What We Told Hedge Funds This Week

A 24-year-old ran $225m to $>40bn in 20 months, got margin-called, and sold his ENTIRE public book to a single buyer.. taking a month of momentum with him. Korea supplied 3-400k mini-Leopolds as the accelerant, the Fed poured real-rate gasoline on it by saying nothing, and the market just paid Microsoft for an accounting trick while punishing Alphabet for honesty. The stupidest market I can remember, in BOTH directions..

Let me get this straight. We have had to endure a MONTH of full-blown mouth diarrhea on everything from memory to fuel cells, and why “the cycle is over”, from people with no clue.. and the proximate cause, per CNBC’s David Faber, is that Situational Awareness sold its ENTIRE public investment book, longs and shorts alike, to a single buyer before Thursday’s session. So it was Leopold who created the July rout? A roughly $40bn AI-thematic book, reportedly levered around 4x gross, long the exact names in our universe (SK Hynix, Micron, Sandisk, Nebius, CoreWeave, Bloom Energy..) and short software names like Adobe that ripped in his face, run by a 24-year-old, got margin-called into the best fundamental quarter the semiconductor complex has ever printed. And the entire commentariat mistook his liquidation for information. The fundamentals didn’t change in July. One guy’s collateral did.

And now we know who bought Leopold’s bags: Ken Griffin’s Citadel, per Bloomberg, which scooped up a big chunk of the AI stocks the fund was forced out of. The same Citadel whose founder was out publicly pushing for a rate hike DAYS earlier. So Citadel lobbied for the exact policy mix that squeezes levered long-duration books, and then bought the largest levered long-duration book on the street at a discount when it duly got squeezed? I asked on X whether I got that right, and I am, of course, just asking questions, as the kids say. But you have to admire the elegance.

Bloomberg reports the fund’s assets have slumped to roughly $10bn, down by more than half, and what remains is mostly the illiquid stuff, private stakes including an Anthropic holding valued around $5bn. So Situational Awareness is, for all public-market intents and purposes, blown up.. surviving as a venture book with a newsletter attached. Leopold’s July letter still closed with a PS telling investors this “seems like a particularly good time to add funds”. Liquidated and upselling in the same paragraph?

On the fundamentals, the adding part is probably RIGHT, which somehow makes the whole thing even more absurd.

The anatomy of a stupid market: one Leopold, 3-400k mini-Leopolds, and a subscription army

Of course, one fund does not create the worst momentum month in recorded history on its own. It needs help, and boy did it get help.

Korea supplied roughly 3-400 THOUSAND mini-Leopolds, i.e. retail margin accounts levered into single-stock ETFs to the tune of nearly 3% of the entire free float, roughly 4-7x the exposure of any other major market, who all got carried out feet first in the space of two weeks.

The Goldman prime book shows the biggest tech-book readjustment in the history of the data set. Layer a forced seller of size on top of a levered retail complex on top of institutional rebalancing after a record quarter, and you get July. None of it, and I mean NONE of it, had anything to do with DRAM contract prices, HBM order books or fuel cell backlogs, all of which went UP while the stocks went down.

Now for the sign-of-the-times part, and forgive me while I get salty. Huge, sophisticated LPs, the kind with investment committees, consultants, and 80-page due diligence questionnaires about your disaster recovery plans, handed BILLIONS to a 24-year-old with zero, and I really do mean ZERO, experience running a levered book. The arc is genuinely spectacular: $225 million at launch in late 2024, roughly $20-24 billion by mid-2026, up some 439% net through June. Columbia valedictorian at 19, a genuinely brilliant essay, a genuinely correct thesis.. and zero drawdown scars. Because make no mistake, the thesis was RIGHT. The book still blew up.

And here is the part that only experience can teach you, and that no investment committee apparently asked about. Leopold’s book was long hardware against SHORTS in software, which on a risk report looks like a “net protected” book.. 4x gross, but only probably around ~1x net, hedged, sleep well.

The problem is that long AI-hardware versus short legacy-software is not a hedge, it is ONE trade expressed twice. It is the same bet on the same theme, and when the theme unwinds the way July unwound, the correlation between the legs goes to 1 and BOTH sides lose simultaneously: the hardware longs cratered 35%+ while the software shorts got squeezed higher in the great de-grossing. Net exposure is a fair-weather statistic.

Gross is what kills you, and gross was 4x. That is the lesson you learn in your first proper drawdown, usually with someone else’s smaller money, not with $24 billion of pension capital. Being right about the destination is worth nothing in levered land if you cannot afford the journey.

And the whole circus is amplified by the subscription-research complex.. people with no experience running money either, pushing TA-based “momentum research” in both directions, drawing trend lines on the way up and drawing the SAME trend lines upside down on the way down.

