DECODING: The DRAM Crisis — Apple, AI, and the Next Global Memory Shortage

DECODING: The DRAM Crisis — Apple, AI, and the Next Global Memory Shortage

Introduction: Micron Points the Finger at Cupertino

The accusation that shook the industry

In the carefully managed world of supplier-client relations in the tech industry, it is rare for a component maker to publicly accuse one of its largest customers of contributing to a global supply crisis. Yet that is exactly what happened in late June 2026, when Micron Technology — one of the three major global manufacturers of DRAM memory, alongside Samsung and SK Hynix — suggested that Apple had helped create the current memory crisis by locking up production capacity for its new devices. The revelation, reported by MacRumors on June 26, 2026, landed like a bomb in a sector already under extreme pressure.

The mechanism is straightforward: when Apple — whose order volumes run into tens of millions of units per product cycle — reserves a large share of memory foundry production capacity, other customers, including server manufacturers, PC makers, and telecom equipment suppliers, are left with insufficient allocations. In a market already stretched by the enormous demand from artificial intelligence data centers, this capacity lockup by a single dominant player adds further pressure to the entire supply chain.

Who is responsible for the DRAM shortage?

The question of responsibility in the current DRAM shortage is complex. Demand for high-performance memory (HBM — High Bandwidth Memory) to feed NVIDIA GPUs and AI accelerators has literally exploded since 2023. Data centers are consuming unprecedented quantities of memory to train and run large language models. According to TrendForce estimates, server DRAM prices climbed more than 60% quarter over quarter in early 2026. Avnet analysts estimated in February 2026 that AI data centers could absorb as much as 70% of global premium DRAM production in 2026.

In this context, Apple is not the sole cause. But its practice of capacity locking — guaranteeing launch volumes by reserving production lines far in advance — has become an aggravating factor in an already depleted market. The Micron declaration is simultaneously an industrial warning signal and a challenge to a commercial practice that the tech industry has never had to justify publicly before.

DRAM: The Invisible Foundation of the Entire Digital Economy

Why DRAM cannot be replaced in the short term

DRAM (Dynamic Random Access Memory) is the type of volatile memory that allows all computing systems to operate at high speed. Unlike SSDs or hard drives, which store data permanently, DRAM is working memory: fast, volatile, indispensable. In an AI training server, the quantity of available DRAM directly determines the size of models that can be trained, the speed of inference, and the number of simultaneous requests that can be handled.

The most critical variant for AI is HBM (High Bandwidth Memory), an advanced form of DRAM stacked in three-dimensional layers to maximize data throughput. HBM3e — the current generation — is produced almost exclusively by SK Hynix and Samsung, with Micron gradually closing the gap. The concentration of this production in two South Korean companies represents a geopolitical and supply risk that few analysts had sufficiently factored in before the current crisis.

A market in extreme seller's territory

According to Avnet data published in February 2026, suppliers are receiving less than 50% of the silicon wafer allocations they request — meaning customers are ordering double what they actually receive. Contractual DRAM prices climbed 50%+ quarter over quarter in early 2026, with some market trackers revising their forecasts upward to 90–95% quarterly increases. These are volatility levels the industry had not seen since the major shortages of the 1990s and 2000s.

This context makes Micron's accusation against Apple all the more significant. In a normal market, capacity lockup by a major customer is routine and manageable. In a market where demand systemically outstrips supply, every gigabyte of capacity captured by one player translates directly into a shortage for others. It is a common-resource problem — and poorly managed common resources lead to collective tragedies.

The Role of AI Data Centers in the Explosive Demand

The limitless appetite of LLMs and GPUs

To understand the DRAM crisis, you need to understand the economics of AI model training. A large language model like GPT-5 or its successors requires thousands of GPUs working in parallel for weeks or months. Each H100 or B200 generation NVIDIA GPU is equipped with HBM3e — typically between 80 and 192 gigabytes per card. A rack of 1,000 GPUs therefore uses between 80 and 192 terabytes of DRAM in HBM form. Hyperscalers like Microsoft, Google, Meta, and Amazon deploy clusters of 10,000 to 100,000 GPUs. The aggregate demand is colossal.

