Apple, two hundred and eighty basis points

2.The arithmetic of the compression

What does one lost point of gross margin cost, in dollars? Added together, those two steps give an underlying compression of about 280 basis points between March and September 2026. On a quarterly revenue base of roughly $110 billion, each point of gross margin is approximately $1.1 billion of gross profit. The order of magnitude of the annualised effect therefore runs into the billions.

The segment detail published in the quarterly report locates the pressure precisely, and it is more instructive than the blended margin. The services margin is perfectly stable, at 75.6% against 75.6% a year earlier. The products margin rises from 34.5% to 40.1%, but it falls sequentially since it stood at 38.7% in the March quarter, and it carries the entire memory cost increase (Apple Inc., 2026a). Put differently, the compression is not diffuse: it is concentrated on hardware, that is on roughly 72% of revenue.

The chief executive himself described the memory situation as a « hundred-year flood », speaking of exponential increases, and confirmed price rises already conceded reluctantly on iPad and Mac. The chief financial officer added that he expects to pay still more for memory beyond September (Apple Inc., 2026b). A company that raises its selling prices and warns it will pay more for its components is describing a compression that has not run its course.

3.Why memory, and why now

Why memory, and why now? The cause is external to Apple and thoroughly documented. Contract prices for conventional memory rose 93 to 98% in the first calendar quarter of 2026, then 58 to 63% in the second, then 13 to 18% in the third (TrendForce, 2026). On flash, the second quarter peaked at a rise of 70 to 75%.

The origin is artificial intelligence, but not through device demand: through the reallocation of capacity. High bandwidth memory for servers consumes far more wafer per useful bit than ordinary memory, and its share of wafer starts at the three largest suppliers has moved from about 18% at the end of 2025 towards 30%. Every wafer switched over is withdrawn from the supply destined for phones and computers.

The strategic consequence is harder than the increase itself. Several large data centre buyers have secured multi-year agreements that contractually cap their increases. The rises therefore concentrate on uncovered buyers and on volumes outside contract. Apple, whose purchasing cycle and volumes do not lend themselves to the same commitments as a data centre operator, ends up on the wrong side of that split.

Several houses did cut their rating in early August, citing that cost line explicitly. China Renaissance moved to neutral on 4 August, target reduced from $329 to $280, citing memory inflation likely to last several more quarters. On 10 August, Jefferies moved from hold to underperform under analyst Edison Lee, target cut from $285.56 to $263.66, on supply chain checks indicating the all-glass anniversary model had been cancelled. The note sees in it « a major setback to efforts to bring in higher-priced iPhones, amid soaring memory costs » (CNBC, 2026; Fortune, 2026). The wording explicitly ties the product argument to the cost argument, that is to the margin bridge of section 2.

A point of proportion is required, without which that paragraph would mislead. This move remains a small minority: of fifty-one analysts covering the name, thirty-two still carried a buy-equivalent rating and only five a bearish one at the time of the note (Fortune, 2026). Those views are reported as news facts attributed to their authors, and are not adopted as our own. We deliberately quote no percentage share price move for that day: the available sources measure it on different bases, close to close for some, close to intraday low for others, and the gap between the two readings exceeds half a point.

4.Opportunity: what the product calendar loads into the December quarter

The starting fact is published and verifiable twice over. The first quarter of fiscal 2026, the December one, ended 27 December 2025, produced $143,756 million of revenue, an all-time record for the company (Apple Inc., 2021-2026). The same amount is recovered by simple difference in the July filing: nine months at $364,357 million, less $111,184 for the March quarter and $109,417 for the June quarter, equals $143,756. Both routes give the same number, which is the most elementary and most reassuring check one can run on a set of accounts.

The mechanism is explicit and Apple describes it itself in its annual report, under a heading devoted to business seasonality and product introductions: holiday demand combines with the timing of new product launches and channel fill (Apple Inc., 2025). The most telling measure of that mechanism is this: the iPhone 17 reached stores on 19 September 2025, that is only eight trading days inside the September quarter. iPhone revenue for that quarter comes to $49,025 million, then $85,269 million in the following quarter, the first full one for the new line, a rise of 73.9%. The product is launched in September; it is paid for in December.

Two precautions are indispensable before drawing anything from this. First, Apple'sfiscal year is a 52 or 53 week period ending on the last Saturday of September, and one December quarter in five or six counts fourteen weeks instead of thirteen: that was the case in 2022, whose $117,154 million are not comparable as is with the neighbouring years. Second, as of 15 August 2026, Apple has announced no date for its September event: its events page carries no dated deadline, and the dates circulating in the specialist press, from 8 to 15 September depending on the outlet, are conjectures that contradict one another (Apple Inc., 2026d).

The analytical interest is therefore not in discovering this seasonality, which everyone knows, but in measuring what it is worth in 2026 against the margin compression. On a base of $143.8 billion, each point of gross margin is worth about $1.44 billion, against $1.1 billion in an ordinary quarter. The quarter in which Apple collects the most is also the one in which the rising cost of memory costs it the most. And the table below adds something the usual narrative ignores: the December quarter premium has been compressing for five years.