In July they collectively decided the cycle was over because the chart said so, while every fundamental input said the opposite. This is what I mean by the stupidest market I can remember, in both directions. The euphoria was overdone on no new information, and the capitulation was overdone on no new information. Fundamentals have rarely mattered less over a three-month window, and that, at least, is an environment you can trade against.

Macro update: Warsh is disconnected from inflation reality.. and it probably lit the match

But there is an uncomfortable macro link as well, which has admittedly made me a bit worried..

Everything has felt disconnected from fundamentals for a while now, and the biggest disconnect of all does not sit in a levered AI fund or a Korean margin account. It sits at the Federal Reserve. Warsh has now delivered two full pressers of saying absolutely nothing, explicitly aiming for a central bank that plays a smaller role in markets (a goal I actually have sympathy for).

The problem is that if you run his “nothings” through a quantitative lens, the rhetoric keeps getting MORE hawkish, and we currently have the biggest disconnect between the Fed’s rhetoric and actual inflation expectations (and our nowcasts, btw) in the history of the time series.

Our nowcasts have US inflation rolling over HARD, the softest prints of the year queued up, and the Fed responding by letting real rates fly higher day in and day out. A central bank hawking into disinflation, by silence.

And that silence was not free. Levered long-duration books, and an AI-thematic momentum fund is about as long duration as equity gets, do not die of old age, they die of real-rate spikes.

The Fed’s say-nothing hawkishness sent long-end reals flying at exactly the moment the most crowded trade in the world was wobbling, and it probably triggered the blow-up of Situational Awareness and the mini-Leopolds alike (Goldman and JPM have reportedly been issuing margin calls across concentrated AI books, not just Leopold’s). The Fed supplied the match, Leopold was the kindling, and Korea was the accelerant.. and Citadel, apparently, brought the marshmallows.

My baseline remains that I struggle to see the Fed hiking this year, that inflation comes down hard if our nowcasts are to be trusted (they typically can be), and that the market’s “panic” over how to read a man who refuses to be read will fade. If you have an explicit 2% target, refuse to forecast, and STILL are not hiking with inflation above target, you are telling me between the lines that you expect inflation to fall.

I am happy to take the other side of the real-rate hysteria while the tourists figure that out.

Portfolio “In Focus”: The scalers.. Alphabet spends honestly, Microsoft spends in a footnote, Meta builds the new AWS

Which brings me to the reporting season, where the market managed to out-stupid even itself this week. We now have all four of the big scalers on the tape, and the divergence in how they ACCOUNT for the same arms race is the real story of the season, in my opinion.

Start with Microsoft, because the quarter itself was excellent. Revenue of 90bn, up 18%. Azure past 100bn in annual revenue for the first time. Commercial RPO at 678bn, up 84% YoY (remember that number). Quarterly CapEx of 41bn, up 70%. And then, buried before the outlook section, the trick: effective FY2027, Microsoft is extending the estimated useful life of its data centres and office buildings from 15 to 25 YEARS.

That flatters depreciation from here, and, more cynically brilliant, it reclassifies a chunk of future data-centre leases from finance leases (which count as CapEx) to operating leases (which magically do not). Poof.. headline calendar-2026 CapEx guidance falls from ~190bn to ~175bn while the actual spending plans are, by Microsoft’s own admission, UNCHANGED. Not one fewer GPU is being bought. And the market’s response to this cosmetic surgery? The stock ripped 7-8% after hours. They spent the same, hid 15bn of it in a lease footnote, stretched the depreciation on the buildings, and got REWARDED for it.

Now compare that with Alphabet last week. Alphabet raised its CapEx guidance honestly, from 180-190bn to 195-205bn, let free cash flow go negative for the first time in 54 quarters, and got absolutely hammered for its transparency. And…. while Microsoft is stretching asset lives, Alphabet is doing the OPPOSITE. The implied asset life on Google’s productive assets has fallen every single quarter since late 2024, from 8.0 years to 6.7. Their depreciation policy has gotten MORE conservative through the build, not less. Same race, same spending, opposite accounting direction. If you are grading earnings quality, and you should be, Alphabet’s earnings are simply higher quality than Microsoft’s from here: more of the true cost of the build-out is running through the P&L today rather than being deferred into the 2030s. Two-thirds of this CapEx is GPUs and CPUs on roughly 7-8 year lives anyway.. the 25-year assumption does not touch the part of the stack that actually gets obsoleted.

So let me say it plainly: Alphabet is my SCALER winner of the reporting season so far. Cloud growing 82%, backlog from 108bn to 520bn in five quarters (6.7x coverage of LTM cloud revenue), return math that clears its cost of capital roughly 3x over (a ~15% IRR and 23.5% steady-state ROIC per dollar of CapEx, per Daniel Bakalarz’s excellent standalone model that went viral earlier this week), and the most conservative depreciation trajectory of the scalers as far as I can judge.

The market sold the honest spender and bought the creative accountant. I know which one I want to own when the dust settles, and it is the one whose earnings I have to adjust the LEAST.

