Micron slipped on Thursday because a private company that has run about 2,000 wafers in its whole life said it can help end the memory shortage. That’s the trade. Kepler Computing came out of stealth this week with a ferroelectric composite material, a 3D scheme that stacks memory on top of logic, and GlobalFoundries as its manufacturing partner. The headlines called it Intel-backed. It is. It’s also AMD-backed and Gates-backed, and GlobalFoundries has money in it too. Intel Capital is one line on a cap table that has raised $468 million.

Kioxia and Sandisk said Thursday they intend to spend roughly ¥5 trillion, more than $31 billion, in Japan through 2032 to expand flash memory production at the Yokkaichi and Kitakami plants. Kioxia said separately that it has begun site preparation for a third fabrication building at Kitakami. Hiroo Ota carried the plan to Prime Minister Sanae Takaichi’s office in person. The wires ran it as a capacity announcement. The two qualifiers attached to it are more informative than the headline number: the spending is contingent on support from the Japanese government, and Fab3 is not expected to begin operations until fiscal 2029.

Nvidia’s quarter was not the problem. Revenue of $96.2 billion, up 106% year over year, earnings of $2.22 per share against $1.05, gross margin at 75.0% versus 72.7%, and a third-quarter guide of $108 billion plus or minus 2% against $57.01 billion a year ago. Colette Kress then told analysts to expect fiscal 2028 revenue growth of 70% while the street was carrying 44%, and added that customer forecasts point to demand roughly doubling, with the guidance reflecting supply constraints rather than orders. The stock rose about 6%. Marvell, Micron, SanDisk and Western Digital did not.

Morgan Stanley’s upgraded CXL forecast is circulating as a Marvell thesis, and the exhibit it comes from does not support the weight being put on it. The chart stacks two lines, memory expansion controllers and switch silicon, and carries the combined market from roughly $50 million in 2025 to about $4.05 billion in 2030. Both halves were raised hard. MXC went from $990 million to $2.1 billion. Switching went from $664 million to $1.9 billion. What did not change is the scale of the thing being doubled. Four billion dollars in 2030 is the entire industry, before Marvell divides it with Astera Labs, Montage, Microchip and Rambus.

Kioxia and SanDisk announced their 9th-generation 2Tb QLC 3D flash memory on August 12, eight days after the same two companies unveiled their 10th-generation QLC technology at the Future of Memory and Storage conference in Santa Clara. The ordering is not a clerical error, and it is the most useful fact in the release. The 10th generation is a node advance: 332 layers, better than 37 Gb/mm², roughly 60 percent more bits per unit area than the 8th generation now in mass production. The 9th generation is not a node advance at all. It bonds an advanced CMOS wafer to the existing memory-array platform, which is to say the same cell array already running in the 8th-generation 2Tb QLC die, and pulls out a 4.8 Gb/s NAND interface, a 33 percent improvement, plus a six-plane architecture, higher read and write bandwidth, and better power efficiency on both paths. Bit density does not move. Hideshi Miyajima said the quiet part in the first quote of the release: the approach delivers high performance while keeping investment costs relatively low. That sentence is the product.

BofA Global Research rebuilt its industry memory model this week on the resumption of SK Hynix coverage, and the headline output is a global DRAM industry sales path of $134bn in 2025 to $573bn in 2026, then $847bn in 2027 and $917bn in 2028. The 2026 figure is 328% year-over-year growth in an industry that grew 52% in 2025 and 86% in 2024. Numbers that size normally signal a modelling error. Here they signal that the entire revenue base repriced inside twelve months.

SanDisk closed its fiscal year with June quarter revenue of $8.965 billion, up 51 percent from March and 372 percent from a year ago. Non-GAAP earnings came in at $39.25 a share against company guidance of $30 to $33. The stock fell after ho urs anyway.

The figure worth sitting with is buried further down the income statement. Cost of revenue for the quarter was $1.383 billion. A year earlier, on revenue of $1.901 billion, it was $1.403 billion. SanDisk sold roughly five times as much product and spent slightly less doing it. Almost every incremental dollar of revenue over the past twelve months fell straight through to gross profit. That is the entire story of this fiscal year, and it explains why the argument about this company has very little to do with how well it is run.

ChangXin Memory Technologies opened at 49.50 yuan against an IPO price of 8.66 and closed the morning session up 531%, carrying a market capitalization near 3.66 trillion yuan and making it the most valuable company listed on the mainland. The headline number is 500%. The number that explains it is 6.73%.

That is the share of post-IPO capital that was unrestricted and tradable on day one — roughly 4.5 billion shares out of a base that runs to nearly 67 billion. Institutional demand for the allocation exceeded 500 times the shares offered; retail subscription ran better than 200 to one. A first trade that added some 2.7 trillion yuan of implied value did so against a sliver of float, before a single wafer of incremental capacity existed or a single customer qualification changed. Price discovery on 6.73% of a company is not price discovery. It is an auction for scarcity, and it resolves when the scarcity does.

The same share of the same company traded at two prices this week, roughly a third apart, and there was nothing anyone could do about it. SK Hynix’s American depositary receipts have run as much as 51% above the Seoul-listed common stock since the July 10 Nasdaq debut, settling around a 33% premium midweek. That gap is not a signal about DRAM pricing, HBM allocation, or Nvidia’s qualification schedule. It is a plumbing failure, and the plumbing is the point.

