ANALYSIS: DRAM Memory, the New Oil of AI — Why the Global Shortage Will Last

ANALYSIS: DRAM Memory, the New Oil of AI — Why the Global Shortage Will Last

Introduction: When Apple Raises Its Prices Because of AI

The June 26, 2026 Shock — Apple Yields to the Chip Shortage

On June 26, 2026, Apple announced price increases on its Macs and iPads, explicitly citing a memory chip shortage caused by the artificial intelligence boom. The company from Cupertino — known for absorbing component shocks without immediately passing them on to consumers — described the situation as an "unprecedented challenge" for the consumer electronics industry. CEO Tim Cook had anticipated the bad news during a Wall Street Journal interview as early as June 17, calling price increases "inevitable." The signal was clear: something structural is fracturing in global technology supply chains.

What is fracturing is the balance between supply and demand for DRAM and NAND Flash memory — the two types of memory chips that equip everything from smartphones to AI servers. Demand from AI-powered data centers has exploded at a speed that manufacturers — Micron, Samsung, and SK Hynix — cannot match. And ordinary consumers are now footing the bill.

DRAM and NAND — The Invisible Bricks of Our Digital World

DRAM (Dynamic Random-Access Memory) is the working memory that allows processors to store and rapidly access data being processed. Without sufficient DRAM, AI systems cannot run the large language models (LLMs) that require hundreds of gigabytes of active memory. NAND Flash is storage memory — it equips SSDs, memory cards, and server storage systems. Both technologies are produced by a narrow oligopoly: Samsung (South Korea), SK Hynix (South Korea), and Micron (United States) together control more than 95% of the global DRAM market.

The Explosion in AI Demand — Numbers That Boggle the Mind

An AI Server Consumes 8 Times More Memory Than a Standard Server

To understand why the DRAM shortage is structural, one must grasp the scale of memory consumption by AI infrastructure. A typical standard data center server contains approximately 256 to 512 gigabytes of DRAM. A specialized AI server — such as those running large-scale language models — can require up to 4 to 8 terabytes of high-bandwidth memory (HBM). Multiply that by the hundreds of thousands of servers that Microsoft, Google, Amazon, and Meta deploy each year for their AI infrastructure, and you get demand that exceeds global production capacity.

ChatGPT alone — the OpenAI service hosted by Microsoft — requires the equivalent of several hundred thousand NVIDIA GPU chips, each accompanied by massive quantities of DRAM. In 2026, industry estimates suggest that AI data centers consume approximately 40% of global DRAM production — a share that was below 10% in 2020. This acceleration is unprecedented in the history of the semiconductor industry.

HBM — The Premium Memory Everyone Wants

At the heart of the shortage lies a specific type of DRAM: HBM (High Bandwidth Memory), a high-performance memory designed to be stacked directly on the GPU and NPU chips used in AI accelerators. HBM3e — the current generation — delivers bandwidth of several terabytes per second, essential for powering the massive matrix computations of AI models. SK Hynix is currently the principal supplier of HBM to NVIDIA, the dominant GPU manufacturer in the AI ecosystem. Samsung is trying to catch up, but still struggles to qualify its HBM4 chips with major customers.

Global HBM production capacity is physically constrained by an extremely complex manufacturing process — stacking dozens of layers of DRAM with through-silicon connections. SK Hynix and Samsung cannot simply "turn on a tap" of additional production — building a new HBM production line takes 3 to 5 years and costs billions of dollars. The current shortage is therefore structurally irreversible in the short term.

The Three Manufacturers — Why They Cannot Keep Up

Micron — The American Player Fighting on Multiple Fronts

Micron Technology, based in Boise, Idaho, is the only American DRAM manufacturer. After years of underinvestment and financial turbulence linked to the semiconductor industry's classic boom-and-bust cycles, Micron announced a massive investment plan — tens of billions of dollars over five years — in new production capacity in the United States and India, partially supported by the American CHIPS Act of 2022. But these investments will only translate into production capacity by 2027-2030.

