The Traders of Huaqiangbei
In the labyrinthine electronics markets of Shenzhen's Huaqiangbei district—the pulsing heart of China's hardware trade—prices now change by the hour. A 16GB DDR4 memory module that sold for 180 yuan in early 2025 commands over 400 yuan today, with premium brands pushing 420 . Samsung's 1TB solid-state drives have vaulted from 200-odd yuan to 620; the 2TB models breach 1,200 . Traders describe storage products as "black gold bars," their value tripling in months . One vendor's refrain has become the district's dark mantra: "yī tiān jǐ gè jià"—several prices a day .
This is not speculation or cryptocurrency mania. This is the leading edge of a global memory famine that threatens to reshape the entire consumer electronics industry, shuttering companies, doubling production costs, and forcing the first sustained contraction in smartphone and PC shipments in a generation. From Detroit assembly lines to Cupertino design studios, the shortage is rewriting the economics of every device that thinks, remembers, or computes. And unlike previous semiconductor crunches—cyclical affairs resolved by capacity expansions—this one has no clear end in sight.
The question facing manufacturers, investors, and consumers is no longer whether memory will become scarce, but whether the shortage will last until 2027, 2028, or, as one major industry chairman now projects, 2030 . The answer will determine which companies survive, which products reach market, and who controls the commanding heights of the AI economy.
The Hunger That Cannot Be Fed
The proximate cause is straightforward: artificial intelligence has developed an insatiable appetite for memory. Where a smartphone might use 8 or 12 gigabytes of DRAM, a single AI training server can consume 80 gigabytes or more of high-bandwidth memory. Enterprise SSD orders for data centres continue to surge, driving what industry analysts describe as sustained, structural demand rather than a temporary spike . The hyperscalers—Amazon, Microsoft, Google, Meta—are purchasing not just capacity but entire production roadmaps, locking in supply years in advance.
Western Digital's chief executive recently revealed that the company's entire hard disk drive capacity for 2026 is already booked . Not mostly booked. Not substantially committed. Completely sold out, more than a year ahead of delivery. In the NAND flash market, November contract prices surged over 60 per cent as wafer supply tightened sharply . Some foundries now demand three-year cash payments upfront, according to Phison's chief executive, Khein-Seng Pua—a terms structure previously unheard of in semiconductor manufacturing .
The shortage extends beyond the chips themselves to the materials that enable their production. Glass cloth, an obscure but essential component in advanced chip substrates, faces severe supply constraints affecting the entire tech industry . These substrates form the physical foundation upon which memory chips are mounted; without them, even functional silicon cannot become a usable product. It is the industrial equivalent of having flour but no yeast—the machinery of production grinds forward, but the final product remains out of reach.
The result is a market that has abandoned conventional pricing mechanisms entirely. According to a DigiTimes report, DRAM prices have begun shifting on an hourly basis as the AI-driven shortage intensifies . Small and medium-sized businesses, unable to compete with the purchasing power of hyperscalers and major OEMs, find themselves fighting for survival in spot markets where prices can move 5 or 10 per cent between morning and afternoon trading sessions. The orderly world of quarterly contract negotiations has given way to something closer to commodity pit trading—frantic, volatile, and fundamentally disconnected from manufacturing cost structures.
The Arithmetic of Disappearance
HP, one of the world's largest PC manufacturers, reported that memory costs increased roughly 100 per cent sequentially from the first to second quarter of this year, now accounting for 35 per cent of the cost of materials needed to build a personal computer . Let that figure settle: more than a third of a PC's bill of materials is now memory alone, up from roughly half that proportion just months earlier. This is not a marginal input cost fluctuation. This is a fundamental restructuring of product economics.
Dell, HP, and other major technology firms have warned publicly of the memory chip squeeze driven by AI demand . China's top chipmaker, Semiconductor Manufacturing International Corporation (SMIC), projects that the shortage may constrain car and electronics production in 2026 . Modern vehicles contain dozens of microcontrollers and processors, each requiring memory; electric vehicles and those with advanced driver-assistance systems require substantially more. A shortage that began in data centres is now propagating through supply chains to affect industries with no direct connection to artificial intelligence.
The International Data Corporation (IDC) has begun revising its forecasts accordingly. The worldwide smartphone market is expected to decline by 12.9 per cent in 2026, with revenues falling slightly by 0.5 per cent . The PC market faces an 11.3 per cent unit decline, though revenues are projected to grow 1.6 per cent as increased average selling prices offset volume contraction . This is the signature of a supply-constrained market: fewer units sold at higher prices, with manufacturers unable to meet demand even as consumers prove willing to pay.
