Citigroup Analysis: AI Begins to 'Consume' NAND, Storage Upcycle Still Ongoing
TL;DR
· Citigroup's report on July 24 indicates that low inventories of NAND and DRAM continue to support the storage upcycle.
· AI Agents are driving up demand for KV Cache, with CMX and near-GPU SSDs starting to consume more NAND.
· Samsung and SK Hynix are still benefiting from the storage boom, but weakening consumer demand, expansion, and HBM certification may affect the pace.
In its global semiconductor report released on July 24, Citigroup concluded that the storage upcycle is not yet over. The key evidence is not mobile phone demand, but rather that inventories of both NAND and DRAM are significantly below normal levels, and the demand for KV Cache and near-GPU storage from AI servers is still expanding.
This directly addresses recent market concerns about the peak of the storage cycle. The controversy mainly stems from two lines of thought: one is that weakening demand for smartphones in China may drag down consumer-grade NAND prices; the other is that the market is worried that NAND channel inventories are beginning to rise, suggesting that the increase in storage prices is nearing its end.
In the report titled "Global Semiconductors: Reality Check: Memory Upcycle Intact on Tight Inventories and Strong KV Cache Demand," Citigroup examines the inventories of suppliers, cloud providers, and channels together. The conclusion is that the current storage supply chain has not yet entered a typical loosening state.
NVIDIA CMX builds KV Cache storage layers through BlueField-4 and NVMe SSD, with AI Agents transmitting more inference demand to NAND.
NAND Supplier Inventory Only 2.6 Weeks, Normal Level About 5 Weeks
The most critical number is the NAND inventory.
Citigroup estimates that by Q3 2026, NAND supplier inventories will only be 2.6 weeks, significantly lower than the normal level of about 5 weeks; major cloud providers have NAND inventories of about 3 weeks, below the normal level of about 7 weeks; channel inventories are about 5 weeks, while the normal level is about 15 weeks.
This indicates that even if there are signs of weakness on the consumer side, the entire NAND supply chain is not loose. Low supplier inventories limit manufacturers' ability to quickly release spot goods; low cloud provider inventories suggest that AI and data center customers may continue to replenish; low channel inventories also reduce the likelihood of prices being suppressed by channel sell-offs.
NAND supplier inventory 2.6 weeks, cloud providers 3 weeks, channels 5 weeks, all below normal levels.
The situation with DRAM is similar. DRAM supplier inventories are about 2.7 weeks, below the normal level of about 5 weeks; cloud provider DRAM inventories are about 2.5 weeks, below the normal level of about 7 weeks; channel inventories are about 4 weeks, while the normal level is about 15 weeks.
This set of data weakens the argument that "the cycle can turn just because smartphone demand weakens." At least from the inventory perspective, both DRAM and NAND have not yet returned to an oversupply state.
DRAM supplier inventory 2.7 weeks, cloud providers 2.5 weeks, channels 4 weeks, all at low levels.
AI Servers Begin to Consume More NAND, Demand for Rubin System SSDs Rises
What distinguishes this storage upcycle from traditional consumer electronics cycles is that AI servers are starting to bring new demand for NAND.
Citigroup highlights two directions: CMX and QLC SSDs. CMX corresponds to what NVIDIA calls the Context Memory Storage Platform, which is a pod-level context and KV Cache storage layer driven by BlueField-4; QLC SSDs are more used for high-capacity storage close to GPUs to enhance AI computing efficiency.
KV Cache can be simply understood as a "context cache" in large model inference. The more frequently AI Agents are used, the longer the task chains, and the more complex the contexts, the more the system needs to store intermediate states and historical information to avoid redundant calculations. Inference not only consumes GPU computing power but also increases memory and storage demand.
A more intuitive hypothesis comes from NVIDIA's next-generation Vera Rubin platform. According to Citigroup's industry estimates, if the Rubin server system uses a 16TB TLC SSD for CMX operation, each system's SSD capacity could reach 1,152TB. NVIDIA's official statement is more conservative, indicating that each GPU pod provides PB-level shared capacity, without confirming that all Rubin systems are fixed to this configuration.
This number explains why AI demand will change the NAND market. In the past, NAND was more easily influenced by the rhythms of smartphones, PCs, and consumer electronics; now, AI servers are also becoming important buyers of high-capacity SSDs. As long as AI clusters continue to expand, storage demand will no longer be solely a function of the consumer side.
According to Citigroup's estimates, the NAND demand brought by CMX will reach 34.6 billion and 115.2 billion 8Gb equivalent units in 2026 and 2027, corresponding to 2.8% and 9.3% of global NAND demand. This is a model assumption, not company guidance, but a nearly 10% share by 2027 is already sufficient to influence supply-demand balance.
Weak smartphones do not mean the storage cycle will peak immediately.
The market's concern about weakening demand for smartphones in China is not without basis. Smartphones remain one of the important terminals for NAND, and if shipments fall below expectations, consumer-grade NAND prices will face pressure.
However, current inventories are too low, leaving little buffer for demand fluctuations. Citigroup mentioned that the supply-demand adequacy of DRAM and NAND suppliers has decreased from 70% to 50%. The surplus that suppliers have to cover demand is decreasing, not increasing.
In this context, weakening smartphone demand is more likely to slow down the pace of price increases rather than directly change the cycle direction. As long as cloud providers and AI servers continue to procure, low inventories will amplify the impact of new demand on prices.
It is also necessary to distinguish between two types of NAND demand: consumer-grade NAND is more affected by smartphones and PCs; enterprise-grade SSDs and AI-related SSDs are more influenced by data center construction, GPU cluster expansion, and model inference demand. The market has previously worried about the former, while Citigroup emphasizes that the latter is still strengthening.
This is also the reason why the report remains bullish on Korean storage stocks. Publicly cited information shows that Citigroup has raised Samsung Electronics' target price to 530,000 KRW and maintained a buy rating; SK Hynix also maintains a buy rating, with a previously reported target price of 3,100,000 KRW.
The individual stock target prices are not the most important part of this report. The real main line is that the market is worried about the peak of the storage cycle, but inventory data and AI storage demand provide contrary evidence.
HBM Certification, Consumer Demand, and Expansion May Disrupt the Pace
Citigroup's conclusion does not mean that storage prices will only rise and not fall.
For Samsung, the main risks include delays in the approval or certification of HBM shipments to key customers, weaker-than-expected PC sales, NAND demand falling short of expectations, aggressive investments by competitors in storage or foundry sectors, and significant appreciation of the won eroding profits.
For SK Hynix, risks are concentrated on declining DRAM demand, weaker-than-expected NAND demand, and weakening global consumption.
These risks indicate that AI demand has not completely eliminated the cyclicality of the storage industry. Storage is still an industry driven by capital expenditure, inventory, and terminal demand. If competitors rapidly expand production or consumer electronics demand further deteriorates, the supply-demand tension may be alleviated.
At least from the inventory numbers provided in this report, the market's concerns about the "peak of the cycle" still lack support from the inventory side. NAND suppliers have 2.6 weeks, DRAM suppliers have 2.7 weeks, and cloud providers' inventories are also below normal levels, indicating that the supply chain has not yet entered a typical oversupply state.
The current state is closer to: the consumer side has noise, the AI side is still placing orders, and the inventory side remains tight. As long as these three conditions do not show a significant reversal, it will be difficult to simply judge the storage cycle based on traditional smartphone cycles.
Disclaimer: This content is provided for general branding and informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online events, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets or to use any services. Crypto assets are highly volatile and may result in loss. WEEX services and online events may not be available in all regions and are subject to applicable laws, regulations, and eligibility requirements. You are responsible for ensuring that your use of WEEX services complies with local laws and for carefully assessing the risks before participating in any crypto-related activities.
You may also like