5.The question everyone asks: can a market rule be drawn from it?

The « sell the news » story goes round every autumn: the share is supposed to rise before the iPhone event and fall after it. We wanted to test it rather than repeat it, and we did so on the raw prices in the EPTA5 base, in market-adjusted returns, that is net of the index the same day, across the nineteen iPhone launches from 2007 to 2025.

The event-day result is negative but not significant: a mean of minus 0.385%, a median of minus 0.733%, a t statistic of minus 0.69, and a 95% confidence interval running from minus 1.48% to plus 0.71%. That interval contains zero, so the hypothesis of no effect cannot be rejected. Fourteen of the nineteen sessions are negative, that is 73.7%, but the Wilson interval on that rate runs from 51.2% to 88.2% and brushes a coin flip.

Two findings finish emptying the story of its substance. First, the effect is specification-dependent: keeping only the September-era events, from 2012 to 2025, the sample falls to fourteen and the result becomes nominally significant at 5%. The only thing that changed is the exclusion of the January 2007 session and its plus 8.39%. An effect that appears or disappears on a discretionary sample cut is not a robust effect. Second, and this is decisive, the subsequent downward drift does not survive market adjustment: over the following twenty-one days Apple's raw return is minus 0.31% on average and minus 1.33% at the median, which looks like selling the news, but once the index is netted out the mean becomes plus 0.21% and the median plus 0.31%, indistinguishable from zero. Apple's apparent weakness after its launches is in fact the seasonal weakness of the whole market in September and October.

As for the « 69% of negative sessions » figure one sees circulating, it comes from an unrefereed working paper and rests on thirteen observations: 69.23% is exactly nine thirteenths, and the corresponding Wilson interval runs from 42% to 87%. It contains the coin flip. Two brokerage notes moreover claim to measure the same phenomenon and contradict each other on the sign, one announcing minus 1.8% in the month after the announcement, the other plus 2.8%.

6.Why the terminal's seasonal ranking does not rescue the story

Queried on Apple, the seasonality module displays a cycle-phase pattern giving plus 50.0% on average over twelve rolling months, computed on eleven observations, with no confidence interval whatsoever. The figure is spectacular and it is worth nothing. Eleven observations spread over forty-four years cover a company that has changed its nature three times, and the same screen queried on a semiconductor produces, at the same instant, an average of minus 0.7% over nine observations.

Figure 3 explains the mechanism. The engine sweeps 2,388 patterns and shows only the top of the ranking. Sweeping two thousand three hundred and eighty-eight calendar windowsmathematically guarantees finding a few showing 100% win rates, even if returns are purely random. The first screen of the ranking contains several, and the same ticker appears twice with two different windows, which is the signature of overfitting rather than of a phenomenon.

The conclusion of sections 4 to 6 is therefore explicit: the opportunity is an accounting one, not a market one. That the December quarter is the largest of the fiscal year is published, cross-checked twice and robust; the idea that an entry and exit rule on the share can be drawn from it survives neither our own event study nor the terminal's ranking read properly.

7.The supplier publishes what the buyer absorbs

If Apple pays more, who collects the difference? The compression Apple describes has an exact counterpart, and it is public. The largest American memory supplier reported for its quarter ended 28 May a gross margin of 84.6% and guides the following quarter to about 86% (Micron Technology, 2026). Its annual growth in memory breaks down into an average selling price up roughly 260% and volumes up only about 20%. One company's margin is the other's cost, and both filings describe the same displacement of value, three weeks apart.

The shift in market shares confirms it. In the second calendar quarter of 2026, global memory revenue splits between Samsung at 39%, SK Hynix at 26% and Micron at 25%, where SK Hynix held 39% a year earlier (Counterpoint Research, 2026). A market recomposing that fast is a market where pricing power has changed hands, and the consumer electronics buyer is not on that side.

This cross reading is also the best way to date the end of the episode. The turn in Apple's margin will not come from a decision by Apple: it will come from the end of the imbalance between wafer supply and server memory demand.

8.Caveats on the terminal's data

Three reservations must accompany any use of the terminal on this file, and they matter well beyond Apple. The first is the vintage trap already flagged in figure 2: the company card shows a last price of $266.25 stamped 20 November 2025, a gap of more than fifteen percent with the mid-August 2026 market. No price level, market capitalisation or multiple from that panel appears in this note.

The second concerns the seasonal ranking and is detailed in figure 3: a volatility column that divides values above 100% by a hundred, an annualised Sharpe ratio sitting beside returns that are not, and a sweep of 2,388 patterns of which only the top is shown.

The third appeared during the analysis and deserves particular flagging. The terminal's economic calendar shows the correct dates for the coming deadlines, Federal Reserve on 16 September, Bank of England on 17 September, employment report on 4 September, but it pairs them with stale values: a US policy rate of 4.38% when the official range is 3.50 to 3.75%, a British rate of 4.25% when the Bank of England holds it at 3.75%, and a payroll figure of plus 143,000 when July destroyed 23,000 jobs. The dates are usable, the values are not.