A gray rhino, telegraphed since November — the cost side, not the crash side

The AI boom just started charging everyone else for memory.

For six weeks the memory story was about chipmakers' share prices. On July 30, 2026 it became a story about what things cost. Apple — the world's largest buyer of consumer hardware components — told investors that memory prices explain more than 100% of its sequential gross-margin decline, and guided margins down again for the September quarter. Scarce memory is now being rationed by contract, not price: the hyperscalers locked in multi-year supply, so from this quarter the increases land on the buyers who didn't.

Server DRAM · 3Q26
+13–18%
QoQ contract · TrendForce Jul 9
Apple adj. gross margin
48.1% → ~46.5%
Jun qtr → Sep guide
PC memory · share of BOM
23%
2026 peak, from 16% in 2025
Core goods PCE · 12m
2.4%
May 2026 · 0.6% a year earlier
DRAM makers · top 3
~90%
1Q26 revenue share
HY credit spread
2.87pp
Jul 29 · reads calm
The bottom line, in plain English

The chips that store data — memory — are in short supply because AI data centres are buying them at any price. The big cloud companies signed long-term contracts that freeze what they pay. Everyone else — phone and laptop makers, small businesses, and eventually you at the till — now absorbs the increases. Apple has put a number on it, and the Federal Reserve's own July report already names computers and electronics as a reason goods prices are rising. This is not a market-crash warning. It is a purchasing-power warning: the AI build-out is showing up as things costing more, in a year when core inflation is already 3.3% and three Fed officials voted to raise rates. What changed and why — not advice, not a forecast of a crash.

01 / The frame

A visible rhino: the shortage was forecast, the incidence was not

A black swan is a shock nobody could foresee. A gray rhino — Michele Wucker's term — is the opposite: a highly probable, high-impact danger everyone can see charging, and ignores anyway. The memory squeeze is squarely the second kind, and it has been on the record for months:

Two disciplined caveats

(1) We did not call this. This desk covered the memory complex as a producer-and-share-price story for six weeks — the June 23 Seoul crash, the July circuit breakers, the record earnings — and first wrote up the consumer-cost leg on July 31, the day after Apple quantified it. The standing danger was visible; claiming we timed it would be hindsight (lesson L-001: compelling ≠ correct).

(2) This is not the crash report. Our June 23 note covered the equity shock in the same industry. This one is about a different mechanism entirely — physical scarcity and who pays for it — and the systemic gauges here read calm: high-yield spreads 2.87pp, the 2s10s curve +0.45pp. The risk described below is to prices and margins, not to the plumbing.

02 / What changed

Record profits upstream, compressed margins downstream

The last week of July made the split unmistakable. The three companies that make roughly 90% of the world's DRAM by revenue reported extraordinary results — SK hynix a record quarter at a 76% operating margin, Samsung an all-time record with ₩89.5tn group operating profit — while the largest downstream buyer of memory on earth guided its margin down for the second consecutive quarter and said it could not build as many iPhones, Macs and iPads as it wanted to.

That is what a genuine input shortage looks like from both ends. It is also why the equity market's reaction has been such a poor guide to the economics: Samsung's shares closed roughly flat on a record quarter and SK hynix fell, because share prices trade the future of a cycle, while the cost pass-through works through contracts signed today and products priced next quarter.

The cost chain, figure by figure. Each row is dated to when the figure was published, and forecasts are labelled as such.

The memory cost chain — upstream price to downstream margin to shelf price
Link in the chainFigureAs of / sourceType
DRAM industry revenue+81% QoQ1Q26 · TrendForce, Jun 1 2026realised
Server DRAM contract price+13–18% QoQ3Q26 · TrendForce, Jul 9 2026supplier survey
2027 RDIMM bit supply+15–20% YoYsupplier guidance via TrendForce, Jul 9 2026forecast — "significantly lagging" CPU shipment growth
PC memory · share of bill-of-materials16% → 23%2025 → 2026 peak · Gartner, Feb 26 2026forecast
Apple adjusted gross margin49.3% → 48.1% → ~46.5%Mar → Jun 2026 qtrs → Sep guide · Apple call, Jul 30 2026reported + company guide single-source (call)
Apple reported gross margin50.1%Jun qtr · Apple newsroom, Jul 30 2026includes ~2pp favourable tariff refunds
PC prices / smartphone prices+17% / +13%2026 vs 2025 · Gartner, Feb 26 2026forecast — see dispersion below
PC units / smartphone units−10.4% / −8.4%2026 vs 2025 · Gartner, Feb 26 2026forecast
PCE core goods inflation0.6% → 2.4%12m to May 2025 → May 2026 · Federal Reserve MPR, Jul 2026official data
03 / The mechanism