In the meantime, Micron is forced to allocate its production primarily to the most lucrative clients — the major hyperscalers (Microsoft, Amazon AWS, Google Cloud) who can pay premiums for firm contracts. Apple, though one of Micron's largest customers, finds itself in direct competition with clients whose orders are larger and who hold more favourable long-term contracts. The result is visible on Apple's price lists in June 2026.

Samsung and SK Hynix — The Koreans Under Pressure

Samsung Electronics and SK Hynix dominate the global DRAM market with approximately 44% and 35% market share respectively. Both companies have their production centres primarily in South Korea — a geography that raises geopolitical risk questions given the proximity of North Korea and tensions in the Taiwan Strait. Both groups are investing heavily in expanding production capacity, but timelines are incompressible. Samsung struggles to qualify its new HBM4 chips with AI customers — a technical failure that cost it significant market share in 2025-2026 to rival SK Hynix.

SK Hynix is in an exceptionally strong — perhaps dangerously strong — position. Its dominance in HBM makes it the indispensable supplier to NVIDIA for its flagship chips. This near-monopoly position on a critical component gives it considerable pricing power — and proportional responsibility in managing the global shortage. The company has announced an acceleration of its investments in HBM production lines but acknowledges it cannot satisfy current demand.

The Impact on Consumers — Who Really Pays the Bill?

Apple — The First Visible Signal for the General Public

Apple's announcement on June 26, 2026 is perhaps the first visible sign for the general public of this structural crisis. The company stated that "the rapid expansion of AI data centers has created extraordinary demand for memory and storage. We have never seen the prices of a component increase this much, this fast." This phrasing — unusual in its frankness — reflects a situation in which even Apple's colossal purchasing power (500 million iPhones and iPads sold per year) no longer allows it to absorb component cost increases.

The announced price increases affect the Mac and iPad lines, which contain more memory than iPhones and are therefore more exposed. High-end models — MacBook Pro with the M4 Max chip, Mac Studio, Mac Pro — which require large quantities of unified memory for professional performance, will be most affected. Creative professionals, video editors, and software developers — privileged customers of these machines — will have to absorb these increases or delay their purchases.

Beyond Apple — The Impact on the Entire Industry

But the memory crisis extends far beyond Apple. HP, Dell, Lenovo, and all PC manufacturers face the same pressures. Automakers integrating embedded AI systems in their vehicles are seeing costs rise. Industrial, medical, and telecommunications equipment manufacturers — all dependent on DRAM for their embedded systems — are absorbing the same shock. The memory shortage is a systemic event that cuts across all industries that have digitized over the past decade.

Memory chip markets are cyclical by nature — periods of shortage and overcapacity have historically alternated. But AI demand has introduced a structurally higher floor under demand, which could extend the shortage period well beyond the usual 18-to-24-month cycles. Some analysts estimate that equilibrium will not be restored before 2028-2029.

AI Data Centers — Insatiable Memory Devourers

Microsoft, Google, Amazon — The GPU Race Accelerates

The three major hyperscalers — Microsoft, Google, and Amazon — have announced data center investment plans of unprecedented scale for 2026-2028. Microsoft alone plans to spend more than $80 billion in fiscal year 2025-2026 on data center infrastructure, primarily to support its AI services. Google and Amazon are on comparable trajectories. Meta announced an investment of $60 to $65 billion in 2025 for its AI infrastructure.

These figures translate directly into massive orders for NVIDIA GPU chips, each accompanied by significant quantities of HBM and standard DRAM. NVIDIA's flagship chip, the H100, integrates 80 gigabytes of HBM2e. Its successor, the H200, scales to 141 gigabytes of HBM3e. The next-generation B200 is expected to exceed 192 gigabytes of HBM3e. Each generation of AI GPUs consumes more memory than the last — a trend that nothing on the visible horizon appears capable of reversing.

Energy and Memory — AI's Two Bottlenecks

Memory is not the only bottleneck for AI data center expansion. Energy is the other major constraint. A large AI data center consumes as much electricity as a mid-sized city. Operators struggle to obtain the necessary electrical connections from grid managers whose infrastructure schedules are calculated over decades — not over the few-quarter expansion cycles of the technology industry. In several regions of the United States and Europe, data center projects are stalled for lack of available electrical capacity.

