[Explanation] Giant Cloud Providers Pivot from HDD to High-Capacity Flash, Following Meta
With the explosive adoption of AI servers, giant cloud providers, led by Meta, are beginning a major shift in recording media from traditional HDDs to high-capacity flash storage (QLC SSDs). The background to this is the stagnation in HDD performance improvements and the critical reality that storage has evolved into an infrastructure component where processing speed dictates the efficiency of AI training and inference.
Key points of the development and the move away from HDD
Meta's large-scale flash adoption and supply chain security
Meta, a global cloud giant, is dramatically shifting its storage strategy to support the exponential expansion of its AI computing capabilities. According to internal documents leaked in July 2026, the company has entered into multi-year, long-term supply agreements with SanDisk, Samsung Electronics, and Sumitomo Electric. Meta plans to double its AI computing power to 7 gigawatts in 2026 and 14 gigawatts in 2027, with annual capital expenditures expected to reach $145 billion, or approximately 23.4 trillion yen. To support this grand plan, Meta is focusing on securing flash memory storage from SanDisk and memory chips from Samsung Electronics. Additionally, its in-house AI chip, Iris, is scheduled for mass production starting in September 2026, as the company advances vertical integration of its hardware. The figure below illustrates the concept of the next-generation AI infrastructure that Meta is building.
Such massive investment is driven by the current structural memory shortage in AI infrastructure, reflecting the fear that if supply capacity is not secured in advance, the expansion plan itself could collapse.
HDD performance limits and new requirements for AI servers
HDDs, which have long been the primary storage medium in data centers, are facing major challenges in the AI era. Although HDD recording density continues to improve, input/output (I/O) performance has not kept pace, resulting in a continuous decline in bandwidth per terabyte (BW/TB). This phenomenon causes frequently accessed hot data to remain trapped within HDDs, creating a bottleneck for the entire system. On the other hand, in AI servers, SSDs have transformed from mere storage locations into critical infrastructure components that support AI processing speeds. In the AI training process, massive datasets must be continuously supplied to GPUs; if storage read speeds are slow, expensive GPUs are left waiting. Furthermore, requirements unique to AI workloads, such as writing checkpoints to save training progress and high-speed searches in vector databases for RAG (Retrieval-Augmented Generation), are exceeding the physical limits of HDDs. Data center engineers are being forced to migrate from HDDs to flash storage to bridge this performance gap.
Technological evolution and the construction of a new storage tier
Realizing a mid-tier storage layer opened up by QLC SSDs
Until now, data centers have primarily relied on cost-efficient HDDs and high-performance but expensive TLC (Triple-Level Cell) flash. However, Meta is proposing the introduction of QLC (Quad-Level Cell) SSDs as a third tier positioned between them. QLC is a technology that holds 4 bits of information per cell and has existed since around 2009, but its adoption in data centers has been limited due to low write endurance and capacity constraints. Recently, however, 2-terabit QLC dies and 32-die stacking technology have become mainstream, leading to a rapid increase in density. Meta is targeting workloads that require performance of around 10MB/s/TB, where HDDs lack the necessary performance and TLC is not cost-effective. Servers adopting QLC can achieve approximately six times the byte density compared to the highest-density TLC servers Meta currently uses. In terms of power efficiency, QLC is also superior to HDDs, and it is expected to provide significant cost benefits, especially for read-intensive AI inference logs and data archiving.
Efficiency through the fusion of software and hardware
To maximize the performance of flash storage, approaches different from conventional SSDs are also being adopted. The DirectFlash technology that Meta is introducing in collaboration with Pure Storage is one such example. Typical SSDs have an internal translation layer to make flash memory appear like an HDD, but this has been a cause of performance inefficiency and reduced reliability. With DirectFlash, this translation process is not performed inside the SSD but is directly controlled by advanced software on the host side. This can improve NAND flash utilization efficiency by approximately 30 percent, making it possible to secure more capacity with the same number of chips. In concrete terms, modules (DFMs) adopting DirectFlash can achieve a 5 to 10-fold increase in density compared to standard SSDs, along with corresponding reductions in space and power. Meta is integrating this technology into its own software stack to efficiently manage storage on the scale of several exabytes. Please refer to the figure below.
Market transformation and future outlook
Structural supply shortages and accelerating chip inflation
The explosive increase in AI demand is bringing about serious structural changes in the NAND flash market. According to 2026 market forecasts, while demand for high-capacity enterprise SSDs for AI servers is surging, expansion of production capacity on the supplier side is limited, and the market is expected to fall into a severe supply shortage. Although demand for consumer electronics, which has supported the market until now, is stagnant, robust demand for data centers is offsetting this, and NAND flash contract prices are expected to maintain an upward trend. This situation is causing chip inflation, where soaring memory prices affect the macroeconomy, and for buyers like Meta, the massive capital expenditure costs are becoming a factor that pressures their finances. Since suppliers prioritize production resources for high-margin AI products, it has been pointed out that this could lead to price increases for consumer products such as PCs and smartphones, or a decrease in shipments of low-to-mid-range models.
Future developments and challenges facing AI infrastructure
The point of interest going forward is how this storage revolution will change the balance of the entire AI infrastructure. The essence of AI is shifting not just to computing power, but to the comprehensive ability to store, move, and efficiently supply massive amounts of data. With the introduction of its in-house Iris chip, Meta is rushing to optimize its storage hierarchy while reducing its dependence on specific semiconductor manufacturers. On the other hand, the widespread adoption of QLC SSDs still faces technical challenges, such as how to manage the performance gap between reads and writes through software. Additionally, physical constraints such as power supply and cooling capacity are becoming major barriers to data center expansion. The arms race in AI infrastructure is entering a new phase that goes beyond simple chip performance competition to competing for supply chain security and overall system energy efficiency. Investors and engineers will need to continue to pay close attention not only to the evolution of GPUs but also to the structural changes in the memory and storage that support data behind the scenes. The graph below shows the forecast for the future supply-demand gap.
Reference Pages
[A case for QLC SSDs in the data center - Engineering at Meta] https://engineering.fb.com/2025/03/04/data-center-engineering/a-case-for-qlc-ssds-in-the-data-center/
[2026 NAND Flash: AI & HDD Shortage Ignite Price Surge - TrendForce] https://www.trendforce.com/research/category/Semiconductors/NAND%20Flash
[Pure Storage at William Blair Conference: Strategic Growth Insights - Investing.com] https://www.investing.com/news/transcripts/pure-storage-at-william-blair-conference-strategic-growth-insights-93CH-4078574
[#AI_Server #SSD #NAND_Flash #Data_Center #Meta #Semiconductors #QLC #Tech_Trends]
Click here for AI slide creation tools (try for free)
Click here for task management and project management tools (30-day free trial)
