"The tightest bottleneck in AI!" The impact of storage has expanded to the macro economy, exacerbating overall inflation.

The AI arms race is igniting a global storage chip crisis, the destructive power of which far exceeds the semiconductor industry itself, and is beginning to spread to the overall macroeconomy. On June 20, according to the Chasing Wind Trading Platform, the Deutsche Bank Aktiengesellschaft Research Institute pointed out in its latest report that storage chips have evolved from cyclical commodities to critical variables with macroeconomic significance, and the scale of this crisis has a clear quantified outline. In 2025, global storage market revenue increased by 35% year-on-year, setting a historical record of $223 billion; SK Hynix, Micron (MU.US), and Samsung's market values have all surpassed $1 trillion, collectively controlling over 90% of the global DRAM market share. Micron's CEO publicly stated that they can currently only meet 50% to two-thirds of the demand from some key customers, calling it the largest supply-demand gap they have ever seen. Deutsche Bank believes that this is not a traditional "boom and bust" cycle replay, but a profound structural supply shock triggered by AI. The bottomless demand of AI for high-bandwidth storage (HBM) is aggressively occupying the production capacity of traditional storage chips, leading to a "storage shortage crisis" that affects the global macroeconomy. Storage chips have evolved from simple commodities to critical macroeconomic variables that determine inflation and corporate profits. Hyperscalers and leading storage vendors (such as Micron) are the absolute winners in this crisis, possessing strong pricing power; while traditional consumer electronics industries like automotive, PC, and smartphones are facing severe profit compression and production allocation. More critically, the soaring cost of storage is transforming into a "chip inflation tax," directly pushing up overall inflation data in places like the United States. Demand Side: AI's Structural Devouring of Storage The demand for storage chips from the AI wave is fundamentally a structural rather than cyclical disruption. Storage chips play the role of "holding and feeding data" in AI systems - AI chips (such as NVIDIA Corporation's GPU) can only process data loaded onto them, and storage is responsible for this process, covering two core dimensions: capacity (how much data can be stored) and bandwidth (the speed of data movement). Without storage, chips cannot train AI models or perform inference tasks. Of particular note is the paradigm shift from "generative" to "agentic AI." Agentic AI can store and recall historical experiences, learn from interactions, maintain conversation context, and requires various types of storage such as DDR5, LPDDR, NAND to operate in synergy, significantly increasing overall storage consumption. Behind this lies an insurmountable "Memory Wall": once computing performance exceeds a certain threshold, if storage bandwidth is not synchronously expanded, the marginal returns of increasing computational power will tend towards zero - the progress of AI is determined by storage, not just computational power. Deutsche Bank's stock analysts predict that the demand for HBM will grow at an annual compound growth rate of around 40% by 2030, while the corresponding growth rate for standard DRAM is around 21%. Hyperscale cloud providers Meta, Amazon.com, Inc., and Microsoft Corporation are paying premiums to secure supply and signing multi-year agreements, further squeezing the market space for other buyers. Supply Side: Wafer fabs expansion unable to keep up with demand pace The core obstacle of the supply gap is time. It usually takes 2 to 3 years from the start of construction to the commissioning of storage wafer fabs, and most of the announced expansion projects will not make a substantial contribution to HBM capacity until at least 2027. The production characteristics of HBM further exacerbate the supply contradiction: the silicon consumption required to produce one unit of HBM is about three times that of ordinary DRAM. This means that each wafer directed towards HBM production is equivalent to squeezing out multiple wafers of standard DRAM/NAND capacity for end markets such as automotive, PC, etc. With the advent of HBM4/HBM4e technology evolution, the required silicon ratio will increase from three times to four times, and the "squeezing effect" will be further intensified. At the same time, the cleanroom space required for wafer processing has reached its limit, forcing manufacturers to make trade-offs on limited production lines. Qualcomm has explicitly stated that by 2026, the size of the smartphone market will be determined by DRAM supply, not consumer demand. DRAM currently accounts for about 70% of the overall storage market, exceeding the historical range of 50% to 60%. To expedite the landing of production capacity, the industry is exploring shortcuts such as "acquiring under-construction or second-hand wafer fabs." Micron acquired an old factory from Taiwan's Powerchip Semiconductor