On Monday morning Micron Technology came within a few dollars of trading above $1,000 a share for the first time in its history. The stock added roughly three percent before the open, SK Hynix gained more than four percent, SanDisk about six. Nasdaq 100 futures were up half a percent, S&P 500 futures a tenth. Measured against the calendar, it was an odd way to start a week: no memory maker was scheduled to report, no central bank met, no statistical agency released anything.
What moved the tape was a report about a company whose shares nobody can buy. Reuters said on Friday that Anthropic is projecting annual revenue of roughly $190 billion to $200 billion by 2028, and that bankers are discussing an eventual listing at up to a $2 trillion valuation. For scale: the annualized run rate the company disclosed in May was $47 billion, and at the end of 2025 it was about $9 billion. For the second quarter of 2026 it expects at least $10.9 billion in revenue and its first quarterly operating profit, at $559 million.
The distance between that headline and the share price of a memory maker in Boise, Idaho, is considerable. It is worth walking, because it exposes the real story of this summer — one that has less to do with artificial intelligence than with how wafer area gets divided up.
Five Links and Not One Contract
The chain behind Monday’s move has five links. Anthropic’s 2028 revenue forecast justifies the compute needed to produce it. That compute justifies the capital budgets of the cloud providers. Those budgets become accelerator orders, and every accelerator needs high-bandwidth memory. Because that memory is fabricated on the same lines as the DRAM in servers, laptops and phones, the chain ends at the price of a memory module.
Not one of those links is a contract. They are inferences, each plausible on its own and collectively a bet that the chain holds. What deserves attention is therefore not the three percent, but the fact that the most heavily priced-in input in a memory stock is now the revenue plan of a private laboratory. The banks valuing Anthropic are, according to the report, applying revenue multiples to forecasts two years out — Palantir currently trades at roughly 53 times expected 2026 revenue, SpaceX and Cloudflare at about 41.6 times. Those multiples travel through capital budgets into the valuation of companies that operate factories.
The second half of the Anthropic report matters more for investors and was barely discussed on Monday. A first quarterly operating profit of $559 million is the first hard counter-argument to the claim that this entire investment wave is financed out of losses. Anyone arguing the circularity case now has a number to explain.
The Numbers Behind the Boom: an 84.9 Percent Gross Margin
What is happening in memory income statements has few precedents in semiconductor history. Micron reported fiscal third-quarter revenue of $41.46 billion — up 346 percent year over year and 74 percent sequentially, its fifth consecutive record quarter. Non-GAAP gross margin came in at 84.9 percent and non-GAAP earnings at $25.11 per share. Operating cash flow reached $25.39 billion; against $7.1 billion of capital expenditure that left $18.3 billion of free cash flow in a single quarter. Guidance for the current quarter is $50 billion of revenue, a gross margin near 86 percent and $31.00 per share.
SanDisk, the pure-play NAND business since its separation from Western Digital, closed fiscal 2026 with $20.25 billion of revenue, up 175 percent. Datacenter, a rounding error two years ago, contributed $5.15 billion — up 437 percent. Fourth-quarter earnings were $39.25 per share on $8.97 billion of revenue against consensus of $34.59 and $8.42 billion. First-quarter guidance is $10.3 billion to $10.8 billion, underpinned by $93.9 billion of long-term agreements.
And still: on August 5, the day of those record numbers, SanDisk fell 5.3 percent to $1,351.76 and dropped another seven percent after hours. Only in the following week did it recover, gaining 35.4 percent to close Friday at $1,641.11 — still around 30 percent below its 52-week high of $2,354.39. Micron closed Friday at $971.66, a fifth below its record close of $1,213.37 on June 25. Record earnings alongside falling share prices is not a contradiction. It is a statement about the calendar.
The Arithmetic of Allocation: 30 Percent of the Area for 13 Percent of the Bits
The defining sentence of this cycle is not in a financial statement but in an allocation table. TrendForce estimates that high-bandwidth memory accounted for about 18 percent of the three big suppliers’ total DRAM wafer input at the end of 2025, will take 22 percent by the end of 2026 and roughly 30 percent by the end of 2027. Its share of shipped bits in those same years is eight, nine and thirteen percent.
