Summer Insects and Ice: The Week the Market Sold While the Order Books Said Sold Out

You cannot discuss ice with a summer insect —
it is bounded by its season.
—— Zhuangzi, “Autumn Floods” (4th c. BC)

What this is

I follow a handful of people doing genuine primary research on X. One is Serenity, whose focus is the upstream of AI infrastructure — optics, memory, packaging, power. He goes to industry conferences to do channel checks, reads earnings transcripts and SEC filings, builds revenue models, and then posts the conclusions publicly.

Through the last week of July and into August, the market ran a violent bout of AI deleveraging. When I lined up the public numbers he pointed at that week, the picture was jarring: prices collapsing, and every supply-side figure saying the same thing — there isn’t enough.

This isn’t a summary of his views. It’s what I saw when I put that week’s public information side by side. If you want the primary material, go follow the account — it’s worth far more than my second-hand tidying.

The four threads I read that week

1. Price and fundamentals came completely unhooked

Those days were indiscriminate global AI deleveraging — high-beta names hit first, regardless of what was underneath.

Meanwhile, that same week: a fuel cell company reported 166% year-on-year revenue growth, expanding margins and raised full-year guidance. A semiconductor test equipment maker reported 104% revenue growth with over 300% EPS growth.

And the most jarring example: Reddit beat substantially on revenue and EPS, guided next quarter above street — and fell 22.6% on the day.

His description of that is right: every time something drops, the market finds a story for it after the fact.

2. Capex isn’t shrinking — it’s being revised up

The same week, the four largest hyperscalers guided 2026 capex to a combined $720–745 billion, up from $695–725bn. Amazon went from $200bn to $220bn, partly because of rising memory costs.

The line that matters most is Amazon’s CEO saying that even at $220bn, they will not have the capacity to meet all of 2026’s demand.

Two Microsoft details worth keeping: they expect free cash flow to be positive in 2027 even with capex rising — meaning this build-out is being funded out of operating income; and roughly two-thirds of quarterly capex is “short-lived assets,” primarily CPUs and GPUs.

(Note also that Microsoft’s 2026 figure moving from $190bn to $175bn is a reclassification of data centre leases — not a cut. That’s the kind of detail you misread if you don’t go to the source.)

3. The bottleneck is migrating upstream, and the gap is widening

This is the most important thing I saw all week.

Memory: SK Hynix, Micron and Samsung have sold out of 2027 DRAM/HBM capacity; current annual NAND capacity is sold out too. The allocation reality: customers are receiving only 60–70% of the volumes they originally requested. Industry commentary points to 2027 being the most acute moment of memory shortage.

Lasers: Lumentum’s CEO warned publicly at an industry summit that the InP laser supply situation for AI data centres is more severe than memory. Even with five InP fabs, shipments are running more than 30% below customer demand.

Put those next to each other and the message is: this cycle has more than one bottleneck, and the further upstream you go, the larger the percentage gap.

Two supporting data points: a probe card and test equipment maker moved its 2026 co-packaged optics revenue expectation from “the high end of the $10–20m range” to “exceeding that range by the end of Q3, and significantly exceeding $20m for the year.” And in the same week, two leading global optical players each signed long-term agreements with the same indium phosphide substrate supplier, paying $22.3m and $43.5m respectively purely to reserve capacity.

When customers pay cash up front to lock capacity, that is more honest than any analyst estimate.

4. A liquidation, and how it got told

The week’s most dramatic event was an AI-focused fund with an exceptional record (+439% year-to-date through June) getting hit in the July drawdown — leverage, liquidity management and hedging all came apart — and the positions being liquidated.

The next day the market ripped, and the liquidated names bounced violently: several up between 17% and 26% in a single session.

Serenity said two things about it that I found more valuable than the market commentary:

First, framing a thesis that worked for years as a “collapse” and a “failure,” then attacking the person after a sharp drawdown, is contemptible. July did expose real problems with that fund’s leverage, liquidity and hedging — that’s a fact. But whether his industry judgment was right is a separate question.