Wheat and soybean prices reach two-year highs due to the impact of the Middle East war

Can Recent Crypto Exchange Closures Signal the End of the Bear Market?

Hyperliquid, Multicoin back CFTC prediction market rules

Crypto is rewriting how Wall Street traders spend their weekends

Elon Musk predicts the end of mandatory work and money within 10 years

XRP Whales Buy While Small Wallets Sell

Will Money Become Obsolete by 2036? Bitcoin as the Ultimate Currency?

Artificial Intelligence Consumes Increasing Amounts of Electricity and Accelerates Investments in Data Centers

Which Cryptocurrencies Does Elon Musk Invest In?

Experts' Opinions on the Future Fed Rate Decision Are Divided

The leading prop trading firms: Multi-asset vs crypto-native compared

Zcash sets Ironwood upgrade for July 28 after Orchard bug

Fake Zoom, Real Thieves: North Korean Hackers from BlueNoroff Scan Your Crypto Wallets

Can You Trade Stocks With USDT? Why Crypto Exchanges Are Moving Into TradFi

Value Investing Guru Warns: AI Bubble May Burst, SpaceX is Nonsense, Bitcoin Will Go to Zero

Dollar: After Another Weekly Increase, the Market Processes Official Signals on the Exchange Rate

Institutions and On-chain Funds Optimistic About Changxin's Continued Surge, Except for Koreans

Who Benefits After AI Creates Trillions in Value? The U.S. Sparks Debate on AI Wealth Distribution Models

Coinbase bets on agentic finance as Base payments cross 100M

Oil Falls More Than 6% After Pause in Middle East Attacks as Markets Surge

First Research Report from Yili Hua's New AI Fund: The Wave of AI Computing Assetization, Axe Compute May Become the Most Undervalued GPU Computing Entry in the US Stock Market

A Collective Confession of the Three Longest-Lasting Crypto OGs – FACE Podcast

BNY unit wins MiCA entry as Europe’s crypto register hits 309

Reduction of VAT, support for IDPs and businesses: frontline communities present key proposals to the government program

Are crypto tokens overpriced when equity owns the real profits?

EU MiCA Transition Period Ends, Many Crypto Service Providers Face Exit Pressure? Understanding MiCA

U.S. Treasuries Pressure Walsh: Hawkish Stance Alone Is Not Enough, Market Wants Rate Hikes

Propinder Launches Free Prop Firm Comparison Tool For Trading Challenges

Nearly 90% of Stolen Funds Cannot Be Recovered: Web3 Attack Targets Shift from 'Code' to 'People' in the First Half of 2026