Thirty percent of the area for thirteen percent of the bits: that is the entire shortage in one ratio. A gigabit of high-bandwidth memory consumes roughly three times the wafer area of a gigabit of DDR5, because the dies are stacked, through-silicon-connected and assembled with meaningful yield loss. Every bit that goes into an accelerator therefore costs the market about three bits that never reach a laptop, a phone or a general-purpose server. Wafer capacity dedicated to high-bandwidth memory is growing roughly 29 percent this year; commodity DRAM capacity grows about ten.
That is the difference between a scarcity and an allocation. A scarcity happens when demand outruns supply. An allocation happens when three companies decide which product a given area of silicon becomes — and that decision is not made in groundbreaking ceremonies, it is made quarterly, in price lists. Micron’s Sanjay Mehrotra put it plainly: over the medium term his company can serve only 50 percent to two-thirds of certain customers’ demand. Contract prices for conventional DRAM rose 90 to 95 percent quarter over quarter in the first quarter of 2026 and another 58 to 63 percent in the second. Industry revenue rose 81 percent in the first quarter alone.
Who Pays for It: 17 Percent on a Laptop
Every allocation has an offsetting entry, and it does not appear in a supplier’s accounts. Gartner expects combined DRAM and solid-state drive prices to end 2026 about 130 percent above 2025 levels. That translates into personal computers costing 17 percent more and smartphones 13 percent more. Volumes move accordingly: PC shipments down 10.4 percent, smartphones down 8.4 percent, the steepest contraction in device shipments in more than a decade. Lenovo, Dell, HP, Acer and Asus have all flagged price increases in the 15 to 20 percent range.
The distribution inside that decline is the interesting part. The surcharge hits entry-level devices far harder in percentage terms than premium models, because memory is a larger share of the bill of materials and there is no margin to absorb it. Gartner expects basic smartphone buyers to leave the market five times faster than premium buyers this year. IDC calculates that the average selling price of a smartphone rises 14 percent to a record $523.
This makes precise something that is usually asserted only as a mood: part of this datacenter build-out is being financed by a transfer from buyers of cheap devices to three memory manufacturers. Not metaphorically — through the bill of materials. Micron’s 84.9 percent gross margin and the 17 percent added to a laptop price are two sides of the same entry.
Why the Supply Answer Does Not Arrive Until 2029
The textbook response to high prices is new capacity, and it is already under way — on a timescale that no longer matters for this cycle. SK Hynix announced roughly $38 billion of new plant investment in early August; its board approved 54.3 trillion won for two fabs, 35.2 trillion for Y2 in Yongin and 19.1 trillion for M17 in Cheongju. Ground is to be broken in July 2027 and the first cleanroom is scheduled to open in June 2029. The Yongin cluster is budgeted at 120 trillion won and targets one million wafers a month by 2030.
The money approved today produces its first bits in 2029. SK Hynix already lifted first-half capital spending 72.7 percent to 18.33 trillion won, Micron is spending around $27 billion this fiscal year and guides to more than $10 billion in every quarter of the next, and Samsung has resumed construction of P5 in Pyeongtaek. Even so, SK group chairman Chey Tae-won considers a supply shortfall of more than 20 percent likely to persist through 2030. Waiting for the capex cycle to fix this market means waiting for an event that sits on the far side of the cycle.
The Self-Correction Runs Through the Wafer, Not the Fab
Which is why the most interesting figure of this summer went almost unnoticed. Since the first quarter of 2026, a wafer processed into 64-gigabyte DDR5 server modules has generated more revenue than a wafer processed into high-bandwidth memory. The profitability of the prestige product has fallen below that of the commodity one.