Second, “his ideas would have played out even if no retail investor had known he existed.” NAND shortages don’t care whether anyone copied the trade, and neither does hyperscaler demand for fuel cells. Reducing a structural call to “everyone was copying his homework” is a disservice to the facts.

What I took away

1. Narrative and ledger fully separated that week, and I need to read them apart.

The media that week: hyperscalers overbuilt and are selling excess compute. The four earnings calls: demand far exceeds capacity, capex goes up.

Institutions that week: emergency hike, three hikes, it’s over. The Fed: unchanged at 3.50–3.75%.

I’ve written this down as a discipline: when narrative and ledger conflict, first check whether the ledger is one I read myself. If my “ledger” is also second-hand narrative, then I’m just picking between two stories — which isn’t independent judgment.

2. Indiscriminate deleveraging is a state of “temporarily not looking at fundamentals,” not a change in them.

Reddit is the cleanest case: comprehensive beat, raised guidance, down 22.6%. What’s being sold in that moment isn’t the company — it’s positions.

The practical implication isn’t “so buy the dip.” He said it himself: “I can’t call the bottom accurately, and neither can anyone else, or they’d just be trading weekly options.” It’s that during an indiscriminate drawdown, price temporarily loses information content. The only useful work in that window is telling apart: has my thesis been falsified, or is somebody being forced to sell? Those require completely different responses.

3. Prepaying for capacity is the strongest demand evidence I know of.

Analyst estimates are wrong; company guidance is both sandbagged and inflated; channel checks have sample bias. But two competitors each putting up tens of millions in reservation payments in the same week, purely to lock future supply of the same substrate — that’s voting with cash.

I’ve promoted this into a checklist item: to test whether a bottleneck is real, look for prepayment. Long-term agreements, reservation fees, offtake, price floors — paying before you receive the goods is much harder to fake than any supply-demand model.

4. Bottlenecks have layers, and the percentage gap tells you which layer you’re on.

Memory: customers get 60–70%. Lasers: shipments more than 30% below demand, described as more severe than memory.

Along one supply chain, different layers are under very different strain — and the tightest layer isn’t necessarily the one everyone is talking about. That week everyone discussed memory; the CEO said lasers were worse.

Which matches what I’ve been doing for six months: walking up the supply chain until you find the link that breaks first when demand doubles. And remember — the upstream gap is usually the larger percentage, because its expansion cycle is the longest.

5. Not cheering for other people’s failure is less a moral point than a discipline.

There was plenty of schadenfreude on X when that fund was liquidated. Serenity chose to separate two things: the leverage and risk management genuinely failed; whether the industry call was wrong is a different question.

That separation is itself an investing skill. Reading “he blew up” as “he was wrong” makes you miss something important: a correct call can be lost through incorrect position management. And the reverse holds too — a wrong call can make money on luck.

The correctness of a judgment and the survival of a position are independent dimensions. Judge only by outcome and you’ll never learn that.

6. People who publish real research for free deserve to be followed, and treated well.

He wrote something rueful that week: going to conferences for channel checks, finding unique supply chain links that later get confirmed, building wafer revenue models, tracking ramp timelines, doing competitor and TAM analysis — and then posting the whole thesis for free.

That landed on me. I’ve spent six months writing this sort of public material and I’m well aware the return on effort doesn’t look good. But people like this are one of the few things in the market that make serious retail investors better. So let me say it plainly: go follow the original account instead of reading my second-hand version.

Notes on sources

• Every figure here comes from public earnings, public industry summit remarks and public news. I noticed them by following the independent researcher Serenity on X — go to the source for the primary material
• Nothing in this piece comes from paid subscription content
• Two related pieces written the same day: Bond Traders Panic When the Fed Doesn’t and The 1873 Railroad Bubble
• The lines from the Zhuangzi are my own footnote

---

Disclaimer: This is a reader’s reflection and general education, not investment advice, an offer, or a solicitation. I have deliberately omitted tickers, recommend nothing, and give no price targets. Figures are relayed from public sources and may contain transcription error — go to the originals. Investing carries risk; judge for yourself against your own circumstances and consult a qualified professional if needed.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.