Micron fell roughly 8% on Wednesday, and the tape assigned two culprits: ChangXin Memory Technologies pricing an ~$8.5 billion IPO on Shanghai’s STAR Market, and reports that Washington may impose new export controls on high-bandwidth memory. Both are real. Neither changed the memory market. The stock repriced; the physical supply, demand, and pricing that define the actual market did not.

The distinction matters because it is the whole argument. The memory market is wafers, contracts, and average selling prices. An IPO is a financing event. Micron shed about $94 billion in market value against a rival worth roughly $85 billion that has not shipped one incremental chip and will not for years. No new supply hit the market because CXMT raised yuan in Shanghai. What moved was sentiment about a future supply path, applied to a name that had run 245% year to date and was overdue for a correction. That is a repricing of positioning, not a change in fundamentals.

Bloomberg reported on Tuesday that Samsung Electronics is in the early stages of exploring an offering of American depositary receipts, having held preliminary discussions with banks without reaching a decision. Sources characterised the process as a review rather than a plan — no bank mandated, no commitment made, and a real possibility that nothing comes of it.

Samsung denied it. A company spokesperson said flatly that Samsung Electronics is not reviewing the possibility of issuing American depositary receipts.

The consensus bull case for memory is that the market is years away from saturation. On the demand side, that is almost certainly correct. It is also the wrong frame, and investors who anchor on it will be looking in the wrong direction when the cycle turns.

Memory downturns have never been caused by demand saturation. They are caused by supply growth outrunning demand growth. Those are different failure modes, and the second one can fire while demand is still compounding at twenty percent.

Every memory forecast published in the last twelve months has been wrong in the same direction: too low. IDC, TrendForce, and Bank of America have each revised DRAM and NAND demand estimates upward multiple times since early 2025, and the pattern is not noise — it is a structural failure of forecasting methodology colliding with a demand curve that refuses to plateau.

The Forecasts Keep Missing In One Direction

The current numbers are already staggering. IDC now expects 2026 DRAM supply growth of only 16% year-on-year, with NAND supply growth at just 17%, both well below the 20-30% historical norms that defined the post-2018 memory market. HBM demand alone is projected to grow 70% year-over-year in 2026, with HBM consuming 23% of total DRAM wafer output, up from 19% the year before. Bank of America forecasts DRAM revenue surging 51% year-over-year and NAND 45%, with ASPs rising 33% and 26% respectively.

Samsung and SK Hynix unveiled a combined roughly $1.3 trillion (2,000 trillion won) decade-long investment plan for new fabs, AI data centers, and chip cluster development. Both stocks fell anyway — Samsung down over 5%, SK Hynix down over 3% on the announcement day, following an even sharper 9%+ plunge earlier in the week. The knee-jerk read: investors think the spending is reckless, a repeat of the 2018-2019 memory bust, or proof the AI trade is cracking.

CXL was sold as a capacity story: extend the memory pool past the DIMM slots soldered to the motherboard. Marvell’s argument with Structera is sharper than that. The pool itself is half-empty. The data sitting in DRAM is compressible, almost no CXL controller touches it, and Structera does — in dedicated silicon, at line rate, invisible to the host.

The number circulating is 3.64x, the top of the range Marvell cites for mixed real-world data types, which it claims match or closely approach what host-side LZ4 achieves in software. Field reporting has been more conservative; ServeTheHome quoted Marvell putting practical ratios at 1.8x to 2x. Both numbers point the same way. Even a flat 2:1 halves the effective cost per gigabyte of a memory pool, and memory is the single largest line item in that pool.

The Wall Street Journal headline frames the memory shortage as a problem to be solved. It isn’t. The more accurate reading of the supply picture is that the crunch is the predictable output of a fixed production base being reallocated toward AI, and there is no near-term lever — industrial or political — that changes that math before 2027. For the three companies that own the supply, that is not a crisis. It is the most durable pricing-power setup the industry has seen in a generation.

The memory industry spent thirty years teaching investors one lesson: never believe “this time is different.” Boom, over-invest, glut, collapse. Price the top early, because the top always comes. That instinct is now the most expensive mistake in semiconductors. The DRAM and NAND supercycle that began in 2024 is not late-cycle. It is early. And the mechanism that has ended every prior memory cycle has been disabled.

The demand is structural, not cyclical

Start with the numbers, because they are not subtle. IDC puts DRAM revenue at $418.6 billion in 2026, up roughly 177 percent year over year, with total memory rising from $226 billion in 2025 to $594.7 billion in 2026 and $790.4 billion in 2027. Bank of America frames the period as a supercycle on the scale of the 1990s boom, with DRAM revenue up 51 percent and NAND up 45 percent. Contract prices through early 2026 rose 90 to 95 percent quarter over quarter. DDR5 spot prices quadrupled from September 2025. Supplier inventories sit at two to four weeks.

The instinct after a move like this is to call the top. SanDisk has gained more than 4,400% over the past year. Micron has added roughly 810%. Both trade within a few dollars of their 52-week highs. Every rule of thumb says a chart like that is closer to its end than its beginning. The rules of thumb are wrong here, and the reason is structural, not technical.

Apple Just Confirmed the Thesis

This week Tim Cook told the Wall Street Journal that price increases across Apple’s lineup are unavoidable, and he named memory as the cause. The September iPhone 18 Pro is expected to carry the first higher sticker price, with TechInsights estimating that preserving Apple’s margin would require adding roughly $270 to the starting price. The market read Apple shares as a wash. It read the memory names as a green light.