← Back to Library

Memory mania: How a once-in-four-decades shortage is fueling a memory boom

Dylan Patel · SemiAnalysis · ·15 min read
Commentary by Hex Index staff

Memory's Bottleneck

Dylan Patel's analysis cuts through the noise around semiconductor shortages with a sobering claim: the memory shortage driving tech markets isn't nearing its peak. What makes this piece notable is Patel's willingness to ground cyclical hype in physical constraints—capacitors holding tens of thousands of electrons, bitlines becoming the dominant bottleneck, scaling laws that once delivered 100× density gains per decade now delivering barely 2×.

The Physics of Scarcity

Patel writes, "DRAM density has increased by only ~2× in total, versus roughly ~100× per decade during the industry's peak scaling era." This collapse of Dennard scaling—the principle that transistor power efficiency improves as they shrink—has transformed memory from a technology-driven deflationary force into a commodity governed by supply-demand mismatches.

Memory mania: How a once-in-four-decades shortage is fueling a memory boom

As Dylan Patel puts it, "The static charge on just a speck of dust might be 10,000x what is stored in a modern DRAM cell." The analogy is visceral: engineers have shrunk DRAM cells into tall, narrow straws holding so few electrons that manufacturing variations or temperature shifts cause errors. Signal integrity, not transistor count, now dictates what's possible.

The static charge on just a speck of dust might be 10,000x what is stored in a modern DRAM cell.

Critics might note that Patel's framing assumes no breakthrough architecture will sidestep these physical limits—yet research into resistive RAM, phase-change memory, and 3D stacking continues. The bottleneck is real, but not necessarily permanent.

Cycle Mechanics

The memory industry's cyclicality stems from capital intensity. Patel writes, "Building leading-edge DRAM and NAND fabs requires multi-billion-dollar investments, multi-year construction timelines, extended yield-learning curves across successive process nodes, and lengthy ramp-up periods before meaningful volume production is achieved." Once built, fabs must run—even in downturns—because sunk costs make idling economically impossible.

Dylan Patel notes, "Node transitions do not halt simply because demand weakens. Consequently, bit supply growth can remain robust well into downturns." This dynamic exacerbates oversupply. Samsung's 1c node delivers 70% more bits per wafer than its 1a predecessor, meaning supply expands even if wafer output stays constant.

Consolidation and Power

The industry has consolidated dramatically. Patel writes, "From roughly 20+ players in the mid-1990s, the number of players contracted to the mid-teens in the 2000s and early 2010s, to fewer than 10 relevant suppliers in the 2020s. Today, there are only 3-4 material suppliers." This concentration gives remaining players pricing power but creates systemic risk—three companies control a critical input for everything from smartphones to AI accelerators.

As Dylan Patel puts it, "By the time pricing rolls over, manufacturers have already committed and deployed multi-billion-dollar capital investments into fabs and equipment that cannot be economically idled." The result is margin compression precisely when balance-sheet stress rises.

Historical Precedent

Patel revisits three supercycles: 1993 Windows PC adoption (DRAM content per PC jumped 4×), 2010 smartphone and cloud buildout (LPDDR standardization compressed pricing power), and 2017-2018 cloud-NAND demand. Each peaked within one or two years as elevated profitability triggered aggressive capacity expansion.

The current AI-driven demand surge mirrors these inflection points. But Patel's warning is explicit: "The scariest thing is that we aren't even close to the peak." Historical patterns suggest the peak will trigger capex surges that eventually flip shortage into oversupply.

Critics might argue that AI workloads differ fundamentally from PC or smartphone adoption—memory intensity per GPU server may sustain demand longer than prior cycles. Yet Patel's data on node migration and yield learning suggests supply will accelerate regardless.

Bottom Line

Memory's physical limits have transformed it from a deflationary technology into a cyclical commodity governed by capital discipline and demand inflections. Patel's analysis is credible because it grounds forecasts in fab-by-fab production data rather than market sentiment. The verdict: buyers should expect continued price pressure through 2026, but the peak will trigger the next oversupply correction—history demands it.

Deep Dives

Explore these related deep dives:

Sources

Memory mania: How a once-in-four-decades shortage is fueling a memory boom

by Dylan Patel · SemiAnalysis · Read full article

Prices of memory are going crazy. SemiAnalysis has been calling this out for over a year since late 2024. The scariest thing is that we aren't even close to the peak. We go through fab by fab production and expansion versus detailed end market demand by memory type to forecast memory revenue, pricing, and margin better than anyone else. This has all been detailed in the SemiAnalysis memory model for a while, but we will share it more publicly today. First some background.

The Inevitability of Memory Cycles: A History of Booms and Busts.

Since its commercial introduction in the 1970s, DRAM has benefited from the two scaling laws that defined the semiconductor industry: Moore’s Law and Dennard scaling. The 1T1C DRAM cell, with one access transistor and one storage capacitor, scaled for decades. Shrinking transistors reduced cost per bit, while clever capacitor engineering preserved sufficient charge to maintain signal integrity.

For much of the industry’s history, DRAM density scaled faster than logic, doubling roughly every 18 months instead of 24 months and driving dramatic cost reductions. As a commoditized product, manufacturers needed to sustain cost-per-bit declines to stay competitive. Suppliers who couldn’t compete on cost fell into a downward spiral: low sales left them short on cash to finance next-generation nodes, which in turn left them further behind on cost-per-bit. Many DRAM producers fell victim and went into bankruptcy, resulting in consolidation to just a few major players today.

For more details on the industry and DRAM basics, check out our technical deep dive:

Yet DRAM scaling has slowed significantly over the past few decades, and density gains over time have shrunk. Over the past decade, DRAM density has increased by only ~2× in total, versus roughly ~100× per decade during the industry’s peak scaling era. Capacitors are now extreme three-dimensional structures with aspect ratios approaching 100:1, storing just tens of thousands of electrons. For comparison, a small static shock when you touch a metal doorknob might involve the transfer of billions of electrons. The static charge on just a speck of dust might be 10,000x what is stored in a modern DRAM cell.

Bitlines and sense amplifiers, once secondary concerns, are now dominant constraints. Every incremental shrink reduces signal margin, increases variability, and raises cost.

An easy way to understand the technical challenge in DRAM scaling is to think of a DRAM cell as a tiny bucket that ...