Actionable insights — Two variables, one credit chart, and a hostile calendar

4:03 1. Reduce the hyped revenue line to two measurable variables — price and volume

1. Find the theme's unit of currency — the thing that is actually sold and counted. For AI he names it explicitly: the token, "sort of the unit of currency for AI." For a cloud it is a compute-hour; for a network it was a bit of bandwidth; for a mine it is a pound.
2. Split revenue into exactly two series and track them separately: price per unit and units produced. Refuse to look at revenue in aggregate, because a collapsing price and an exploding volume net out to a number that tells you nothing.
3. Date both series from a fixed reference point so the comparison is honest. His is end of May: price per token −50%, tokens produced +2.5x.
4. Compute the ratio and state the threshold in advance: is "the dropping in costs being offset by a more than offsetting increase in tokens?" A 50% price cut needs more than a doubling of volume just to stand still.
5. Identify what is driving the price series, since that tells you whether the fall is one-off or structural. Here it is open source / open weights — a permanent supply of free substitutes, not a promotional discount.
6. Look for a step-change in the volume driver that can outrun price. His is the workload mix: "the agentic phase uses 10 to 100 times more tokens than the chat-based AI phase," and it only began on January 30th.

• Published price-per-token curves flattening (the substitute is done cutting) or, worse, volume growth decelerating while price keeps falling — the combination that ends the trade; the agentic share of workloads stalling; any month where the ratio of volume growth to price decline drops below 1.

4:57 2. Require a demand test and a profit test in the same print

1. Score the segment on the change in growth rate quarter over quarter, in percentage points, not on the level. "Fast" is priced; "accelerating" is new information. Big three clouds: 35% → 43%.
2. Then pull the same segment's operating margin for both quarters and require it to be up as well. Here: "more importantly, the operating margins also expanded about 2 percentage points."
3. Reject each half alone. Accelerating growth on a flat or falling margin means the growth was bought — with price cuts, subsidised capacity or absorbed hardware cost. A rising margin on decelerating growth is harvesting.
4. State the reason the pair matters before you run it: "you may be generating revenues, but if you're losing tons of money, it really doesn't matter."
5. Run it across every competitor in the layer in the same quarter and rank them — the spread between peers is the signal, not any single print.
6. Then add a third line the income statement hides: free cash flow. A segment can accelerate and expand margins while the parent's cash flow goes negative because the capex is capitalised. That divergence is the whole story of this episode.

• The second derivative turning — acceleration going negative while the level still looks high; margin expansion sourced from longer depreciation schedules rather than pricing; the gap between segment operating margin and consolidated free cash flow widening further; "supply constrained" language disappearing from the calls.

40:48 3. Read the credit market before the equity market — hyperscaler CDS vs the IG index

1. Build a short daily watchlist of the credit default swaps of the theme's biggest borrowers and check it every morning. His list: "Microsoft and CoreWeave and all the rest of them, Nvidia… What's the cost to insuring that debt?"
2. Never read the level in isolation — compare it to a benchmark index. His is the 5-year North American investment-grade CDS average. The signal is the spread to benchmark, and specifically the moment a mega-cap prices wide of the average investment-grade company.
3. Justify the priority with the payoff asymmetry, which is why credit leads: a lender's upside is capped at being repaid, so "they have to feel darn sure that they're going to get paid back." Equity and venture investors can afford nine zeros out of ten; "that's not the way it's going to work for a credit investor."
4. Ask how the build is being funded and flag the transition. Dot-com was "funding this through free cash flow"; this build is increasingly debt plus "circular stuff." A theme that switches from cash-funded to debt-funded has acquired a new failure mode that did not exist before.
5. Anchor the whole thing to the risk-free curve, since that sets the cost of everything: "the price of the 10 and 30-year treasuries, that determines everything… the 30-year is up at the highest level since 2007."
6. Escalate when three things coincide: rising CDS, a debt-funded capex programme, and high-yield/high-risk financing appearing in the capital stack. "If the cost of money is getting more expensive and you're having to fund a data center buildout with high-risk loans and your credit default swaps are going up, that's an issue."
7. Note the precedent that validates the screen: "credit default swaps were something that everybody was looking at during the global financial crisis to see, okay, which companies are potentially in trouble."

• An AI borrower's CDS gapping rather than drifting; the spread to the IG index widening for two or more names at once; new issuance pricing concessions; SPV / off-balance-sheet or vendor-financed data-centre structures; any hyperscaler guiding capex down — his stated end-game, where "they'll all probably go ahead and slam the brakes on at the same time."