SpaceX itself, which merged with xAI in February 2026, operates massive AI data centers like Colossus 2 — whose supply of NVIDIA GB300 chips is the subject of a $6.3 billion agreement with Reflection AI. The explosion in AI memory demand is not a passing phenomenon: it is a structural trend that will accelerate with every new generation of larger, more powerful models requiring even more DRAM.

TrendForce analysis and projections through 2027

TrendForce, the Taiwan-based semiconductor analysis firm, projects that tension on DRAM markets will persist at least until 2027, possibly beyond. The structural reason is simple: adding new DRAM production capacity requires massive investment in fabs and construction lead times of two to four years. Samsung, SK Hynix, and Micron are investing tens of billions in new capacity, but those investments will not translate into additional volume until 2027–2028.

In the meantime, the industry is in a classic semiconductor feast-and-famine cycle. Today: famine. Prices are rising, allocations are insufficient, customers are fighting over available wafers. In two to three years, if investments materialize and demand does not accelerate further, there could be temporary overcapacity — and a price collapse. The violent cyclicality of the DRAM market is nothing new, but its amplitude in the current cycle is unprecedented.

The Capacity-Locking Strategy of Tech Giants

How Apple, NVIDIA, and others protect themselves at everyone else's expense

The practice of capacity locking (capacity reservation) is neither illegal nor new. Major tech players negotiate multi-year agreements with their suppliers guaranteeing production volumes. Apple has historically been expert in this field — its scale allows it to impose favorable terms on suppliers and monopolize entire production lines during new product launches.

But when the entire market is under stress, these practices create a crowding-out effect on less powerful buyers. PC manufacturers, industrial equipment makers, and electric vehicle producers who need memory for their embedded systems all find themselves at the back of the queue. In the medical sector, in defense systems, in critical infrastructure — wherever DRAM is an essential component, the shortage is felt. It is a negative externality of the tech giants' dominance over supply chains.

Strategic stockpiles and public policy questions

In response to this situation, several governments are considering strategic stockpile policies for critical semiconductors, modeled on strategic petroleum reserves. The concept is straightforward: build buffer stocks of essential components to absorb shortage shocks. But implementation is complex: electronic chips age and become obsolete quickly, and storing significant volumes represents a colossal investment.

The United States with the CHIPS Act, Europe with the European Chips Act, Japan with its subsidies to TSMC — all are attempting to partially reshore semiconductor production. But for DRAM specifically, ambitions are more limited. Micron is the only significant American player, and its market share remains below that of Samsung and SK Hynix. Memory independence is a 2030 target at best — and even that requires massive ongoing investment.

The Implications for the AI Industry and Its Users

The companies paying the highest price

The DRAM crisis translates concretely into exploding operating costs for companies that depend on AI inference. Startups building AI services — from virtual assistants to content generation tools, medical systems, and financial analytics platforms — see their processing costs rise in direct proportion to memory price increases. For companies building their business model on thin margins, this is existential pressure.

Hyperscalers (Microsoft, Google, Amazon, Meta) absorb these increases more readily thanks to their purchasing power and long-term agreements. But even they pass a portion of the costs on to their enterprise customers. The cost per token — the unit of processing for language models — is directly tied to DRAM availability and pricing. A prolonged shortage could slow AI adoption in less capital-intensive sectors.

Innovation constrained by physical resources

There is a deep irony in this situation: the AI revolution, promised as a purely software and algorithmic transformation, is running into fundamental hardware constraints. Algorithmic advances can improve model efficiency, reduce memory requirements per parameter, and optimize architectures — and those advances are real. Mistral, DeepSeek, Llama, and others have demonstrated that much is achievable with less.

But frontier models continue to grow in size, power, and memory demand. GPT-5, the successors to Gemini Ultra, future multimodal systems — all require more DRAM than their predecessors. The next shortage of 2027, if current trends hold, could be even more severe than the one in 2026. Companies that plan ahead now — diversifying suppliers, signing long-term agreements, investing in memory-efficient architectures — will be better positioned to weather it.

