NextFin News - The semiconductor industry is running a race in which demand has already booked every seat through 2026, yet the market keeps pricing in a crash. SK Hynix has sold out its entire DRAM, HBM and NAND production for next year. Micron says high-bandwidth memory alone will be a $100 billion market by 2028, up from $35 billion in 2025. Samsung is spending a record $73 billion to catch up. And still, the Philadelphia Semiconductor Index has given back roughly a fifth of its gains since its late-June record. The question is not whether chips are in demand - the question is whether this is the rare cycle that does not end the way every memory cycle before it has ended.
The answer, based on the supply-side mechanics now visible across the industry, is that this supercycle is structurally longer and flatter than the boom-bust pattern etched into semiconductor history - but it is not immune to a correction. The binding constraint has shifted from demand to supply: advanced packaging, HBM wafer conversion and sub-2nm capacity cannot be built quickly, and every wafer allocated to HBM is a wafer that cannot produce conventional memory. That crowding-out effect is what separates this cycle from 2017 and 2021. It also means the endgame will likely look less like a cliff and more like a long, gradual normalization - unless the hyperscalers funding this buildout blink first.
The Numbers Behind the Supercycle Claim
The evidence for a structural shift is unusually concrete because it is expressed in capital commitments rather than forecasts. SEMI's 300mm Fab Outlook, released in October 2025, projects global 300mm fab equipment spending of $374 billion from 2026 through 2028 - the first time 300mm spending will top $100 billion in a single year, reaching $107 billion in 2025, then $116 billion in 2026, $120 billion in 2027 and $138 billion in 2028. Logic and micro components lead with $175 billion as foundries race toward sub-2nm, gate-all-around and backside power delivery, with 1.4nm entering volume production by 2028-2029; memory follows with $136 billion, driven by HBM bandwidth needs for training and expanding NAND requirements for inference-generated content.
At the company level, the commitments are just as large. Samsung Electronics plans to spend more than 110 trillion won, about $73.3 billion, on chip capacity expansion and research in 2026, a 22 percent increase, in an effort to retake the AI chip lead from SK Hynix. TSMC raised its 2026 capital expenditure forecast to between $60 billion and $64 billion, with 70 to 80 percent directed at advanced process technologies and 10 to 20 percent at advanced packaging, where CoWoS capacity is expanding by more than 80 percent annually through 2027 and still cannot meet demand. UBS estimates global AI capital expenditure will grow 60 percent in 2025 to $360 billion and a further 33 percent in 2026 to $480 billion, with sovereign and enterprise investors joining hyperscalers in the buildout.
The demand side validates the spending. NVIDIA reported second-quarter fiscal 2026 revenue of $46.7 billion, up 56 percent year over year, with $41.1 billion coming from its data center segment. About half of that data center revenue came from large cloud service providers. NVIDIA's chief financial officer, Colette Kress, told investors the company expects revenue to grow approximately 70 percent in the next fiscal year, though it remains constrained by memory supply bottlenecks. In other words, the customer with the biggest order book is itself supply-constrained.
"DRAM inventories have dropped to extremely low levels, and for DDR5 products, newly produced chips must be shipped to customers immediately to meet demand," said Kim of SK Hynix, after the company reported a record quarterly operating profit of 11.4 trillion won, up 62 percent from a year earlier.
Why This Cycle Is Different - and Why That Matters
Every memory cycle in the modern era has followed the same script: a demand-driven boom lasting four to seven quarters, followed by an oversupply-driven bust lasting four to eight quarters, with revenue declines of 25 to 40 percent and margin compression from peaks above 50 percent to the low 20s or worse. In the 2017-18 cycle, DDR4 retail prices doubled into early 2018, industry inventories fell to three to four weeks against a normal eight-week average, and the market peaked - only for DRAM prices to crash 60 to 70 percent in 2019 as capacity, ordered during the boom, arrived roughly 12 months later. Micron's stock fell 54 percent peak to trough. The 2021 cycle peaked around month 20 and then collapsed. The pattern is so reliable that it has its own leading indicators: spot DRAM prices top three to six months before revenue, and inventories balloon to 20 to 30 weeks before a downturn.
Three things are different now, and each one attacks a different link in that chain.
First, the nature of the demand. AI-driven demand is tied to long-lived infrastructure programs - data centers that operate for a decade, not consumer devices replaced every two years. That gives suppliers confidence to add capacity in controlled stages rather than the blind overspend that caused the 2019 glut. Micron pulled its $100 billion HBM market forecast forward two years, from 2030 to 2028, implying a compound annual growth rate near 40 percent. SK Hynix expects DRAM shipments to grow more than 20 percent year over year in 2026, and HBM products - which make up about 20 percent of its DRAM shipments - contribute roughly half of its operating profit. HBM's wafer intensity is about four times that of conventional DRAM, so industry estimates put AI's share of global DRAM wafer capacity at nearly 20 percent in 2026 on an adjusted basis. That is a structural reallocation of the industry's capital stock, not a seasonal inventory restock.
Second, the supply response is physically constrained. Total DRAM wafer starts are growing only about 6 to 8 percent year over year exiting 2026, by industry estimates, and even the most aggressive investor cannot add meaningful new supply before 2027 or 2028 at the earliest. Every wafer allocated to HBM is a wafer that does not produce conventional DDR5, creating a crowding-out effect that pushes standard memory prices up to meet HBM. Advanced packaging - the CoWoS bottleneck at TSMC - cannot be spun up overnight; it is expanding more than 80 percent a year and still cannot satisfy demand. This is the mechanism that makes the cycle longer: supply cannot respond elastically to price, so prices stay elevated longer than the historical pattern would suggest.