Rationed by contract: who is shielded, and who pays

This is the part that has had the least attention and matters the most. In a normal shortage, price rises until demand gives way, and everyone faces roughly the same price. That is not what is happening. Several US cloud providers signed multi-year long-term agreements that restrict suppliers from raising prices to them. TrendForce's July 9 assessment is explicit about the consequence: from the third quarter of 2026, the primary source of server DRAM price increases shifts toward customers without long-term agreements, plus any incremental volume bought outside an existing agreement.

So the scarcity is not being shared. It is being allocated by who had the balance sheet and the foresight to lock in supply eighteen months ago — and the residual lands on everyone who buys memory in the open market. You can see the adaptation already: cloud buyers and server makers are stepping down module sizes, from 96GB and 128GB RDIMMs to 32GB and 64GB, to hold procurement costs flat. When a hyperscaler is downsizing its memory configurations, a laptop vendor with a $499 price point has no room at all — which is precisely why Gartner expects the sub-$500 entry-level PC segment to disappear by 2028, and expects buyers of basic smartphones to exit the market five times faster than premium buyers this year.

Incidence of the shortage — who absorbs the price, per TrendForce's Jul 9 2026 assessment and Gartner's Feb 26 2026 forecast
BuyerProtectionWhat they absorbObservable evidence
US hyperscalers / large CSPsmulti-year LTAscapped contract price; incremental volume outside the LTA still repricedAmazon raised 2026 capex to ~$220bn from ~$200bn; Microsoft held spending steady
Server OEMs & buyers without LTAsnonethe 13–18% QoQ increase, and further upward revisionsRDIMM configurations stepping down to 32/64GB to hold cost
Premium device makersscale + margin headroommargin compression, partly passed onApple: memory >100% of sequential GM decline; MacBook/iPad prices raised
Entry-level device makersnonesegment viability — memory at 23% of BOM leaves nothing to absorbGartner: sub-$500 PC segment gone by 2028; AI-PC penetration delayed to 2028
Householdsnonehigher shelf prices, fewer choices, longer replacement cyclesPC lifetimes +20% for consumers by end-2026; retailer ASPs guided up
04 / The macro question

The Fed has already written this down

The strongest evidence that this is more than a trade-press story is that it appears in the Federal Reserve's own July 2026 Monetary Policy Report. Core goods inflation, the Fed writes, "moved up notably over 2025 and continued to rise in the early months of this year": the 12-month change in PCE core goods prices was 2.4% in May, against 0.6% a year earlier. And on the cause, the report is specific — rapid gains in consumer prices for "computers and other electronics" contributed to core goods inflation, gains that "likely reflect the surge in demand for semiconductors and other components important to the buildout of data centers."

Why this matters more than a tariff line-item

"Although a large portion of high-tech goods are imported, most of these products are exempt from tariffs, so changes in tariff rates are not an important factor in their prices." Board of Governors of the Federal Reserve System, Monetary Policy Report, July 2026

Read plainly: this particular price pressure is not a trade-policy artefact that a court ruling or a renegotiation could unwind. It is a physical allocation problem, and interest rates do not manufacture DRAM.

Set that against the week's macro data and the discomfort is obvious. Core PCE eased only to 3.3% year-on-year in June from 3.4%; second-quarter GDP growth came in at an annualised 1.5%, below the 1.8% expected; and on July 29 the FOMC held at 3.50–3.75% with three regional presidents dissenting in favour of a hike — the most officials pushing one direction since September 2016. A cost-push impulse in durable goods is the least convenient thing that could arrive in that configuration.

How big, honestly? Estimates diverge enormously, and averaging them would be false precision. Gartner's February forecast is +17% on PC prices and +13% on smartphones. IDC has put PC price increases at 4–8%; Counterpoint has average smartphone prices up 6.9% versus 2025. both secondary-sourced That is a two-to-fourfold spread between reputable houses on the same question, nine months into the episode — which is itself the most honest single fact in this report. PCs and phones are also a small share of a consumer basket dominated by housing and services, so even the high end of that range is a fraction of a percentage point on headline inflation, not a regime change.

05 / Pre-mortem

The four ways this note is wrong

Written before the fact, not after, and in descending order of how likely each is to embarrass us.

