Investing / The Compound Effect
A letter from 2031, about the week the machines ran out of memory
| 4 min | By Heath J. Harris
Nvidia booked a $96 billion quarter and warned it cannot buy enough memory. So we wrote ourselves a letter from five years out about what this week looks like from there.
A note before the letter. This is a thought experiment. The numbers about this week are real and sourced. Everything about 2031 is imagination, because nobody gets tomorrow's paper early, including us.
Dear reader in August 2026,
By the time you get this, you already know how the week felt. Nvidia reported $96.2 billion in quarterly revenue Wednesday night, up 106% from a year earlier, and guided next quarter to $108 billion (CNBC, August 26, 2026). Thursday, the market added more than $400 billion to its value, pushing it back above $5.5 trillion, the largest company on earth. Friday it gave a chunk back on questions about how its AI customers finance their purchases (Yahoo Finance, August 28, 2026). One company moved more value in two days than most S&P 500 members are worth in total.
From where we sit in 2031, that is not the part anyone remembers.
The number nobody led with
Buried in the earnings call was a line item: $279 billion in memory purchase commitments, plus a warning that memory shortages could persist into fiscal 2028 (24/7 Wall St, August 27, 2026). The most powerful chip company in the world stood up and said, in effect, we cannot buy enough memory.
That should have been the whole story. An AI server uses roughly eight to ten times the memory of a traditional server, per industry estimates. The three companies that make advanced memory (SK Hynix, Samsung, Micron) had already sold out of DRAM through 2027 (Motley Fool, August 6, 2026). SK Hynix alone controlled 58% of high-bandwidth memory, the specialized kind AI chips cannot run without (Counterpoint Research).
And here is the strange part you may have noticed at the time: the market shrugged at the sellers of the scarce thing. Micron opened up 3% on the Nvidia news Thursday and erased most of the gain within minutes (24/7 Wall St, August 27, 2026). It finished the week near $933, about 26% below its June high, despite being up roughly 220% on the year. It traded around 6.5 times next year's expected earnings while the S&P 500 traded near 20 (StockAnalysis; FactSet). Investors were paying a premium for everything with AI in the story except the companies whose product was actually sold out.
Why the market argued with itself
There were two readings of that discount, and in August 2026 you could not know which was right.
Reading one: memory is a commodity, and commodity booms end the same way every time. Sold out today means overbuilt tomorrow. The market kept a long memory of 2000, when the fiber-optic build-out was real and the fiber stocks still fell 90%. Cheap on peak earnings is the oldest trap in cyclicals.
Reading two: this build-out was different in one specific way. It was being paid for out of the operating cash of the most profitable companies in history. Microsoft, Alphabet, Amazon, and Meta had guided to roughly $725 billion of combined capital spending for 2026, up 77% in a year (Tom's Hardware). Booms die when the money runs out. This money had not run out.
We would love to tell you which reading won. We are not going to. Partly because we promised not to spoil it, and mostly because the useful lesson turned out to be the same in both futures.
What survived both versions of the future
In the version where it ended badly, the people who got hurt were rarely the ones holding some Nvidia inside an index fund. They were the ones who noticed the AI trade had quietly become an outsized share of their net worth and decided that was fine, because it had been right so far.
In the version where it kept going, the people who suffered were the ones who sold everything because a stock they never planned to own hit $5 trillion and made them feel late.
Both groups made the same mistake. They let the most exciting thing on television set their allocation. The clients who did fine in either version had a written target, a band around it, and a rule: when the band breaks, we trade. When it holds, we do not touch it. Boring saved them, or boring cost them a little upside. Boring never took them out of the game.
The week the machines ran out of memory, the question that mattered was never what memory chips would do next. It was whether your plan depended on knowing.
With hindsight and affection,
Your future self
(Signed from 2031. Mailed from 2026.)
If you want help pressure-testing your own allocation against more than one future, our team at Compound Advisory does this work every week. You can schedule a complimentary assessment at compoundadvisory.co/retirement-clarity-assessment.