This dual constraint — memory and energy — creates additional pressure on semiconductor manufacturers to develop more energy-efficient architectures. NVIDIA and its competitors are working on architectures that improve the performance-per-watt and performance-per-memory ratios. But these innovations take time — and in the meantime, the DRAM shortage translates into rising prices for end consumers.

Supply Chains — A Global Redrawing Under Way

The American CHIPS Act — The Semiconductor Reindustrialization

Faced with the geographic concentration of semiconductor production — primarily in East Asia (Taiwan, South Korea, Japan) — the United States and the European Union have launched massive reindustrialization programmes. The American CHIPS and Science Act of 2022 allocates approximately $52 billion to semiconductor production on American soil. Intel, TSMC, Samsung, and Micron have announced American factories partially supported by these funds. The European Union launched its own European Chips Act, with a goal of doubling Europe's share of global semiconductor production by 2030.

But these investments take time. A next-generation fab (semiconductor manufacturing plant) takes 3 to 5 years from initial decision to commercial production. New American and European capacities will not be operational until 2027-2030. In the meantime, geographic dependence persists, and the 2026 DRAM shortage is the visible symptom of this structural vulnerability.

India and Southeast Asia — The New Players

Facing geopolitical risks linked to concentration in South Korea and Taiwan, some manufacturers are exploring alternative production sites. Micron announced a significant investment in India, where the Modi government offers generous incentives to attract the semiconductor industry. Japan has attracted TSMC, Samsung, and Sony to build new production capacities as part of its policy to secure technology supply chains.

These geographic diversifications are positive for long-term resilience. But they will not resolve the current DRAM shortage — the new factories will not be operational for several years. And production of HBM, the critical memory for AI, remains concentrated in South Korea for reasons of technological expertise and industrial know-how that are not easily relocated.

HBM Memory — Technological Race and Geopolitical Stakes

HBM4 — The Next Technological Frontier

The next generation of high-performance memory for AI is HBM4, whose commercial deployment is expected for 2026-2027. It promises even higher bandwidth and reduced energy consumption per gigabyte. SK Hynix and Samsung are locked in a technological race to be the first to deliver commercial volumes of HBM4 to NVIDIA and other AI chip manufacturers. Samsung's lag on HBM3e cost the company several quarters of lucrative contracts — a lesson its engineers have clearly absorbed.

But HBM4 is not merely a technical improvement — it is another leap in manufacturing complexity. Stacking ever-thinner layers of DRAM connected by thousands of connections traversing the silicon (TSV, Through-Silicon Vias) is one of the most complex manufacturing challenges in the industry. Even SK Hynix, the current leader, acknowledges that ramping up HBM4 production will take time. The shortage will likely intensify before it begins to ease.

China in the Memory Race — An Absent but Threatening Player

China has for years sought to develop an independent semiconductor industry but remains significantly behind global players in advanced memory. YMTC (Yangtze Memory Technologies Co.) has made remarkable progress in NAND Flash production, enough to be placed on the restricted entities list by the American Commerce Department in 2022. For advanced DRAM — and even more so for HBM — China remains far from global standards.

But the ambition is there, and the financial resources deployed by the Chinese state to close this gap are considerable. If China manages to develop indigenous HBM production capability within five to ten years, global market dynamics will be profoundly transformed. For now, this threat is more potential than real — but decision-makers in Seoul, Boise, and Brussels must keep it in mind when formulating their investment strategies.

The Impact on European Companies and Digital Sovereignty

Europe — A Spectator of the Shortage?

Europe is almost entirely absent from advanced DRAM production. The European Chips Act, launched in 2023, aims to double Europe's share of global semiconductor production — from 9% to 20% by 2030. But this objective concerns primarily advanced logic chips (TSMC announced a factory in Germany, Intel in Germany and Poland). DRAM and HBM production in Europe remains virtually non-existent.

European companies — whether in fintech, digital health, connected automotive, or telecom — all depend on the same Korean and American suppliers for their memory components. The DRAM shortage is therefore a European digital sovereignty problem as much as an economic one. But here too, deploying solutions will take time that market urgency does not allow.