1 · It shows up as fewer sales, not higher prices. This is the most probable failure, and it is in our own evidence: Gartner's forecast has unit shipments falling 10.4% and 8.4% precisely because buyers refuse the price. TrendForce's July 3 note says memory price gains are already moderating as consumer demand weakens and base effects bite. If the shock is absorbed by demand destruction and downgraded specifications, it lands in corporate volumes and household deferral — visible in shipment data, largely absent from CPI. The savings risk would then be to earnings in the device chain, not to purchasing power.

2 · Memory is the most cyclical commodity in technology. Every previous DRAM squeeze ended with a capacity glut and a violent price collapse. Samsung is guiding capex into the high-₩40tn range; China's CXMT went from roughly 3% to 8% of the market in a year. The 2027 tightness in this note rests on supplier-provided 2027 guidance, and suppliers benefit from signalling scarcity. Treat the +15–20% bit supply figure as an interested party's estimate, not a fact.

3 · Part of this is a reversal, not an increase. Consumer electronics have delivered decades of falling quality-adjusted prices. Some of the current core-goods contribution is the absence of the usual deflation rather than genuine inflation — a real change in the arithmetic, but a smaller one than the headline percentages imply.

4 · The framing could be too calm. We lean on high-yield spreads at 2.87pp and a positively sloped curve to say the plumbing is fine. Those gauges are lagging by construction and have read calm into previous turns. If the memory shock becomes a cost-of-capital story for the AI build-out itself, credit would confirm late and this note's separation of "cost problem" from "systemic problem" would look naïve.

What the desk is actually claiming. That an input shortage created by AI infrastructure demand is being allocated by contract rather than price; that the burden therefore falls disproportionately on buyers without long-term agreements, including households; that this is already measurable in official core-goods inflation data by the Fed's own account; and that its magnitude is genuinely uncertain, with respectable estimates differing by a factor of two or more.

What it is not claiming. Not that a market crash follows. Not that inflation re-accelerates to a new regime. Not that any security is a buy or a sell. The protective posture this suggests is the dull one: if you were planning to replace memory-heavy hardware, the price and availability trend has been running against you for three quarters and is guided to continue — a budgeting fact, not a trade.

Falsifiable claims logged to the calibration record on publication, with resolution criteria fixed in advance: P-0057 TrendForce reports a further quarter-on-quarter rise in 4Q26 server DRAM contract prices, p=0.80 · P-0058 PCE core goods 12-month inflation stays at or above 2.0% in the August 2026 reading, p=0.70 · P-0059 Apple's reported September-quarter gross margin lands inside its guided 47–48% range, p=0.65 · P-0060 disconfirmer — full-year 2026 worldwide PC shipments fall by less than 8%, i.e. Gartner's −10.4% proves too pessimistic, p=0.55. These are scored whether they land or not; the failures are kept permanently.

06 / Reference

Glossary and sources

DRAM
Dynamic random-access memory — the working memory in phones, laptops and servers. Distinct from storage (SSD/NAND), which holds data when the power is off.
HBM
High-bandwidth memory — the premium DRAM stacked next to AI accelerators. It consumes the same fabrication capacity as ordinary DRAM, which is why AI demand raises the price of the memory in a laptop.
Contract vs spot price
Contract prices are negotiated with large buyers, usually quarterly; spot is the open market. Contract prices are what device makers actually pay.
LTA (long-term agreement)
A multi-year supply contract that fixes or caps price. The central mechanism in this note: LTA holders are shielded, so the increase concentrates on everyone else.
RDIMM
Registered memory module — the server format. "Bit supply" is total memory capacity shipped, the honest measure of whether supply is actually growing.
Bill of materials (BOM)
The component cost of building a device. Memory rising from 16% to 23% of a PC's BOM is what removes a vendor's ability to absorb the increase.
Core goods inflation
Price change of physical goods excluding food and energy. The component where a component-cost shock would show up first.
Gray rhino
A highly probable, high-impact risk that is visible and ignored — as opposed to a black swan, which is unforeseeable. Michele Wucker's term; the house lens.

Confidence flags used above

multi-source / primary document single-source — not independently verified claimed but unsourced — excluded from this note

One attribution caution, stated rather than smoothed over: the "hundred year flood" phrasing from Apple's July 30 call is reported by different outlets as coming from chief executive Tim Cook and from chief financial officer Kevan Parekh. Because we cannot corroborate which executive said it, the phrase is attributed here to Apple management on the call and to no named individual. attribution conflict across outlets