6:45 4. Date the bottom by finding the leveraged forced seller, not by valuation

1. Separate the two drivers before doing anything else: "there's only two things that drive stocks… earnings [and] the multiple on those earnings." A drawdown driven by the multiple is a flow problem; one driven by earnings is a thesis problem.
2. When the multiple is compressing while fundamentals hold, go looking for who is being forced to sell. "The multiple… gets over inflated because you have individuals as a group or funds as a group doing stupid things. And stupid things usually means you've levered up."
3. Do the wipeout arithmetic on the suspected holder rather than guessing: at 4x leverage, a 25% drawdown is a 100% loss and mandatory liquidation. Work out what decline in the underlying triggers the margin call at each plausible leverage level.
4. Look for the double squeeze that accelerates it — shorts rallying while longs fall. "Software was rallying and your longs going down. In this case, semiconductors were getting crushed. You have both sides of your book working against you."
5. Confirm the fundamentals are intact independently before treating the selling as mechanical (here, via insight 1: price down, volume more than offsetting). Only then is the forced sale a buy signal instead of a warning.
6. Publish/act on the anticipation, not the confirmation — he put out the near-bottom note on July 29th "because I had been hearing about funds in trouble"; the sale to Citadel was announced the next day.
7. Treat the liquidation as risk-reducing: once the forced holder is gone, so is the supply. "The market obviously ripped the next day, especially in a lot of those positions that were getting unwound."

• Prime-broker chatter about funds in trouble; crowded-position overlap across leveraged holders; a long/short book whose shorts and longs are the same trade in mirror; leveraged-ETF assets in a single theme; the next concentrated AI holder large enough to move the tape when it unwinds.

21:38 5. Recognise the state-sponsored entry pattern — and price the incumbent for its third run

1. When a consensus says a commodity industry has permanently stopped being cyclical, treat that as the setup rather than the conclusion — and go looking for a subsidised entrant, because that is what historically restarts the cycle.
2. Check the base rate for the specific industry. In memory it has now run twice: US → Japan in the 1980s (Intel had "75% market share in the DRAM industry" and was pushed to the brink), Japan → Korea in the 1990s ("by the 2000s you had the Japanese companies being driven out of business"). Both entrants were state sponsored, both were "several generations behind," both caught up.
3. Ignore current share; measure the slope. CXMT at "about 8% of the global DRAM market… is nothing, but it is ramping incredibly fast."
4. Take the entrant's own published capacity targets as the forecast to test — wafer starts 300,000 → 500,000 by end of next year; YMTC intending to be "bigger than Samsung or SK Hynix in the NAND business by the end of next year." Then ask what the incumbent's price assumption is worth if those targets are simply met.
5. Track the quality gap separately from the volume gap, and apply the good-enough test: high-bandwidth memory "yielding very very well… only a generation behind" is sufficient wherever the application is not bleeding-edge — "sometimes a Ford will work just as well as a Ferrari."
6. Score the entrant's motivation, because a strategic motive removes the price discipline a commercial one imposes. Cut off from US chips, "having their own semiconductor supply is as important as having an aircraft carrier or nuclear weapons… more of a defense technology in some ways."
7. Scale the threat to the entrant's capability, not the last one's: China dwarfs Japan or Korea in GDP, population, land and the ability "to put up these massive factories."
8. Convert to a sequencing view: if the entrant is credible, the commodity leg of the theme can break before the theme does — "I'm actually starting to wonder if the semiconductor bubble breaks before then."

• CXMT/YMTC wafer-start and yield disclosures against their stated targets; Chinese HBM qualification wins at real customers; commodity DRAM/NAND spot prices rolling while the shortage narrative persists; incumbent capex announcements that assume the shortage; export-control changes that either slow the ramp or harden the strategic motive.

45:54 6. Construct the portfolio from downside protection outward, and flex the short book

1. Start from the loss function, not the idea list: "if you go down 100%, doesn't matter if you're up a billion percent, you still are wiped out." Asymmetry of ruin sets the constraint before any name is chosen.
2. Make gross and net exposure the first decision, driven by your read of the environment: "if I think things are massively oversold, I may have almost close to no shorts. And then if I think we're in a period of time like we are now… I may have more shorts on than longs." The short book is a dial, not a fixed sleeve.
3. Run one research process for both sides. Screening for longs generates shorts as a by-product: "I thought this was a long but actually it's probably better off as a short because it is doing AI but it's actually probably one of the guys that's going to get displaced."
4. For any AI-touching name, ask the displacement question explicitly — is this company using AI, or being consumed by it? Same evidence, opposite conclusion.
5. Grade the private-market rivals on hard financials, not narrative: Anthropic "turned profitable in Q2," OpenAI "lost even more money in Q2 relative to Q1." Relative profitability trajectory decides which side of a pair you are on.
6. Layer the macro/credit picture on top of stock selection rather than beside it — "the part that a lot of individual investors, or even some of the quote unquote fundamental investors, don't focus on enough is what's the big picture? What's the credit picture?"
7. Accept that the framework is personal and say so: "Stephen Curry likes to score points a lot differently than Shaquille O'Neal used to." A process you cannot execute under stress is not your process.

