"AI's biggest bottleneck"! The impact of storage has extended to the macro economy, exacerbating overall inflation.

The AI arms race is igniting a global storage chip crisis, with destructive power far exceeding the semiconductor industry itself and beginning to spread to the overall macro economy.

On June 20, according to Wind Trading Desk, Deutsche Bank Research pointed out in its latest report that storage chips have evolved from cyclical bulk commodities to critical variables of macroeconomic significance, and the scale of this crisis now has a clear quantitative outline.

In 2025, global storage market total revenue will increase 35% year-on-year, setting a historical record of $223 billion; the market values of SK Hynix, Micron, and Samsung—the three major giants—have all surpassed $1 trillion, and together they control over 90% of the global DRAM market share.

Micron's CEO openly stated that they can currently only meet 50% to two-thirds of key customers' demands, calling this the largest supply-demand gap he has ever seen.

Deutsche Bank considers this is by no means a replay of the traditional "boom and bust" cycle, but a profound structural supply shock triggered by AI. AI's insatiable demand for high bandwidth memory (HBM) is fiercely squeezing traditional storage chip production capacity, resulting in a "storage shortage crisis" impacting the global macroeconomy.

Storage chips have evolved from simple commodities to key macroeconomic variables determining inflation and corporate profit. Hyperscale cloud providers and leading storage manufacturers (like Micron) are absolute winners in this crisis, possessing strong pricing power; traditional consumer electronics like cars, PCs, and smartphones are facing severe profit compression and rationed production capacity. Worse still, the surge in storage costs is turning into a "chip inflation tax," directly pushing up overall inflation numbers in the US and elsewhere.

Demand Side: AI's Structural Consumption of Memory

AI's surging demand for storage chips is essentially a structural, not cyclical, disruption.

Storage chips play the "holding and feeding data" role in AI systems—AI chips (such as Nvidia GPUs) can only process data already loaded onto them, storage is responsible for this process, covering both capacity (how much data can be stored) and bandwidth (how fast data moves). Without storage, chips can neither train AI models nor run inference tasks.

Particularly noteworthy is AI's shift from "generative" to "agentic AI" paradigms. Agentic AI can store and call historical experience, learn from interactions, and maintain conversational context, requiring synergistic operation across DDR5, LPDDR, NAND, and other types of storage, massively increasing overall memory consumption.

Behind this is the hard-to-surmount "memory wall": After computing performance exceeds a certain threshold, if memory bandwidth isn't simultaneously expanded, the marginal gains from increased compute will trend towards zero—AI's progress rate is decided by memory, not just by compute.

Deutsche Bank's equity analysts forecast that HBM demand will grow at a CAGR of about 40% to 2030, standard DRAM at about 21%. Hyperscale cloud providers Meta, Amazon, and Microsoft are paying premiums and signing multi-year agreements to lock in supply, further squeezing other buyers' market space.

Supply Side: Wafer Fab Expansion Can't Keep Up with Demand

The core obstacle to the supply gap is time.

It typically takes 2 to 3 years from breaking ground to production for a memory wafer fab, and most announced expansion projects won’t make substantive contributions to HBM capacity until at least 2027.

HBM’s manufacturing characteristics further amplify supply contradictions: Producing one unit of HBM consumes about 3 times the wafer volume of ordinary DRAM. Every wafer directed toward HBM production essentially crowds out several wafers that could be used for standard DRAM/NAND for automotive, PC, and other end markets.

As HBM4/HBM4e technology generations advance, the required silicon ratio will climb from 3 times to 4 times, further intensifying the squeeze effect. Meanwhile, wafer processing cleanroom space is approaching its limits, forcing manufacturers to make trade-offs on limited production lines.

Qualcomm has clearly stated that the scale of the smartphone market in 2026 will be determined by DRAM supply, not consumer demand. DRAM currently accounts for about 70% of the total memory market, higher than the historical range of 50% to 60%.

To accelerate capacity rollout, the industry is exploring shortcuts like "acquiring under-construction or used wafer fabs." This year, Micron acquired an old PSMC factory in Taiwan for $1.8 billion, saving about two years compared to building a new factory from scratch.