The Bill for European AI Startups

For European AI startups — which attempt to build models and services competitive with American giants — the memory shortage is an additional burden. GPU and memory costs for training and inferencing AI models represent a growing share of development budgets. American hyperscalers, which have contracted priority component deliveries, hold a competitive advantage further reinforced by the shortage. The gap between tech giants and emerging players is widening precisely because of these supply constraints.

Initiatives like Mistral AI (France), Aleph Alpha (Germany), and Stability AI (UK) attempt to develop open and sovereign AI models. These efforts deserve political and economic support. But if access conditions to basic components remain as unfavourable for non-American players, the competition will be structurally asymmetric.

Memory Market Cycles — Forecasts for 2026–2030

A Return to Equilibrium Before 2028 Is Unlikely

Semiconductor market analysts agree on one point: the return to equilibrium between DRAM supply and demand will not occur before 2028 at the earliest. The new production capacities announced by Micron, Samsung, and SK Hynix take 3 to 5 years to deploy. Meanwhile, demand from AI data centers will continue to grow. The business models of major AI players — which monetize AI services at high margins — allow them to absorb component price increases that ordinary consumers cannot afford.

The medium-term risk is a market bifurcation: a premium market for high-performance AI components, at high prices and reserved for major players, and a standard market for consumer products, forced to absorb cost increases or to reduce product specifications. Apple chose the first option — passing price increases on to its customers. Other manufacturers in more price-sensitive segments may choose to reduce memory quantities in their products — thereby lowering performance.

Technology Trends That Could Change the Game

Several technology trends could mitigate the shortage over the longer term. Processing-in-Memory (PiM) — which integrates computing capabilities directly into memory chips — could reduce data transfers between processor and memory, easing bandwidth demand. More memory-efficient AI architectures — such as Mixture of Experts (MoE) or model compression techniques — could reduce per-model memory requirements. Photonics — using light rather than electricity for inter-chip interconnections — could revolutionize memory architectures by 2030.

But these innovations are 3 to 7 years away. In the near term, the shortage is real, prices are rising, and consumers who want a MacBook will pay the price. The AI revolution has a material cost that enthusiastic technology discourse has long ignored.

Geopolitical Implications — Memory and National Security

DRAM as a Strategic Resource

The 2026 DRAM shortage has accelerated awareness in Western capitals that memory components are a strategic resource — on a par with oil, rare earths, or critical medicines. The American Department of Defense and European defence ministries themselves use massive quantities of DRAM in their communications, command-and-control systems, guided weapons systems, and surveillance infrastructure. A DRAM shortage is also a national defence risk.

The Pentagon has for several years had dedicated programmes to secure semiconductor supply for defence applications. The American CHIPS Act includes specific provisions for military-use semiconductors. But defence needs represent a fraction of the overall market — and the overall market is currently being absorbed by civilian AI demand. The military finds itself competing with hyperscaler data centers for the same components.

Upcoming Trade Tensions

The DRAM shortage risks generating new trade tensions. The United States, through the CHIPS Act, directly subsidizes semiconductor production on its soil and imposes export restrictions on advanced technologies toward China. These industrial-protectionist policies, while understandable from an American national security perspective, create market distortions and friction with European and Asian allies.

If the shortage intensifies, pressure may build for American and Korean manufacturers to allocate their production preferentially to domestic or allied markets. This kind of directed allocation of critical components — analogous to OPEC's oil export quotas — would set a major precedent in the semiconductor industry and transform global technology competition into an open trade conflict.

The Consumer Response — Adaptation or Waiting?

Delaying Purchases or Accepting the Increases

Faced with the price increases announced by Apple and other manufacturers, consumers have two main options: accept the increases and proceed with planned purchases, or delay their decisions in the hope of price normalization. Historical data on consumer behaviour during electronics component shortages — notably the semiconductor shortage of 2020-2022 — shows that purchases tend to be deferred, particularly for long-lifecycle products like professional computers.

