Data as of 1 October 2026: market values and expected earnings for 97 Nasdaq-100 companies, plus the latest filings of Micron, NVIDIA and Apple.
Micron is one of every 36 dollars in the Nasdaq-100. It is also one of every five dollars of the earnings growth the index is supposed to deliver next year. That mismatch is the whole story, and Micron's margins now sit far above their two previous peaks.
So the question is not whether the memory cycle turns. It always has. The question is what it costs the index when it does, and the answer depends almost entirely on why it turns.
What the index is promising
On the analyst estimates compiled by Finviz, the Nasdaq-100's earnings should rise about 50% over the next fiscal year. At today's prices the index trades at 20.7 times those expected earnings (31.1 times what it actually earned over the last twelve months). That gap is the promise: you are paying today for profit that has not happened yet.
Now look at who is supposed to deliver it. The 22 chip and memory companies in the index are 32.5% of its value but carry 77.2% of the promised growth. Strip them out and the rest of the index is promising +16%, at a richer 25.1 times expected earnings. The rest of the index is not cheap; it is just not where the excitement is.
NVIDIA alone carries 28.3% of the promise. The four memory and storage companies (Micron, SanDisk, Seagate and Western Digital) carry 26.9% while weighing 4.2%. Memory is a small part of the building holding up a large part of the roof.
Two ways it ends, and they look nothing alike
The first is a glut. The three big memory makers finish the factories they are building, there are more chips than buyers, and prices fall. Micron, SanDisk and SK hynix lose. NVIDIA, the big clouds, Apple, Dell and the PC makers win, because memory is a cost inside everything they sell. For the index it is mostly profit changing hands.
The second is an AI spending slowdown. The companies paying for the party cut back. Memory falls first and hardest, and NVIDIA, Broadcom, Marvell, TSMC and the chip-equipment makers follow. Nobody wins, which tends to simplify the arithmetic.
Here is what each case would cost, measured one way only: how far the index would have to fall to stay exactly as expensive as it is today, once there is less profit to pay for.
| Scenario | Times expected earnings | Promised growth | Fall to stay as expensive |
|---|---|---|---|
| Today, as it stands | 20.7 | +50.2% | — |
| Memory stays where it is | 22.7 | +36.7% | −9% |
| Memory earns half of today | 23.5 | +32.6% | −12% |
| Memory earns a fifth, as in 2019-20 | 23.9 | +30.1% | −13% |
| AI slows: chips and memory stop growing | 27.9 | +11.4% | −26% |
| Real slowdown: chips −20%, memory ÷5 | 31.0 | +0.5% | −33% |
| For comparison: the index without chips or memory | 25.1 | +16.4% | — |
A memory collapse on its own, even a brutal one, costs the index around 10-13%. What costs it a third is the slowdown, because it takes NVIDIA, Broadcom and the rest down at the same time. These figures assume the market keeps paying today's multiple. If it also decides to pay less for each dollar of profit, the fall is larger, so read them as the minimum damage, not as forecasts.
How to tell which one is happening: if memory prices fall while NVIDIA keeps announcing more sales, it is the first. If NVIDIA or the big clouds announce less and memory falls at the same time, it is the second.
Micron's cycle, in one line
The two previous peaks left Micron with 61 dollars of every 100 in sales after manufacturing costs in 2018, and 47 in 2021. In 2023 it was losing money on making the chips alone. Today it keeps 86.8. Whatever this is, it is not normal.
The margin it expects for this quarter is already slightly lower: 86 against 86.8. But its sales are still growing 22% on a like-for-like number of weeks. The warning that worked in 2018, 2021 and 2022 was a different one: the quarter after the peak brought in less revenue than the one before. $7,913M against $8,440M in 2018, $7,687M against $8,274M in 2021, and $6,643M against $8,642M in 2022. That signal has not fired yet.
What NVIDIA's own accounts say
From its quarter ended 26 July and the CFO's commentary:
- It has tied itself to expensive memory. Supplier commitments went from $119bn to $279bn in a single quarter, mostly memory by its own description: $87bn for its next fiscal year and $88bn for the one after. Some can be cancelled, at an additional cost. In 2025 it already booked a $4.5bn charge on committed purchases of a chip it could not sell in China.
- It is passing the cost on, just. It kept 75 of every $100 after manufacturing last quarter, unchanged, and expects 74 for this one, in the middle of the jump in the price of the memory it puts inside every chip.
- It is collecting later and guaranteeing clients. It earned $63.7bn from the business and $24.1bn came in as cash. It now takes 60 days to get paid, up from 45, and it has guaranteed up to $105bn for a data-centre complex built for OpenAI.
- Consumers are noticing. In its own words, its PC sales slowed because of elevated memory prices.
None of this is a problem while demand holds. All of it becomes one the quarter demand does not.
Who is paying for all this
NVIDIA and Micron alone already sell the equivalent of more than 60% of what Microsoft, Google, Amazon, Meta and Oracle invest. And that is before SK hynix, Samsung, Broadcom, the buildings and the electricity.
It adds up because there are other buyers: only 55% of NVIDIA's data-centre sales go to those five. The other 45% goes to new AI clouds, governments and companies, the part financed partly with debt and, in some cases, with NVIDIA's own help.
