NextFin News - Memory chip prices are rising at a pace the semiconductor industry has not seen since the 2017-2018 shortage, and the executive who leads the chipmaking supply chain's main industry group says the surge is being driven by a structural shift in how AI infrastructure consumes chips. Ajit Manocha, president and chief executive officer of SEMI, discussed memory pricing and the semiconductor supercycle in a Bloomberg Television interview on September 2, 2026. His comments land as the data confirms the unusual shape of this cycle: record pricing, record investment, and barely-growing capacity, all at once.
The backdrop is stark. Gartner expects global semiconductor revenue to reach $1.6 trillion in 2026, up 92% from $809 billion in 2025, with memory alone forecast at $837 billion this year and more than $1 trillion in 2027. Memory, which accounted for 27% of semiconductor revenue in 2025, is projected to make up 54% in 2026. That concentration is the heart of the story: the industry is not just growing — it is being rewired around AI data centers.
The Price Data Does Not Look Like a Normal Upcycle
The contract-price tape is unambiguous. TrendForce reported that conventional DRAM contract prices jumped 90% to 95% quarter-over-quarter in the first quarter of 2026, followed by a 58% to 63% increase in the second quarter. For the third quarter, the pace is expected to moderate but remain positive: server DRAM contract prices are projected to rise 13% to 18% quarter-over-quarter, with NAND Flash up 10% to 15%. Three consecutive quarters of double-digit increases is not how a textbook memory cycle behaves.
The cause is a capacity squeeze with a specific name: high-bandwidth memory. HBM, the vertically stacked DRAM used in AI accelerators, consumes far more wafer area per unit of output than conventional DRAM — industry estimates put the multiplier at three to four times. As Samsung Electronics, SK Hynix, and Micron Technology — which together account for more than 90% of DRAM production — reallocate their best capacity toward HBM and server DRAM, the supply of memory for PCs, smartphones, and automotive applications shrinks even as total industry output grows. TrendForce estimates HBM wafer input among the top three suppliers will reach 22% of total DRAM wafer input in 2026, up from 18% at the end of 2025, and 30% by the end of 2027.
Micron has said it expects DRAM and NAND bit shipments to grow about 20% in calendar 2026, yet still told customers it is "disappointed to be unable to meet demand from our customers, across all market segments." That gap — rising output alongside unmet demand — is the signature of a supply-constrained market, not a demand-led boom.
SEMI's equipment forecasts quantify the investment response. In its April 2026 300mm Fab Outlook, the industry association projected that worldwide 300mm fab equipment spending would rise 18% to $133 billion in 2026 and 14% to $151 billion in 2027. Memory is a central beneficiary: 300mm memory equipment investment is expected to surpass $50 billion for the first time in 2026, reaching $52 billion, before climbing another 11% to $57 billion in 2027. DRAM equipment spending alone is forecast at $37 billion in 2026, up 29%, with 3D NAND at $14 billion, up 28%.
Yet that investment is not translating into near-term capacity relief. SEMI's outlook shows 300mm memory capacity reaching 4.1 million wafers per month in 2026 and 4.2 million in 2027 — growth of just 2.4% year-over-year. For a market where AI data-center demand is compounding, a 2.4% capacity response is thin. The combination of record pricing, record investment, and barely-moving capacity is what makes this cycle different from the ones before it.
Why the Transmission Mechanism Has Changed
In a conventional memory upcycle, demand rises, prices climb, manufacturers add capacity, and prices fall back toward marginal cost. The 2017-2018 DRAM shortage, driven by smartphones and cloud servers, followed that script: prices roughly doubled over about 18 months before supply caught up. The 2020-2021 chip crunch, fueled by pandemic demand and crypto mining, lasted 18 to 24 months and then normalized.
This time, the supply response is structurally muted for two reasons. First, HBM does not simply absorb a share of existing capacity; it destroys effective output. Because HBM stacks multiple dies vertically and requires through-silicon-via bonding and advanced packaging, a wafer devoted to HBM yields far fewer sellable gigabytes than a wafer devoted to DDR5. Every percentage point of HBM conversion therefore removes more than a percentage point of conventional DRAM supply. The result is a market where total wafer starts can rise while sellable gigabytes for non-AI applications fall.