Phison's Pua went further, warning that the flash memory shortage could lead to the shutdown of consumer electronics companies in 2026 . Not production cuts. Not delayed product launches. Shutdowns—permanent exits from markets where margins have compressed to zero or below, where securing memory allocation requires cash reserves and credit lines that smaller players simply do not possess.
The Horizon Dispute
When the shortage will ease has become the industry's central debate, and the answers span nearly a decade.
Micron's chief executive, Sanjay Mehrotra, expects the RAM shortage to last through 2027, with conditions beginning to ease in 2028 . Samsung, the world's largest memory manufacturer, expects the DRAM supercycle to fade away by 2028, a projection that has already led the company to adopt cautious investment in expansion plans . These timelines reflect the brutal arithmetic of semiconductor manufacturing: even if a decision to expand capacity is made today, new fabrication plants require two to three years to construct and bring to volume production.
But SK Group's chairman, Chey Tae-won, offered a darker assessment: the global memory chip shortage will persist for another four to five years, extending to 2030 . His reasoning centres on a supply-demand mismatch that cannot be quickly resolved. Wafer supply currently trails demand by 20 per cent, he noted —a gap so substantial that even aggressive capacity expansion would require years to close, assuming demand remains constant. If AI adoption accelerates further, the timeline extends accordingly.
A competing view comes from Kye-hyun Kyung, a former Samsung executive, who believes that memory prices will fall in the second half of next year . His argument rests on China's emerging DRAM manufacturing capacity, which he suggests could "crush the 414 per cent DDR5 price spike within a year" . Chinese producers, operating with state support and lower cost structures, could flood the market with sufficient volume to break the shortage—or at least moderate its most extreme price dislocations.
This divergence is not merely academic. If Kyung is correct, companies hoarding memory or paying spot-market premiums will find themselves holding overpriced inventory as prices collapse. If Chey is correct, firms that fail to secure long-term supply agreements now will spend the next five years unable to manufacture products at competitive costs. The uncertainty itself becomes a strategic problem, forcing companies to make billion-dollar bets on fundamentally unknowable futures.
The Cartel Question
Into this fraught landscape arrives a lawsuit that asks whether the shortage is entirely organic. Samsung, SK Hynix, and Micron—the three companies that collectively control roughly 95 per cent of the global DRAM market—are being sued for allegedly price-fixing RAM amid the global memory supply shortage .
The timing is provocative. Memory prices have indeed surged to unprecedented levels, and the industry has a history: all three companies have faced price-fixing allegations and settlements in previous cycles. The fundamental structure of the DRAM market—high barriers to entry, massive capital requirements, a small number of sophisticated players—creates conditions where coordinated behaviour becomes possible even without explicit collusion. When only three firms control nearly all supply, their individual capacity decisions collectively determine market outcomes.
The legal complaint must contend with a genuine, observable shortage driven by AI demand. Proving that prices reflect coordination rather than scarcity requires demonstrating that supply has been artificially constrained—that manufacturers possessed the ability to produce more but chose not to in order to sustain elevated prices. Samsung's decision to adopt "cautious investment in expansion plans" despite expecting the supercycle to fade by 2028 could be interpreted as prudent capital discipline or as strategic supply management, depending on one's perspective.
What is indisputable is that the three defendants are indeed rethinking expansion plans, worried that the demand boom won't last "too long" . This conservatism is understandable from a shareholder perspective—memory manufacturers have repeatedly overbuilt capacity in previous cycles, leading to brutal price collapses and years of losses. But it also ensures that supply will remain constrained, prices will remain elevated, and the shortage will persist longer than if all players built capacity aggressively.
"Western Digital's entire HDD capacity for 2026 is booked out. Not mostly booked. Completely sold out, more than a year ahead of delivery."
The Valve in the Machine
Amid this crisis, one data point stands curiously isolated: Valve will release the Steam Controller on 4 May 2026, before the Steam Machine or Steam Frame . The announcement, seemingly tangential to memory markets, actually illuminates the shortage's reach. Valve, a company known for hardware ambition—the Steam Deck handheld console, the Index VR headset—is staggering its product launches in ways that suggest constrained component availability.
The Steam Controller requires minimal memory compared to a full gaming console; prioritising its release allows Valve to bring a product to market without competing for the high-capacity memory modules that more ambitious devices demand. This is adaptation to scarcity: companies are redesigning roadmaps, launching simpler products first, and shelving memory-intensive projects until supply conditions improve.