Market Trends to Watch Through 2027

Early warning signals of the next shortage

Several early signals indicate that the next DRAM shortage is already forming for 2027. First, the massive investment announcements by the three major manufacturers (Samsung pouring hundreds of billions of won into new fabs, SK Hynix building new HBM4 facilities, Micron accelerating its American construction program) will not yield results until 2028 at the earliest. Second, demand continues to accelerate with new generations of GPUs (NVIDIA Rubin, AMD MI400) whose memory specifications exceed those of previous generations.

Third, emerging AI applications — autonomous vehicles, humanoid robots, autonomous defense systems, and medical AI — will all represent new sources of memory demand that barely existed at scale in 2024. The next wave of edge AI — AI running on the devices themselves rather than in the cloud — will create demand for energy-efficient mobile memory that will compete with data center server demand, adding another dimension to an already complex supply picture.

HBM4 and the stakes of the next generation

The transition to HBM4 — the next generation of high-bandwidth memory — is a major industrial competitiveness issue. SK Hynix and Samsung are racing to be first to volume production of this technology. Micron, historically behind on HBM, is working to close the gap with its American and Japanese fabs. Geopolitics enters here as well: semiconductors produced in the United States or Japan are less vulnerable to disruption tied to a crisis on the Korean peninsula or in the Taiwan Strait.

For the AI industry, the availability of HBM4 in volume will determine the feasibility and cost of the next generations of foundation models. If this transition proceeds smoothly, the current shortage could ease. If it is delayed by technical problems or geopolitical shocks, we enter an era of structural memory rationing for AI — with profound implications for the technological competitiveness of nations.

The Geopolitics of Silicon: Systemic Risks for the West

DRAM production concentration as a strategic vulnerability

Behind the commercial shortage lies a geopolitical vulnerability that few Western governments have seriously addressed. More than 70% of global DRAM memory production is handled by two companies: Samsung and SK Hynix, both Korean, with their main facilities located in South Korea — a country 60 kilometers from the North Korean border whose security depends on a defense treaty with the United States. The third player, Micron, produces in the United States, Japan, and Taiwan. The dependence on Taiwan for cutting-edge technologies is particularly concerning given persistent tensions with Beijing.

In the event of a conflict in the Taiwan Strait — a scenario that American and Japanese military planners now treat as a real rather than theoretical possibility — semiconductor supply chains, including memory, would be severely disrupted. Simulations conducted by the CSIS (Center for Strategic and International Studies) in 2025 estimated that such a conflict would produce a global economic shock of several trillion dollars and paralyze Western high-tech industries for a minimum of 12 to 18 months. The current AI-driven DRAM shortage is a minor commercial disruption compared to that extreme scenario.

Policy responses: the CHIPS Act and its limits for memory

The American CHIPS and Science Act of 2022, endowed with $52 billion in subsidies, has primarily benefited logic chip production (TSMC, Intel, Samsung Logic) rather than DRAM memory. Micron announced investments of $40 billion in the United States enabled by the CHIPS Act, but its new advanced DRAM-producing facilities will not enter production until 2027–2030. The European Chips Act is even further behind on executing its ambitions. Resilience of DRAM supply will therefore remain limited for the next several years.

South Korea, aware of its pivotal position, announced in March 2026 a "Semi-Shield" program — a mechanism allowing the government to coordinate semiconductor exports in a geopolitical crisis, modeled on the oil export control systems of the 1970s. This mechanism would give Seoul unprecedented negotiating leverage in its relations with Beijing, Washington, and major industrial customers. Memory is becoming an instrument of state power.

Alternative Players: China's Rise in Memory

CXMT and Chinese ambitions in DRAM

Changxin Memory Technologies (CXMT), China's leading DRAM manufacturer, crossed several technological milestones in 2025–2026 that would have seemed impossible five years earlier. Its DDR5 chips — the current standard generation — are reaching performance levels comparable to Samsung products from two years ago. American export controls have slowed CXMT but have not stopped it: the company has turned to Dutch (ASML), Japanese, and Korean alternative equipment makers, and has developed some processes internally. It is not yet competitive on cutting-edge HBM, but it is progressing at a pace that alarms industry analysts.