Professionals who depend on high-performance machines — creatives, developers, scientists, researchers — cannot always defer purchases. They will absorb the price increases in their budgets, often by charging more to their end clients. A video editor whose MacBook Pro costs €500 more will pass that cost on in invoices. This is the inflationary logic of component shortages — it diffuses throughout the entire economy, often invisibly.

The Secondary Market and Alternatives

The DRAM shortage and new equipment price increases have logically boosted the secondary market for refurbished equipment. Well-maintained used Macs and PCs have seen their resale value increase significantly since 2025. For users whose needs do not require the latest performance generations, buying refurbished is an economically rational alternative.

Alternative manufacturers — notably Chinese brands like Huawei and Xiaomi for personal computers — could benefit from Apple's price increases if they manage to maintain more competitive prices. But these manufacturers face the same memory shortages — their margins are simply different and their exposure to American export restrictions creates additional uncertainties in their supply chains.

Massive Investment in New Capacity — The Race Against the Clock

The Expansion Announcements of the Three Giants

Facing the explosion in demand, Samsung, SK Hynix, and Micron have announced unprecedented investment programmes in their production capacities. SK Hynix announced the opening of a new mega-factory at Cheongju, South Korea, dedicated to HBM4 production from 2027. Samsung is massively investing in modernizing its production lines at Pyeongtaek. Micron, backed by American CHIPS Act funds, is building a new plant in Boise, Idaho, that will begin producing advanced DRAM in 2028.

These investments run into the tens of billions of dollars each — a scale that illustrates the colossal barriers to entry in this industry. Building a state-of-the-art DRAM factory takes three to five years from initial decision to commercial production. That is precisely why the current shortage cannot be resolved quickly: the investment decisions that could have anticipated the AI boom would have had to be made in 2021-2022, when no one yet grasped the full scale of the ChatGPT revolution.

New Entrants and the Impossibility of Short-Term Diversification

Several countries and companies have attempted to enter the advanced memory market to break the Korean-American oligopoly. CXMT (ChangXin Memory Technologies) in China has made progress in standard DRAM production but remains several generations behind on HBM. Taiwan Semiconductor Manufacturing Company (TSMC), specialized in advanced logic, does not produce DRAM — its expertise and production lines are fundamentally different. Europe has no significant player in advanced memory production, despite the ambitions of the European Chips Act.

This reality means that diversification of advanced memory supply sources is structurally impossible in the short term. Governments concerned about the geographic concentration of this industry must accept that even with massive investments starting today, it will take at least a decade for a new player to reach industrial maturity in HBM production. In the interim, dependence on Samsung, SK Hynix, and Micron is an inescapable geopolitical fact to be managed rather than denied.

The Secondary Market and Consumer Adaptation Strategies

The Resurgence of Refurbished and Rental

Faced with rising new-device prices, consumers and companies are developing adaptation strategies. The refurbished device market — refurbished iPhones, MacBooks, enterprise servers — has been growing rapidly since the first shortage announcements. Platforms like Back Market, Swappie, and Decluttr saw their transaction volumes increase significantly in the first half of 2026. Consumers who previously renewed devices every two to three years are extending their replacement cycles to four or five years.

For companies — notably SMEs and startups that lack hyperscaler resources — the response to the shortage increasingly involves cloud services. Rather than buying ever-more-expensive servers, they rent computing power from AWS, Azure, or Google Cloud — which displaces the shortage problem onto the hyperscalers but externalizes it financially. This shift toward cloud computing as a response to hardware shortages paradoxically reinforces the dominant position of major technology players.

Implications for SME and Startup Innovation

One of the least-discussed consequences of the memory shortage is its impact on innovation at the base of the technology pyramid. AI startups that needed affordable servers to develop and train their models face hardware and cloud costs that have risen 20 to 40% depending on configuration. For a seed-stage startup, that is a significant difference in cash runway. Some promising AI projects, carried by teams of fewer than ten people, may not survive this cost increase.

This pressure on small players mechanically favours market concentration in AI around large companies that can absorb cost increases. It is the inverse of the democratization promise carried by open-source AI and accessible models. The AI revolution risks becoming a privilege of large companies and wealthy countries — not the universal tool it promised to be. This concentration risk deserves urgent political attention.

