Even with the market still booming, memory has rarely sustained positive growth for more than five consecutive years. That is the central argument in a TechFlowPost market analysis by Takashi Yunokami, who says the semiconductor industry — and memory in particular — is now in an expansion so steep that it falls outside the range of what the past three decades would normally suggest.
Memory broke away from the industry’s long-term pattern after around 2024
Using World Semiconductor Trade Statistics, or WSTS, data, the article says the four major semiconductor categories — micro devices, memory, logic and analog — generally moved along a steady upward trend from 1991 through May 2026. In the author’s view, that was the industry’s normal development path despite changes in the broader economy.
What changed, he says, is the shape of the curve after around 2024. Logic, and especially memory, began to rise in a near-vertical fashion. Memory stood out most. Its growth nearly pierced the top of the chart, making the cumulative expansion of the prior 30 years look modest by comparison.
Yunokami writes that, based on his years of watching the semiconductor market, memory has never shown a climb this steep. He describes it as a historical anomaly that is shaking the structure of the industry itself.
Monthly memory shipments rose from $5.6 billion in 2016 to $63.3 billion in May 2026
The article narrows in on the post-2016 period. Monthly MOS memory shipments were about $5.6 billion in 2016. After a short-lived boom during the 2017-2018 memory bubble and a later slump, shipments fell to just $5.8 billion at the start of 2023, which the author treats as the bottom of the downturn.
What followed was not a normal recovery. By May 2026, monthly memory shipments had climbed to $63.3 billion. That is more than 11 times the 2016 level, according to the figures cited in the piece, and about 10.7 times the trough seen in early 2023.
Logic also expanded, but far less dramatically. The article says the logic business, including NVIDIA GPUs, grew from $13.3 billion to $31.6 billion. Yunokami calls that the combined effect of NVIDIA GPUs and AI, though he says the rise still looks moderate when set against the move in memory.
Year-over-year growth reached 285%, far above prior memory-cycle peaks
The distortion becomes clearer in growth-rate terms. The article says memory’s latest year-over-year growth rate reached 285%, a record high.
For comparison, Yunokami says the previous memory bubble, around 2017, peaked at roughly 60% annual growth. Other semiconductor categories remained well below that level in the current cycle. Logic was around 40%, while microcontrollers and analog were around 14% to 19%. In his reading, only memory has moved onto what he calls another level.
DRAM and NAND are driving the surge
Within memory, the article identifies DRAM and NAND flash as the two products behind the expansion. Using TrendForce data, it sketches out how sharply those markets have rebounded from the industry downturn.
At the start of 2023, DRAM had fallen to $9.7 billion and NAND to $8.7 billion. The author describes that period as a dark phase marked by losses across memory makers and forced production cuts.
By the second quarter of 2026, TrendForce is projected to put the DRAM market at $145 billion and NAND at $81 billion. Compared with the 2023 trough, that would represent roughly 15x growth for DRAM and about 9x growth for NAND. Together, the two would total $226 billion for the quarter, implying an annual pace above $900 billion.
The article argues that memory, once treated as a low-cost commodity, is now preparing to challenge logic and MPUs for leadership inside the semiconductor industry.
Pricing, not just volume, is the core reason revenue has exploded
Yunokami says the market’s expansion cannot be explained by shipment volume alone. In his view, the biggest factor is the extreme jump in memory pricing.
Citing TrendForce DataTrack, the article says DRAM spot pricing was just $4.70 in early 2025 and has recently surged to $46.00, about a tenfold increase. NAND 1Tb TLC wafer pricing rose from $2.40 to $25.00 over the same period, also close to 10x.
That means a tenfold increase in market size does not imply shipments also rose tenfold. A large share of the jump came from unit pricing. The article states plainly that even if volume had stayed flat, revenue could still have multiplied if prices rose by nearly the same amount.
For memory manufacturers, that has created an unusually favorable setup. Prices increased without requiring a comparable jump in capital spending, and margins improved quickly. The article later links the rally in memory-related stocks to that same profit expansion.
Capex from Amazon, Google, Microsoft and Meta sits at the center of the demand shock
The article traces the root of the price surge to demand far outstripping supply, and then traces that demand back to investment by hyperscale data center operators.
It names Amazon, Google, Microsoft and Meta as the four companies at the center of the buildout. Their combined capital expenditures were just $21 billion in 2015. After OpenAI released ChatGPT in November 2022, the article says, generative AI enthusiasm pushed that investment curve sharply higher.
Yunokami cites projections showing the group’s capex rising to $355 billion in 2025 and $755 billion in 2026. That is about 36 times the 2015 level over a little more than a decade.
The article adds that $755 billion exceeds JPY 120 trillion and says four companies alone are preparing to spend, in one year, an amount comparable to Japan’s national general-account budget on data centers and AI infrastructure.
AI data centers are absorbing GPUs, HBM and SSD supply like a “black hole”
Yunokami describes AI data centers as a black hole for semiconductor demand. As hyperscalers race to build AI capacity, GPUs used for training and inference, HBM attached to those GPUs, and large-capacity SSDs built with NAND flash are all being pulled into that system one after another, he writes.