Third, the financial structure is stronger. In 2000, the semiconductor boom was financed with debt and equity issuance from companies with little cash flow. Today's flagship chipmakers earn tens of billions in free cash flow, and the AI buildout is funded largely from operating profits. An 85 percent crash - the kind that defined 2000 - requires valuations and financing structures that 2026 does not have.
The Counter-Thesis: Why the Market Is Pricing a Crash Anyway
The bear case is not a strawman, and it deserves its weight. A compilation of semiconductor bear markets going back to 1995 records the index down roughly 20 percent from its late-June 2026 record in barely three weeks, after the sector gained more than 80 percent in the first half of the year - a pace of ascent not sustained since 1999-2000. The triggers stack up like a checklist from past cycles. Reports suggested Meta Platforms plans to resell excess AI computing capacity, crystallizing the question of whether Big Tech has overbuilt AI infrastructure. The first large wave of custom AI accelerators from Amazon, OpenAI and others is shipping against NVIDIA's near-monopoly, threatening margin compression at the top of the stack. A disappointing Samsung quarter reset memory expectations. And TSMC's decision to raise its 2026 capex to between $60 billion and $64 billion reads, to bears, as the classic capex-bust trigger that preceded the 2000 collapse - the index roughly doubled in the year into March 2000 and rose about 180 percent in the twelve months into June 2026, with record retail and momentum participation in both.
The strongest version of the bear argument is simple: capex is a leading indicator of supply, and supply is a lagging indicator of price. When Samsung, SK Hynix, Micron and TSMC collectively commit hundreds of billions of dollars to new capacity, that capacity arrives with a lag, and when it arrives, it arrives all at once. The 2019 DRAM crash was not caused by a collapse in demand; it was caused by capacity ordered in 2017 showing up in 2019. The same physics applies now. If hyperscaler AI spending slows - because the return on AI infrastructure fails to meet the hurdle rates that justify $480 billion in annual capex - the industry will be left with too much capacity chasing too little demand, and the historical pattern will reassert itself.
The counter-thesis has one quantifiable point of failure. The bust pattern requires supply to arrive faster than demand. The falsifying signal is DRAM inventory weeks: if inventory stays below 10 weeks through the end of 2027 - well below the 20 to 30 weeks that preceded every past downturn - while HBM remains sold out and wafer starts grow no faster than 10 percent annually, the oversupply mechanism that has ended every prior cycle simply cannot fire, and the bear case collapses. Conversely, if DRAM inventory weeks climb above 15 for two consecutive quarters, or if HBM order books begin to slip into 2027, the structural thesis is wrong and the historical pattern is back in force.
What Comes Next: Scenarios by Time Horizon
Short term - the next two to four quarters - the picture is the clearest. Production is sold out through 2026 at all three major memory producers. HBM4 is entering volume production, with SK Hynix beginning shipments in the fourth quarter and Samsung announcing mass production of commercial HBM4 in February. Memory industry profits are not expected to peak before the fourth quarter of 2027, according to industry analysis. The near-term risk is not fundamentals; it is valuation and positioning. The roughly 20 percent drawdown in the semiconductor index is a reminder that stocks discount turns before they show up in the numbers - Micron's stock peaked in May 2018, roughly two quarters before its fundamental peak in the fourth quarter of calendar 2018.
Medium term - 2027 to 2028 - is where the cycle gets tested. New fabs come online: Micron's first Idaho fab is expected to produce its first wafers in mid-calendar 2027, pulled in from the second half of 2027, and a second Idaho fab will begin construction in 2026 and be operational by the end of 2028. Micron is also preparing a $9.6 billion HBM facility in Hiroshima, Japan, with construction expected to begin in May 2026 and shipments around 2028. If demand grows as forecast, this new supply is absorbed and the upcycle extends. If hyperscaler capex growth decelerates from 33 percent toward single digits while capacity additions accelerate, the margin-compression phase begins on schedule.
Long term - beyond 2028 - the structural question resolves around whether AI compute demand proves to be a permanent step-up in the industry's growth rate or a one-time infrastructure buildout. The base case is that AI becomes a durable, compounding demand layer - inference at the edge, enterprise deployment and sovereign AI programs broaden the customer base beyond a handful of hyperscalers. The downside case is a 2000-style overbuild if the economics of AI services fail to justify the capital deployed. The upside case is that the $100 billion HBM market of 2028 becomes the floor, not the ceiling, as memory becomes the binding constraint on AI performance and pricing.
The beneficiaries and the exposed are not symmetric. Memory producers with HBM exposure - SK Hynix, Micron and Samsung - hold the pricing power as long as the crowding-out mechanism holds. Equipment makers benefit from the $374 billion fab spend regardless of which memory vendor wins. Foundries and advanced-packaging leaders, led by TSMC, sit at the narrowest point of the supply bottleneck and can price accordingly. The exposed are companies betting on a smooth, uninterrupted ramp: any supplier whose business plan assumes demand growing 30 percent a year without a single inventory correction is pricing in a miracle.
So is the chip supercycle here to stay? The structural leg - AI as a permanent, capital-intensive demand layer with supply that cannot respond elastically - is real and will outlast the cyclical leg. But "supercycle" does not mean "no correction." It means the correction, when it comes, will be shallower and the recovery faster than history suggests. The market is not wrong to fear the cycle; it is wrong to assume this cycle must end the way the last one did.
Explore more exclusive insights at nextfin.ai.