Education and Training — Preparing Tomorrow's Professionals

A Skills Deficit in the Semiconductor Industry

The DRAM shortage is not only a production capacity crisis — it is also a human skills crisis. Training an engineer specialized in designing or manufacturing advanced memory takes five to ten years after a basic university degree in electrical engineering or materials physics. The sudden explosion in AI demand has created a demand for qualified talent that vastly exceeds available supply in producing countries. Samsung and SK Hynix are fighting to recruit the same engineers from Korean, Taiwanese, and American universities, with sharply rising salary offers.

This talent pressure has cascading effects. Universities that train semiconductor engineers see their programmes overloaded and their faculty approached by industry for career transfers. Countries that invest today in specialized training programmes — notably the United States with the education component of the CHIPS Act and South Korea with its university-industry partnerships — are building a durable competitive advantage. Europe, which lacks a strong tradition in this industry, is falling behind on this crucial axis.

University-Industry Partnerships as a Strategic Lever

Innovative training models are emerging to accelerate talent production. SK Hynix launched an industrial doctoral programme with several leading Korean universities, allowing students to conduct their research directly in the group's factories, on real HBM production problems. Micron has established university chairs at several major American universities to develop specific competencies in advanced memory. These partnerships reduce the adaptation time between academic training and real industrial needs.

For Europe, the question is whether the European Chips Act will include sufficient investments in specialized training — not just in factory construction. A semiconductor factory without engineers to operate it is a sterile investment. European plans include training components, but their ambition falls short of the projected needs to develop a competitive industry by 2030.

Conclusion: Memory Is the Oil of the AI Era — and We Have No OPEC

A Structural Shortage With No Quick Resolution in Sight

The DRAM and HBM memory crisis is not a temporary market accident — it is the symptom of a structural transformation of the global technology economy. AI demand has created a memory consumption floor that nothing can saturate in the short term. Manufacturers cannot accelerate production capacity at the pace of this demand. And end consumers — from iPad users to industrial companies — bear the consequences of a resource allocation they do not control.

Apple's announcement on June 26, 2026 may be the starting gun for a lasting restructuring of consumer electronics prices. If the shortage lasts until 2028-2029 as some analysts project, years of price increases lie ahead. Consumers who thought the digital era meant continuous performance improvements at constant prices will need to revise their expectations.

What Governments Must Do — and Quickly

Faced with this reality, governments — American, European, Japanese, Korean — must accelerate their investments in semiconductor supply chain diversification, support R&D on alternative memory architectures, and build strategic reserves of critical components analogous to strategic oil reserves. These policies exist in embryo — the CHIPS Act, the European Chips Act, Japan's reindustrialization programmes. They must be accelerated and amplified. Memory is the new oil. We cannot wait for the next crisis to take that seriously.

By Maxime Marquette, columnist

Columnist's transparency note

Stance and Method

This analysis comes from an observer of technology and geopolitical trends, not a semiconductor engineer. I rely on information available in public journalistic and industry sources. My pro-Western positions on technological sovereignty questions shape my analysis. I believe in the necessity of diversifying technology supply chains to reduce geopolitical vulnerabilities.

What I Do Not Know

I cannot precisely predict the timelines for resolving the shortage or the scale of forthcoming price increases. Information on the technology roadmaps of memory manufacturers is partially confidential. The market data I use is based on analyst estimates that may be revised. This analysis represents the best possible judgment from sources available on June 27, 2026.

Sources

Primary sources

ABC News — Apple increases prices for Macs and iPads, blaming a shortage of memory chips — June 26, 2026

9to5Mac — Apple confirms price increases due to RAM shortage — June 17, 2026

The Register — Apple passes 'Rampocalypse' costs on to consumers — June 25, 2026

Secondary sources

The Guardian — economic and technology analysis — June 2026

Foreign Policy — geopolitics of semiconductors — 2026

Axios — technology coverage and semiconductor markets — June 2026

19FortyFive — national security and technology supply chains — 2026