What happened after the last two peaks
| Twelve months after | Peak of 2018 | Peak of 2021 | Already collapsed (June 2022) |
|---|---|---|---|
| Micron | −5% (at worst −30%) | −29% | +15% |
| NVIDIA | −38% (at worst −52%) | −41% | +179% |
| Chips in general (SOXX) | +16% (at worst −15%) | −28% | +47% |
| Nasdaq-100 | +3% (at worst −17%) | −25% | +33% |
| Apple | +1% (at worst −30%) | −2% | +43% |
Total return with dividends from the month-end close of the month of Micron's warning (September 2018, September 2021, June 2022). "At worst" is the worst month-end close within the twelve months.
Two occasions are a precedent, not a statistic. Both times the Fed was raising rates in the year that followed, and on 16 September 2026 it raised them for the first time since 2023. What the two cases show is that, when both things coincide, NVIDIA did worse than Micron itself. The best moment to buy all of it was when memory had already collapsed, which is the kind of advice that is only useful afterwards.
The dates that decide it
- 29 October, SK hynix: price per chip and new factories. If prices stop rising, the glut has started.
- Late October, Microsoft, Google, Amazon and Meta: planned investment for 2027. The first to cut is the warning of a slowdown.
- Late October, Apple: how much it keeps from every $100 of devices. In June it kept 40, about 37 without a tariff refund, against 34.5 a year earlier. Apple is 11% of the index and is the first test of whether expensive memory is reaching the consumer.
- Late November, NVIDIA: margin and supplier commitments. Below 74, or cuts to memory commitments, and the shortage is over.
- Around 30 December, Micron: revenue guidance. Less than the quarter before is the signal that worked in 2018, 2021 and 2022.
What this means for a company that is not listed
None of this stays on the stock exchange. The same accounts that tell an investor what the Nasdaq-100 is promising tell a finance director what is coming down the supply chain. Five readings we would take into a client's planning this quarter:
- If you buy hardware, budget the price in two scenarios. NVIDIA says in its own words that expensive memory slowed its PC sales, and laptops, servers and devices carry the same memory. If a hardware refresh is in next year's budget, price it at today's memory cost and at a lower one in case the glut arrives. Buying now or waiting is a cash decision, and it deserves a date rather than a hope.
- If you sell into the data-centre build-out, your forecast has a calendar. Electrical contractors, cooling, construction and staffing firms depend in large part on what Microsoft, Google, Amazon and Meta decide to invest. Their late-October results can tell you which scenario you are in before your own customers do.
- Collections are slowing at the top of the chain. NVIDIA now takes 60 days to get paid instead of 45. When the strongest company in a chain collects later, the pressure can travel down it. A 13-week cash flow forecast that schedules each customer on the day they actually pay sees that before the bank balance does.
- Purchase commitments are a liability that does not look like one. Of NVIDIA's $279bn of supplier commitments, only some can be cancelled, and at an additional cost. Most smaller companies have their own version: minimum-volume contracts, take-or-pay clauses, annual orders. They belong in the forecast, with the exit cost written next to them.
- Money got more expensive. The Fed raised rates on 16 September, and our Governor, the rule that decides each week whether surplus cash can be put to work, is now red. A credit line drawn calmly costs less than one drawn in a hurry.
The method is the one used above: name the two scenarios, pick the signal that tells them apart, and agree in advance what you will do when it fires. That is most of what a fractional CFO does for a company of this size, and the part a spreadsheet cannot do on its own.
In one sentence
The Nasdaq-100 promises 50% earnings growth and three quarters of that promise is chips and memory: if memory falls because too many factories were built, it costs the index around 10-13% and helps NVIDIA and Apple; if it falls because AI spending slows, it costs a third and, as in 2018 and 2021, drags NVIDIA down harder than Micron, so watch what moves first, the price of memory or what the big clouds invest.
Frequently asked questions
How does the memory chip cycle affect a company that is not in the chip business?
Through three channels: the price of the hardware it buys, the demand of its customers if it sells into data-centre construction, and the payment terms that travel down the supply chain. In its quarter ended 26 July 2026, NVIDIA said expensive memory slowed its PC sales and took 60 days to collect from customers, up from 45.
What should a CFO do when a key industry cycle may be turning?
Name the two scenarios, pick the signal that tells them apart, and agree in advance what the company will do when it fires. For memory in October 2026 the signal is which moves first: the price of memory, or what Microsoft, Google, Amazon and Meta say they will invest.
How can a company tell a supply glut from a demand slowdown?
In a glut, prices fall while the biggest buyers keep announcing more spending. In a slowdown, the buyers announce less and prices fall at the same time. A glut lowers costs for companies that buy the product; a slowdown hits everyone in the chain.
Why do purchase commitments matter for a smaller company?
Because they are a liability that does not appear as debt. NVIDIA's supplier commitments rose from $119bn to $279bn in one quarter, and only some of them can be cancelled, at an additional cost. Minimum-volume and take-or-pay contracts work the same way at any size, and belong in the 13-week cash forecast with their exit cost.