effect. The projected trough is approximately 71% in the fourth quarter of fiscal 2027.

Nvidia also disclosed a detailed breakdown of its supply and capacity commitments through fiscal 2032, totaling $279 billion. Management stated this expenditure is primarily for memory procurement. Approximately $92 billion of this budget is allocated to the remainder of fiscal 2027, followed by $87 billion and $88 billion for the subsequent two years. This signals that Nvidia is reserving a significant portion of the memory market through fiscal 2029, as the company does not trust current market dynamics to deliver Blackwell and Vera Rubin systems on schedule.

CEO Jensen Huang further clarified that unconstrained demand suggests growth well beyond the 70% Nvidia is willing to officially guide. Supply—specifically memory components—is the governing factor. Nvidia did not merely place a casual purchase order; it placed a multi-year claim on one of the scarcest assets in the semiconductor value chain: memory bits.

What do hyperscalers need more of?

Hyperscalers do not purchase “memory” as a single entity; rather, they acquire different components to perform distinct functions.

HBM represents the scarcest tier of the memory stack, sitting directly atop GPU packages and dictating the speed at which models can train. Without HBM, accelerators are essentially expensive chip clusters. DRAM acts as the workhorse surrounding this package, feeding system memory for central processing units (CPUs) and inference hosts. Meanwhile, NAND supports the storage infrastructure, housing training data, checkpoints, and logs. While all three are vital, they are not equally impacted by Nvidia’s warning.

Nvidia is securing supply because GPU shipments rely on available stacked memory and DRAM, not because it cannot find solid-state drives (SSDs). Micron stands as one of only three HBM suppliers and a top-tier DRAM vendor capable of providing this capacity.

While Sandisk is a prominent NAND manufacturer and AI factories will continue to purchase flash, storage sits further downstream in data center architectures. If a GPU cannot ship, orders for SSDs can be delayed. However, if HBM is unavailable on time, the entire rack shipment is delayed. This is why Nvidia’s $279 billion memory budget resonates far more strongly for Micron than for Sandisk.

Why Micron looks like the better buy over Sandisk

Despite its emerging critical role in hyperscaler data centers, Micron is not yet priced like a compounding growth stock. Although shares have experienced a significant run over the past year, Micron’s forward price-to-earnings (P/E) ratio hovers around 6. This valuation is modest compared to other category-leading chip stocks like Nvidia, Broadcom, or Taiwan Semiconductor Manufacturing. The disparity in valuation multiples stems from the market treating memory as a boom-bust commodity.

MU PE Ratio (Forward) data by YCharts.

Micron is positioned to benefit from several secular tailwinds throughout the AI infrastructure era. Hyperscaler capital expenditure increasingly treats memory as a first-order constraint, inference deployments are scaling and consuming more bits per token, and contractual demand provides multi-year visibility.

Ultimately, this dynamic supports an environment where earnings remain elevated long enough for the market to expand its valuation multiple and cease applying a deep cyclical discount.

While Sandisk will continue to benefit from higher flash prices and an increasing mix of enterprise data center revenue, NAND remains a secondary winner. Nvidia did not write a quarter-trillion-dollar check because SSDs are difficult to procure; it did so because memory scarcity is the bottleneck determining how many AI systems will actually be built on time. This is Micron’s opportunity, making it the superior stock to buy and hold for the memory boom.

Source link