Second, producer discipline has replaced share-seeking. After the 2022-2023 downturn, when memory prices collapsed and producers burned cash, the three major manufacturers have prioritized profitability over volume. Market intelligence citing Goldman Sachs estimated the 2026 global DRAM deficit at 4.9% and the NAND deficit at 4.2% — the widest gaps since 2011 — yet capacity additions remain measured. This is not coordination; it is a rational response to a demand curve the producers now believe is permanently higher.
"Strong demand for high bandwidth memory and other advanced memory technologies is reshaping investment priorities across the semiconductor supply chain. As AI infrastructure expands, memory manufacturers are accelerating investments in both capacity and technology migration to support the next wave of data-intensive applications."
That is Manocha, in a statement accompanying SEMI's second-quarter 2026 outlook. The short version: this is a supply-constrained market where the constraint is technological, not just financial. New capacity takes three to five years to build, and the most advanced capacity is increasingly reserved for customers willing to sign long-term agreements.
Cyclical or Structural? The Call
The central judgment: the memory supercycle is a cyclical price wave riding on top of a structural regime shift, and the structural leg is the one that matters for anyone with a horizon beyond the next few quarters.
The cyclical leg is real and will revert. Contract price increases of 90% in a single quarter cannot persist; the third-quarter moderation to 13% to 18% is the first evidence of mean reversion. When the DRAM and NAND capacity commissioned in 2027 and 2028 comes online, and if AI infrastructure spending pauses, memory prices will fall from these peaks. Historical cycles teach that memory is never immune to overshoot, and the summer of 2026 already showed how quickly sentiment can turn: Samsung, SK Hynix, and a basket of memory stocks fell more than 20% from recent highs even as Samsung posted record profit.
But the structural leg does not revert. Three pieces of evidence support that call. First, the demand base has changed composition: Gartner expects the AI data-center ecosystem to grow from 36.5% of semiconductor revenue in 2026 to more than 53% by 2030. That is a permanent shift in where value is created, not a one-time inventory restocking. Second, the pricing mechanism has changed: Samsung and SK Hynix are moving away from one-year contracts with big-tech customers toward long-term agreements spanning three to five years, with some deals including prepayments of 10% to 30% of contract value and minimum price floors. Those structures replace the quarterly spot pricing that used to transmit boom-bust signals through the market. Third, the technology frontier has changed: HBM and advanced packaging have raised the barrier to entry, concentrating supply among three producers and making capacity additions slower and more capital-intensive.
A structural claim requires evidence that history no longer applies. Here it does. The old cycle was driven by consumer electronics — PCs, phones, tablets — whose unit growth has stalled. The new cycle is driven by AI training and inference, where the number of accelerators per data center keeps rising and each accelerator consumes an expanding amount of memory. A data center built in 2024 does not stop needing memory in 2027; it needs more. That is a demand floor that did not exist in previous cycles.
The consumer market is already paying the price for that floor. TrendForce reported in February 2026 that memory's share of a mainstream smartphone bill of materials has surged from a historical 10% to 15% to 30% to 40%, and it has cut its 2026 global smartphone production forecast to a 10% year-over-year decline, to approximately 1.135 billion units. When a component shortage starts destroying demand in the end market, it is no longer just a supply story.
The Second-Order Question: Who Actually Captures the Value?
The market has already priced the first-order conclusion — memory is scarce, so memory stocks rise. The second-order question is whether the capital being spent to relieve that scarcity will earn its keep, and the answer is not uniform across the supply chain.
Consider the capex math. SEMI's $374 billion 300mm fab equipment forecast for 2026-2028, and the $133 billion and $151 billion annual figures, represent the largest peacetime investment wave in semiconductor history. But equipment spending does not equal profit. The producers adding the most capacity — Samsung, SK Hynix, Micron — face a race between falling unit costs from new process nodes and falling prices when that capacity arrives. If memory prices normalize in 2027-2028 just as new fabs ramp, the return on that capital could disappoint even if revenue hits record levels.
The more durable value capture sits one layer up the chain: in the equipment and advanced-packaging suppliers that sell the tools regardless of which memory technology wins, and in the producers with the most advanced HBM capability. HBM4 supply negotiations for 2027 launched in the second quarter of 2026, and TrendForce expects suppliers to push through substantial contract price hikes reflecting the acute supply-demand imbalance and rising manufacturing complexity. The company that can ship HBM4 at volume first captures pricing power that a commodity DRAM producer cannot.