Memory makers have responded by prioritizing higher-margin HBM and high-performance DRAM and NAND aimed at data centers. The article presents that as a rational business choice. But it also means production capacity has been shifted toward AI products, leaving much less available for other applications.
The industries taking the hardest hit are PCs, smartphones and game consoles. According to the article, DRAM and NAND supply for digital consumer electronics has become severely tight, to the point of being completely inadequate.
With supply limited and demand concentrated in AI infrastructure, competition for remaining output has intensified and prices have moved higher again. The article says PC and smartphone manufacturers have already begun sounding alarms, saying they cannot secure the memory they need and are being forced to pass higher procurement costs on to product prices.
The author’s 2023 forecast for 2032 was overtaken years early
The piece also revisits a forecast Yunokami made in 2023 for the global semiconductor market through 2032. At the time, he combined what he called the PC effect, internet effect, smartphone effect and the coming AI semiconductor effect, and projected that the market would roughly double every decade, reaching about $1.2 trillion in 2032.
He now says that forecast turned out to be completely wrong — not because it was too optimistic, but because it was too conservative.
Based on WSTS’s spring 2026 outlook, the article says the global semiconductor market reached $630.5 billion in 2024, is expected to hit $795.6 billion in 2025, and is forecast to jump to $1.5112 trillion in 2026. In 2027, it is projected to rise further to $1.9137 trillion, close to $2 trillion.
In other words, the level he had once assigned to 2032 was surpassed in 2026. The article argues that the issue was not excessive optimism in the original forecast, but the extent to which the AI boom overturned established industry assumptions.
Most of the industry’s growth is concentrated in memory and logic
Yunokami says the current expansion in the broader semiconductor market is being driven mainly by two categories: memory, including DRAM and NAND, and logic, including GPUs.
He writes that memory could exceed $1 trillion in 2027, while logic could move above $500 billion. Analog and micro semiconductor segments, by contrast, have stayed largely flat.
That leaves the industry with an uneven growth structure. Rather than broad-based strength across all categories, the market is being carried by two AI-linked segments growing at an exceptional pace.
History shows memory has never managed more than five straight years of positive growth
The article’s key question is how long this AI-linked boom can last.
To answer that, Yunokami looks back over roughly 35 years of annual memory-market growth, from 1991 onward. He revisits a series of past cycles, including the boom after the launch of Windows 95 in 1995, the 2000 IT bubble and its collapse, the 2017-2018 memory bubble, the shock of Lehman Brothers in 2008, and the pandemic-era boom that began in 2020.
His main takeaway is not the height of any single peak, but the duration of each positive-growth phase. Across that full period, he says, memory has never posted more than five consecutive years of annual positive growth. There has never been a six- or seven-year positive streak.
He attributes that pattern to the silicon cycle: demand surges, prices rise, companies increase equipment investment, supply eventually overshoots, and prices collapse. In his view, memory’s nature as a product makes that sequence difficult to escape.
By that historical pattern, the current boom could end by 2028 at the latest
The article says the current AI wave began lifting the memory market from the 2023 trough and moved into a strong expansion phase in 2024. If the historical pattern still holds, Yunokami argues, the upcycle should end by 2028 at the latest and could peak as early as 2027.
He acknowledges the common argument that this time is different, or that AI is uniquely powerful, but says every bubble has generated similar claims. For that reason, he does not treat those arguments as grounds to dismiss the cycle.
“The higher the mountain, the deeper the valley”
The article closes with another rule Yunokami says the semiconductor industry keeps repeating: the higher the peak, the deeper the decline that follows.
He points to the IT bubble, when annual growth climbed above 50% in 2000 before plunging to -49.5% in 2001. He also points to the 2017-2018 memory bubble, which topped 60% before dropping 33% in 2019. In his reading, larger booms have generally been followed by steeper and longer downturns.
Applied to the present cycle, the implication is severe. A 285% annual growth rate is well above the peaks of prior bubbles. By the logic laid out in the article, the downturn that follows this boom could be deeper and harsher than anything the industry has experienced before. Yunokami says companies should be preparing now for a serious recession beginning around 2027 to 2028.
Memory stocks have surged, but the author urges companies to prepare for the reversal
The article also notes that memory manufacturers’ stock prices have risen broadly, which Yunokami says is no surprise given the increase in profits tied to higher pricing.
As a symbolic example, he cites Kioxia. The company’s stock rally, he writes, has left 600 investors sitting on unrealized gains of more than JPY 1 billion from their holdings, effectively creating a group of “billionaires.” The wider industry, he says, is now celebrating at a scale not seen before.
Still, his conclusion is a warning rather than a celebration. Stock prices reflect expectations for the future, and when those expectations become too elevated, even a small miss against them can trigger a rapid collapse.
That is why he ends by calling on companies enjoying the boom to make practical preparations for the downturn while conditions are still favorable. In his framework, memory has never sustained positive growth indefinitely, and the bigger the peak, the worse the adjustment that follows.