Manufacturing Corporation (PSMC) for $1.8 billion this year, saving about 2 years of time it would take to build a new factory from scratch. Deutsche Bank's latest estimate shows that global monthly wafer capacity for DRAM will increase by about 1.475 million wafers over the next five years, but the demand growth rate will continue to exceed the pace of supply expansion. Spillover Effects: From chip crisis to all-encompassing inflation Deutsche Bank states that the essence of the storage crisis is a zero-sum game: each wafer used for AI server HBM means less storage available for smartphones, PCs, or automobiles. Hyperscale cloud providers, with their ability to price AI computing power services, are able to pass on upstream costs to users, making them the most resilient group in this crisis; but a more widespread group of enterprises and consumers are suffering from allocation-based squeezing. The report emphasizes that price shocks have spread from the chip end to end products and macroeconomic price indexes: TrendForce predicts that in the second quarter of 2026, standard DRAM contract prices will increase by 58% to 63% quarter-on-quarter, while NAND flash contract prices will increase by 70% to 75% quarter-on-quarter. In terms of consumer electronics and PCs, Deutsche Bank estimates that total revenue from consumer terminal markets in 2026 will decrease by 15% year-on-year, but is forecast to recover to 9% year-on-year growth in 2027. Apple Inc. CEO has publicly warned of storage cost pressures; Apple Inc. quietly reduced the maximum memory configuration of some Mac Studio models, Microsoft Corporation lowered the entry-level memory of its new Surface business laptop from 16GB to 8GB, and Dell Technologies, Inc. Class C is also cutting back on product configurations - companies are generally opting to "downgrade specifications" rather than raising prices directly. It is worth noting that Lenovo, Dell Technologies, Inc. Class C, Asus have warned that they may implement a 15% to 20% price increase starting from July this year. In the automotive industry, the expected rise in DRAM costs is projected to increase the prices of regular vehicles by $150 to $300, and advanced autonomous vehicles by $400 to $600. Companies like Aptiv, Aumovio, Ford have issued signals of tight DRAM supply. Deutsche Bank analysts predict that inventories will be depleted by the end of 2026, and the substantial impact on car production volumes will begin to show from 2027. Automakers are faced with three options: absorb cost pressures to compress profits, pass on price hikes to consumers, or directly remove features such as L2+ level autonomous driving or in-car AI chat Siasun Robot&Automation to reduce DRAM-intensive functions. The Producer Price Index for Electronic Components and Accessories in the United States rose by 26.9% year-on-year in May 2026, far exceeding the 5.9% in January. Rising car prices may further drive consumers to extend loan terms, increasing overall interest expenses over the lifespan of the vehicle. Nine U.S. trade associations representing industries such as automotive, consumer electronics, medical devices, telecommunications, and retail have jointly written to Treasury Secretary Benson and Commerce Secretary Lutnick this month, issuing formal warnings about the potential impact of AI-driven storage disputes on the U.S. economy. Breaking the Deadlock and Potential Risks: Fab Construction, Algorithms, and the Hidden Peril of the "AI Bubble" The storage shortage is reshaping the geopolitical landscape of global technological competition. South Korea is heavily exposed to the global AI capital expenditure boom, with SK Hynix and Samsung accounting for 69% of global DRAM production. However, this double-edged sword also makes the South Korean economy extremely vulnerable: on June 8, when the semiconductor sector collapsed, the technology-heavy South Korean Kospi index plummeted by 8.29%, marking its ninth-largest single-day drop since 1980. Deutsche Bank believes that while giants are mitigating capacity constraints through massive capital expenditures and acquiring second-hand wafer fabs, they should be highly vigilant of the devastating overcapacity that could arise if AI demand slows down. To break the bottleneck, the three giants are aggressively increasing capital expenditures. In addition to physical expansion, the optimization of software algorithms has also caused market upheaval. In March of this year, Alphabet Inc. Class C released the "TurboQuant" algorithm, which reduces the memory requirements for large models during inference, leading to a sharp drop in stock prices of Samsung (-6%), Micron (-7%), and SK Hynix (-7%) on that day. Although the algorithm only targets the inference stage's KV cache and does not affect training memory requirements, and efficiency improvements may ultimately increase total demand due to the "Jevons Paradox," this signals that the tech industry is trying to find ways to reduce its excessive reliance on HBM. This article is reprinted from "Chasing Wind Trading Platform"; GMTEight Editor: Huang Xiaodong.