There is also a geographic second-order effect. Regional 300mm investment from 2026 to 2028 is forecast at $94 billion in China, $86 billion in Korea, $75 billion in Taiwan, and $60 billion in the Americas. That dispersion is not purely economic; it is policy-driven, the result of subsidy programs and supply-chain localization mandates. Policy-driven capacity is slower to build and less responsive to price signals — which reinforces structural tightness even as it raises long-run capacity.
The Counter-Thesis: This Is Still a Cycle, and Cycles End
The strongest case against the structural call is simple: memory has always been cyclical, and the current tightness is being managed by the same three producers who will flood the market the moment pricing looks attractive enough. Skeptics point to the 2017-2018 shortage and the 2020-2021 crunch as proof that every memory upcycle ends in overshoot. They also note that AI infrastructure spending is itself cyclical — hyperscalers are making multi-hundred-billion-dollar capital commitments that cannot be sustained indefinitely, and any pause in data-center buildout would leave memory producers with excess capacity and falling prices.
That argument is correct as far as it goes, and it is the right lens for the next 12 to 18 months. The 13% to 18% third-quarter DRAM increase already shows the peak rate of price growth is behind the market. If AI accelerator orders slow, or if HBM yield improvements release more effective capacity than expected, the cyclical leg will reassert itself quickly. Memory stocks can give back a large share of their gains on a single shift in sentiment.
But the counter-thesis conflates price levels with industry structure. Even if memory prices fall 30% from their 2026 peaks, the floor will be higher than in previous cycles because the demand base has shifted toward AI data centers under multi-year contracts. The bust, if it comes, will be shallower and shorter than the 2022-2023 downturn. That is the structural claim in its weakest, most defensible form: not that prices only go up, but that the cycle's amplitude has been compressed.
Two falsifying signals would break that judgment. First, if 300mm memory capacity grows faster than 10% year-over-year in 2027 while HBM wafer input stays below 30% of total DRAM input, the structural-tightness thesis is wrong — it would mean capacity is responding elastically and the old boom-bust mechanics are intact. Second, if hyperscaler capital expenditure growth falls below 10% year-over-year for two consecutive quarters, the demand-floor assumption breaks.
What Comes Next
The practical implication is a split verdict across time horizons. In the short term, memory prices will moderate from their parabolic first-half pace; the third-quarter 13% to 18% DRAM increase is the transition. That is negative for the momentum trade in memory stocks but not negative for the industry's underlying earnings power. Over the medium term, through 2027 and 2028, the producers with the most advanced HBM capability and the equipment suppliers that enable the capacity buildout are the most direct beneficiaries. Over the long term, the structural shift toward AI data centers — projected to absorb more than half of semiconductor revenue by 2030 — means the memory market's center of gravity has moved, and with it the definition of what "normal" pricing looks like.
The exposed parties are the buyers of conventional memory: PC and smartphone OEMs facing steep mobile DRAM price increases, and automotive and industrial customers who sit at the back of the allocation queue. TrendForce expects smartphone makers to adjust production plans and raise end-product prices to protect margins as memory costs climb.
Base case: memory prices grind higher through the rest of 2026 at a moderating pace, capacity additions remain disciplined, and the industry reaches the $1 trillion memory revenue mark in 2027 as Gartner forecasts. Upside case: HBM4 ramps slower than expected, keeping the supply-demand gap wide and extending pricing power into 2028. Downside case: a hyperscaler capex pause or a faster-than-expected HBM yield improvement triggers a 2027 price correction of 25% to 35% from peak levels.
What to watch: the 300mm memory capacity figures in SEMI's next outlook, the 2027 HBM4 contract settlements, and quarterly hyperscaler capital expenditure guidance. Any two of those pointing toward looser supply would be the first real evidence that the supercycle is ending.
The memory market has spent fifty years teaching investors that no shortage lasts forever. The lesson for this cycle is different: the shortage may ease, but the floor has moved. This is not the 2017 cycle with better marketing — it is a market where AI infrastructure has turned memory from a commodity into a strategic input, and commodities do not command strategic pricing for long.
Explore more exclusive insights at nextfin.ai.