Deutsche Bank’s latest estimates show that global monthly DRAM wafer capacity will increase by about 1.475 million wafers over the next five years, but demand growth will still outpace supply expansion.

Spillover Effects: From Chip Crisis to General Inflation

Deutsche Bank says the essence of the memory crisis is a zero-sum game: Every wafer used for AI server HBM means less memory available for smartphones, PCs, or cars.

Hyperscale cloud providers, with their ability to price for AI compute services, can pass upstream costs onto users—they are the most resilient group in this crisis; but broader enterprises and consumers are suffering rationed squeeze-outs.

The report highlights that price shocks have already spread from chips to end products and macroeconomic price indicators:

TrendForce predicts that standard DRAM contract prices will rise 58% to 63% quarter-on-quarter in Q2 2026, NAND flash contract prices will rise 70% to 75% QoQ.

In consumer electronics and PCs, Deutsche Bank estimates that total annual revenue in the consumer terminal market will decline 15% year-on-year in 2026, with 2027 expected to recover to a 9% YoY increase.

Apple’s CEO has openly warned about memory cost pressure in earnings calls; Apple has quietly reduced the maximum memory configuration for some Mac Studio products, Microsoft lowered the entry-level memory for its new Surface business notebooks from 16GB to 8GB, and Dell is also trimming product specs—companies are generally choosing "downgrade specs" rather than direct price increases.

Notably, Lenovo, Dell, Asus have warned they may implement price hikes of 15% to 20% starting in July this year.

In the automotive sector, rising DRAM costs are expected to push up regular vehicle prices by $150 to $300, and advanced autonomous vehicles by $400 to $600.

Aptiv, Aumovio, Ford and other companies have signaled DRAM supply tightness. Deutsche Bank analysts forecast that inventory will run out throughout 2026, with significant impact on car production starting in 2027.

Car makers face three choices: absorb costs and compress profits, pass price increases onto consumers, or directly cut DRAM-intensive functions like L2+ auto driving and in-car AI chatbots.

The US electronic components and accessories Producer Price Index (PPI) rose 26.9% YoY in May 2026, far higher than 5.9% in January.

Increases in car prices may further push consumers to lengthen loan terms, raising total lifetime interest payments.

Nine US trade associations representing the car, consumer electronics, medical equipment, telecom, and retail industries have jointly written to Treasury Secretary Bessent and Commerce Secretary Lutnick this month, issuing a formal warning about the potential impact of AI-driven memory competition on the US economy.

Breaking the Deadlock and Potential Risks: Factory Construction, Algorithms, and AI Bubble Concerns

Memory shortages are reshaping the geopolitical landscape of global tech competition.

South Korea is heavily exposed to the global AI capex surge, with SK Hynix and Samsung accounting for 69% of global DRAM production. But this double-edged sword also renders the Korean economy extremely fragile:

On June 8, when the semiconductor sector crashed, Korea's Kospi index—dominated by tech stocks—plummeted 8.29%, marking the ninth largest single-day drop since 1980.

Deutsche Bank believes that although the giants are alleviating capacity bottlenecks through massive capex and secondhand wafer fab acquisitions, there is a high risk of catastrophic overcapacity if AI demand slows.

To break the bottleneck, the three giants are frantically increasing capex. Aside from physical expansion, software algorithm optimization is also shaking the market.

In March this year, Google released the "TurboQuant" algorithm, which reduces memory usage for inference of large models, causing Samsung (-6%), Micron (-7%), and SK Hynix (-7%) stock prices to plunge that day.

Although this algorithm only targets KV cache during inference and doesn't affect training memory demand, and gains in efficiency could ultimately increase total demand due to the "Jevons Paradox", it signals that the tech sector is trying everything to escape excessive dependence on HBM.

~~~~~~~~~~~~~~~~~~~~~~~~

All the above exciting content is from Wind Trading Desk.

For more detailed interpretations, including real-time insights and frontline research, please join [Wind Trading Desk▪Annual Membership]

Risk Warning and DisclaimerThe market has risks; investments need caution. This article does not constitute personal investment advice and does not take into account the special investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article fit their specific circumstances. Investments made accordingly are at one's own risk.