That, and not construction, is this market’s actual self-correcting mechanism. When conventional memory earns more per square millimeter, suppliers shift area back — not out of concern for laptop buyers, but because the ranking of products on the same wafer has flipped. A shift like that takes a quarter, not three years. It is precisely why TrendForce now models only 13 to 18 percent server DRAM contract price growth for the third quarter, against 90 to 95 percent in the first. At the same time, high-bandwidth memory contract prices are expected to rise several-fold in 2027, which would pull the allocation back the other way. So the market does correct itself — but through relative prices between two products rather than through volume.
The Stocks Involved, and the Single-Digit Multiple Trap
The most instructive way to hold this cycle in an American portfolio is to separate the three exposures it actually contains. The first is price: Micron and SanDisk earn on dollars per bit, which is why their earnings look like a different company’s from one quarter to the next. The second is substitution: as flash gets expensive, nearline hard drives absorb the overflow, which is why Western Digital and Seagate have participated in this move without touching a DRAM fab. The third is pass-through: Dell, HP and the PC brands are on the paying side of this trade, taking a component cost increase into a price-sensitive market, which is the opposite position to the one their suppliers hold.
Then there is the trap. After a record quarter Micron trades at roughly six times expected earnings for the next twelve months, while its trailing multiple sits near 21. In a cyclical business, a single-digit multiple on peak earnings is not a bargain signal; it is a probability statement that some of those earnings will not recur. How much is anyone’s guess: 46 analysts polled by S&P Global average a $1,502 price target, with a low of $361 and a high of $2,200. A six-fold spread is not a disagreement about the company, it is a disagreement about the date of the turn.
That has a practical consequence for anyone whose exposure is passive. Semiconductor index funds and the technology weighting inside a typical retirement account now carry memory positions whose earnings power was multiplied by a price move, not by a product. Nobody chose that exposure; it arrived by weighting. The next scheduled test is Nvidia’s report the week after next, followed by the Federal Reserve’s Jackson Hole meeting — the first tells you whether accelerator demand is still accelerating, the second whether the discount rate applied to it is about to move.
What Argues Against This Reading
The weightiest objection comes from Citigroup. Analyst Atif Malik cut his Micron target to $1,150 from $1,400 and dates the cycle peak to the second quarter of 2027: prices keep rising but more slowly, and the biggest long-term risk is growing Chinese capacity. That deserves respect — ChangXin Memory is now the world’s fourth-largest DRAM maker with roughly eight percent share, debuted on Shanghai’s STAR Market in July with a 466 percent first-day gain, and aims to supply about 17 percent of global DRAM by 2028.
Yet the same case shows the limit of the argument. The Chinese entrant supposedly about to break pricing is sold out through 2027 and, according to available figures, prices comparable 64-gigabyte DDR5 server modules above Samsung’s roughly $1,240. New capacity depresses prices only once it exceeds demand; while shortage lasts, it pays the imitators too. The second objection is demand destruction, already visible: ten percent fewer PCs and eight percent fewer phones is not a rounding error, it is a shrinking market for the bits that new fabs will produce. The third stands at the top of this article — a chain of five inferences without a single contract in it.
Three Measures That Reveal the Calendar
Anyone watching this market should track the second derivative of prices rather than prices themselves. Whether contract prices rise 90 percent or 15 percent in a quarter matters far less to next year’s earnings than the fact that the acceleration has faded. Cycles do not turn at the peak of prices; they turn at the peak of the rate of change.
The second measure is the ranking on the wafer. As long as a DDR5 server module earns more per unit of area than high-bandwidth memory, the allocation mechanism is working against the shortage. If that flips back in 2027 because high-bandwidth prices rise several-fold, commodity memory tightens again — and so does the surcharge on devices.
The third measure is PC and smartphone unit volumes. They are the honest gauge of how much of this boom is price and how much is genuine demand. If shipments fall faster than forecast while prices are still rising, the cycle is closer to its end than any gross margin suggests. If they fall as forecast and prices keep climbing, the uncomfortable version of this story holds: that a datacenter in Virginia and an entry-level smartphone in Manila compete for the same wafer area — and the allocation has already been decided.
Try TradingView Free for 30 Days
Plus get a $15 discount on your first subscription through this link.