• Your own net exposure drifting with the tape rather than with the view; a long thesis that survives only if the company's AI supplier stays a supplier; software's relative underperformance (4% YTD vs 60%+ for semis) closing, which removes the margin of safety; the "safe space" argument becoming consensus and therefore priced — as he says security already has.

53:18 7. Stack independent calendar odds instead of arguing a single factor

1. Refuse the single-variable debate on principle: "there's never one factor that you're looking at… It's not any one of those things. But you put them all together."
2. Collect factors that are independent of each other and each carries its own base rate. His four: monetary, seasonal, political and geopolitical.
3. Monetary — start from "don't fight the Fed," then read the chair's own words as the guidance: Warsh tying "65 months of sustained elevated inflation… squarely with the central bank" is a pre-announced hike. Then use the calendar to place it: two meetings (Sep 16, Oct 28) and a Nov 3 election, and no chair hikes a week before a vote. Conclusion: September 16th, then a pause.
4. Seasonal — use long, stated samples rather than impressions: since 1957, September is the only month down on average and the only month more likely down than up.
5. Political — quantify the midterm overlay: peak-to-trough losses of 10% from July 31st to November 9th since 1990, versus 5% in non-midterm years.
6. Geopolitical — reason from precedent about incentives, not headlines: Iran held the hostages until Reagan's swearing-in, so expect "flare-ups right up until the midterms" because "if you end up with a change in the political climate, that's much better for Iran."
7. Add the theme-specific political risk on the same clock — the data-center backlash is bipartisan into the midterms (Gallup: 71% against data centers vs 53% against nuclear) and structurally short-lived.
8. Combine with valuation and express it as odds, not certainty. The poker frame: "when you know you're going to lose seven out of eight hands, it becomes very crucial on focusing on how much do you lose on the seven hands."
9. Put an expiry date on the bearishness so it does not become a permanent posture: negative "between now and the midterms on November 3rd," constructive for "at least another year" after.

• The September 16th decision itself (the single highest-conviction, checkable call in the hour); market pricing of the hike moving toward his view before the meeting; polling shifting far enough to change the midterm assumption; the data-center issue surviving past November, which would convert a short-term risk into a thesis risk; Iranian escalation timed to the vote.

55:52 8. Audit your own conviction for survivorship bias — and check it against a live loser

1. Whenever a "just hold it forever" argument appears, ask what the denominator is. The winners get told; the failures do not: "you never have people on that say AOL was a buy and hold, Yahoo was a buy and hold, Nokia was a buy and hold, Cisco was a buy and hold" — or IBM.
2. Count the survivors honestly. Microsoft has "done great through three different decades… But that's one company." One survivor from a long list is a base rate, not a template.
3. Adopt the posture explicitly: "strong conviction but loosely held" — "as the facts change, you want to change, too."
4. Test the discipline on your own book, not on a museum piece. He volunteers Nike unprompted — "one of my big disasters this year" — and diagnoses it plainly: lost share to newer brands, "years of mismanagement," a former market-share leader that being cheap does not rescue.
5. Separate brand durability from financial durability. Disney is the archetype — "you've got to put it away for your grandkids. Well, that hasn't worked out so well."
6. Turn the humility into a position rule rather than a sentiment: because "you don't know if you're holding the next Google or… the next Yahoo," size for the possibility that you are wrong.
7. Ban leverage at the personal level, with the concrete failure mode: margin plus leveraged ETFs on top of margin destroyed "millions of retail accounts in Korea." "You never want to put your family at risk."
8. Do the inoculation exercise before believing that smart operators must be right: pull the primary document from the last cycle — Cisco's own earnings release going from +70% to −30% bookings growth in months (25:32).

• A holding you have stopped re-underwriting because the brand feels permanent; a thesis whose evidence is a single surviving analogue; leverage anywhere in the personal account, including leveraged ETFs bought on margin; the moment the facts change and the position does not.