Across the industry, similar calculations are underway. Smartphone manufacturers are reducing the number of SKUs they offer, cutting configurations with 12GB or 16GB of RAM and focusing production on 6GB and 8GB models. Laptop makers are extending product cycles, refreshing chassis and displays while leaving memory configurations unchanged. Even planned features—on-device AI assistants, high-resolution video recording, advanced multitasking—are being deferred because they require memory that cannot be sourced at acceptable cost.
The shortage is thus not merely inflating prices; it is actively shaping what gets built, what reaches consumers, and what technological futures remain accessible in the near term.
The Unseen Hand
What makes this shortage historically significant is its origin outside the memory industry itself. Previous semiconductor shortages emerged from supply shocks—a fire at a fabrication plant, an earthquake disrupting logistics, underinvestment during a downturn. This shortage emerges from a demand shock of unprecedented scale and uncertain duration.
Artificial intelligence is not a product cycle that peaks and fades. It is a fundamental reordering of how computation happens, shifting workloads from human-facing devices with modest memory requirements to massive data centres where memory is the primary constraint on model size and inference speed. Every percentage point improvement in model capability requires exponentially more memory. Every new AI application—code generation, image synthesis, conversational agents—requires inference infrastructure that consumes memory at scale.
The hyperscalers understand this, which is why they are locking in supply years in advance, paying premiums that would bankrupt smaller buyers, and vertically integrating into chip design to optimize for their specific workloads. Their memory purchases are not inventory speculation; they are strategic investments in AI platform control. Whoever secures memory secures the ability to deploy models, serve users, and extract value from the AI transition.
This creates a zero-sum dynamic. Memory allocated to a data centre is memory unavailable for smartphones, PCs, automobiles, industrial systems, medical devices, or any of the thousands of other applications that have come to depend on cheap, abundant semiconductors. The AI boom is cannibalising the rest of the electronics industry, and there is no obvious equilibrium where both can be fully satisfied.
What Comes Next
The industry is preparing for a prolonged siege. SMIC's warning about constrained car and electronics production in 2026 reflects a growing consensus that conditions will worsen before they improve. Phison's warning about potential company shutdowns is not hyperbole; it is risk assessment. Smaller players, unable to secure memory at any price, will face a stark choice: exit the market or accept permanent second-tier status, designing products around whatever scraps of supply remain after the giants have taken their fill.
For consumers, the implications are rising prices and diminishing choice. The smartphone market's projected 12.9 per cent contraction in 2026 will manifest as fewer new models, longer replacement cycles, and devices that cost more while offering less. The PC market's shift toward higher average selling prices despite unit declines means budget systems will become scarce, pushing entry-level buyers toward older technology or out of the market entirely.
For manufacturers, survival requires securing supply now, even at punitive prices, or accepting that production will be supply-limited regardless of demand. HP's experience—memory costs doubling in a single quarter to consume more than a third of materials costs —is becoming universal. Companies are redesigning products to use less memory, negotiating multi-year supply agreements, and in some cases vertically integrating into memory production itself.
The wild card remains China. If Kyung's assessment proves correct and Chinese DRAM production scales rapidly , the shortage could moderate faster than the pessimistic projections suggest. But Chinese producers face their own constraints—access to advanced manufacturing equipment, technology transfer restrictions, and the sheer difficulty of matching the quality and reliability of Samsung, SK Hynix, and Micron's cutting-edge processes. A flood of cheap but lower-specification Chinese memory might relieve pressure on budget devices while leaving the high-performance segment still constrained.
The Price of Intelligence
In Huaqiangbei, the traders adjusting prices by the hour understand something the broader market is still absorbing: memory has become the bottleneck resource of the digital economy. Not talent, not algorithms, not energy—memory. The ability to store and rapidly access vast quantities of data determines who can train models, deploy applications, and compete in the AI era.
This was not supposed to happen. Memory was supposed to be a commodity, governed by Moore's Law, perpetually falling in cost and rising in abundance. The entire architecture of modern computing assumed cheap memory, designing systems that used it profligately because efficiency was unnecessary. Now efficiency is mandatory, abundance is over, and the industry is repricing everything accordingly.
The shortage will eventually ease—perhaps in 2027, perhaps in 2030, perhaps when Chinese production reaches scale or demand growth finally slows. But the transition period will determine which companies survive, which industries adapt, and which technological futures remain accessible. Those holding memory hold leverage. Those without it must wait, pay, or disappear.
In the meantime, prices change by the hour, capacity books out years in advance, and the electronics that have become infrastructure for modern life grow scarcer and more expensive. The AI revolution has arrived, and it is hungry. What it hungers for is memory, and there is not enough to go around.