If CXMT — with massive state backing — achieves global competitiveness in standard DRAM by 2028–2030, the geopolitical implications are considerable. China would become partially self-sufficient in memory, reducing its exposure to Western sanctions. More critically, it could offer a lower-cost alternative to countries in the Global South seeking to avoid Western technological dependency. The balkanization of semiconductor supply chains, already underway for advanced processors, would extend to memory.

The impact on global prices and demand in the medium term

The entry of CXMT into the global DRAM market would have a deflationary effect in the medium term — if and when it materializes. TrendForce analysts estimate that if CXMT achieves 10% global market share by 2029, average standard DDR5 prices could fall 15 to 20% relative to current projections. For AI server manufacturers looking to cut infrastructure costs, this prospect is attractive. For Western players who have invested massively in new production capacity, it represents a significant margin threat.

The reality is that the DRAM memory market has entered an era of structural turbulence of which the current shortage is only one episode. The convergence of explosive AI demand, the geographic concentration of production, China's rise, and geopolitical risks on supply chains creates durable instability. Tech companies that depend on access to competitively priced DRAM — meaning all tech companies — will need to manage this uncertainty as a top-tier strategic risk in their 5 and 10-year planning.

Conclusion: Memory Is the Oil of the AI Era

A strategic resource not taken seriously enough

DRAM — and in particular HBM — has become a critical strategic resource on par with oil for the 20th-century industrial economy. Except that unlike oil, we have not yet built the institutions, strategic reserves, and industrial policies to manage this dependency in a mature way. Micron's accusation against Apple is a symptom of a dysfunctional market in which the most powerful players — Apple, the hyperscalers, GPU manufacturers — capture available resources at the expense of everyone else.

The correction will come, as it always does: through new investments, new capacity, perhaps architectural innovations that reduce memory demand per computation. But the 2026–2027 window will be difficult. And the next shortage, already gestating, could be even more severe. Governments, companies, and investors who have not yet grasped this had better update their understanding — and fast.

What decision-makers must do now

The right policy response to the DRAM crisis combines several levels. At the industrial level: accelerate investment in new production capacity and diversify supply chains geographically. At the regulatory level: examine whether the capacity-locking practices of tech giants constitute anticompetitive behavior during shortage conditions. At the strategic level: integrate the availability of critical semiconductors into national security planning alongside rare earth metals and energy raw materials. DRAM is not just a computing component — it is the infrastructure of tomorrow's economy.

By Maxime Marquette, columnist

Columnist's transparency note

Expertise and limits on this subject

Maxime Marquette is not a semiconductor engineer. This analysis rests on recognized specialized sources in the electronic components industry — TrendForce, Avnet, official statements from Micron, and MacRumors — rather than on first-order technical expertise. Market projections for 2027 are by nature uncertain and could be invalidated by unforeseen technological developments or investments. I hold no financial position in any of the companies cited.

Sources and verification

The price and allocation figures cited come from specialized industry publications. Data on AI data center demand consists of analyst estimates, not certified figures. Apple's implication in the shortage rests on Micron's statements as reported — I was unable to independently verify the details of commercial agreements between Apple and its memory suppliers.

Sources

Primary sources

MacRumors — "Micron Suggests Apple Helped Cause Memory Crisis" — June 26, 2026

Avnet — "Riding the AI Supercycle: Navigating the 2026 Memory & Storage Shortage" — February 2026

Secondary sources

CNBC — AI investment and infrastructure demand data (SpaceX/xAI Colossus) — June 22, 2026

OpenAI — Daybreak initiative and AI infrastructure requirements — June 22, 2026

Defense News — context on critical technology supply chains — June 2026

Foreign Policy — geopolitical analyses of semiconductor supply chains — June 2026

Kyiv Independent — context on strategic technology vulnerabilities in conflicts — June 2026