Meta, for completeness, sits in the (mostly) honest-and-punished bucket alongside Alphabet, just with less to show for it this quarter. Revenue beat at 60.8bn (ads at 59.3bn), but the EPS print of 6.18 missed badly against 7.14, dragged by a 2.4bn legal charge and severance costs, free cash flow shrank to a rounding error (784 MILLION, from 8.5bn a year ago), and they raised the FLOOR of the CapEx range to 130-145bn, which mechanically means 40bn+ per quarter through year-end…

But then read Thursday’s regulatory filing, because THIS is the real Meta story: almost 700 BILLION dollars of future spending commitments. 349.3bn of non-cancelable contractual commitments (third-party cloud, servers, network gear), plus another 347bn in leases that have not even STARTED yet, of which 68bn was added in July ALONE, with payments starting in 2027-28. Nobody commits 700bn of compute, on top of the El Paso gigawatt campus with BlackRock owning 80% of the vehicle, to serve their OWN apps. My reading is that Meta is building the new AWS.

They are locking in capacity at today’s prices because they are confident they can charge MORE than their contractual commitments if they re-lease that compute to third parties into a supply-constrained market. Zuckerberg is not overspending on Reels.. he is quietly standing up hyperscaler number four Ior five.. I have lost count by now), and the income statement pain of 2026 is the entry ticket. The market is pricing the commitments as a liability. I suspect they are an asset with a lag.

And right on cue, Amazon just reported as I am writing this, and the numbers are a gift to that exact thesis. AWS revenue of 42.2bn against 40.6bn expected, up 37% ex-FX versus 31% expected, the fastest growth in years. Group net sales of 200.6bn (beat), operating income of 27.5bn against 23.6bn expected, and a headline EPS of 5.75 against a 1.82 estimate, though do remember that number is flattered by the mark-up on their Anthropic stake, so grade the operating lines, not the headline. Q3 guided to 197-202bn. The read-through is simple: the incumbent cloud is growing 37% at a 42bn quarterly run-rate WITH a 364bn+ backlog behind it, i.e. demand still comfortably exceeds supply.

That is precisely the market Meta wants to sell capacity into, and precisely why the “capacity limit on CapEx” consensus is wrong. In Bakalarz’s words on Alphabet: “The risk was never the return. It’s the size and irreversibility of the commitment.” Granted.. but irreversible commitments with excellent returns are still excellent investments. The market just needs a couple of quarters to remember that part, and probably also a Fed that does not allow real rates to rise every single day.

My data is SCREAMING that the market is reading this wrong, and I think the turn comes within a few months (on hyperscaler credit). Whether the Fed also stops reading it wrong, I cannot guarantee.

Portfolio Update: Buying the remainings of Leopold, while waiting for the USD weakness

So what do we actually DO about all of the above? Let me start with the obvious one.

Everything “Leopold” held is, in my view, a good buy here. His public book reads like an overlap report of our own AI-bottleneck and electricity themes: Micron, Sandisk, SK Hynix, Nebius, CoreWeave, Bloom Energy. The fundamentals of those names IMPROVED during the month in which they were forcibly dumped, the forced seller is now gone (his inventory sits with Citadel, who are many things but not forced sellers), and the single largest technical overhang of the cycle cleared in one transaction. When the most informed buyer on the street hoovers up a liquidation, you generally want to be on their side of the trade rather than the commentariat’s.

HOWEVER.. and this is where we refuse to be cocky: MUCH damage has been done in July, both to the credibility of the theme and to the technical picture. Trend lines are broken, momentum models are max short, a generation of Korean retail has been carried out, and every allocator who got talked into an AI fund at the top will be selling strength for months.

That kind of damage does not repair in a week, whatever the fundamentals say. So we are buyers of the Leopold complex on OUR timeline, not the market’s: scaling in patiently, adding on stabilisation rather than lumping into the falling knife, and keeping a little powder for the retests. Waiting a little is a position too.

In the macro book, we are feeling BETTER about USD shorts here, and we keep that lean in the portfolio. The ingredients are all lining up: our nowcasts have US inflation rolling over hard while the rest of the world re-accelerates, the Fed is (silently) hawkish but will be overtaken by the data, and Warsh’s whole project of a smaller Fed footprint is structurally USD-negative once the real-rate hysteria fades.

Notice that we already got a weaker dollar out of the last presser DESPITE rising long-end reals.. The JPY leg looks particularly attractive, with speculative positioning (on a stronger USD) still stretched... And it is also why we added Barrick Mining to the book lately: gold and gold miners are the cleanest expression of the decoupling/debasement theme for the part of the portfolio that does not want to fight the AI tape every morning.

As long as the long end of the USD curve re-steepens, the USD can sell-off while the “carry unwind” contagion risk is very low, which in my opinion remains the case. I therefore think the USD can weaken without triggering a major “carry unwind” here.

The regime picture supports the same conclusion. Growth probabilities are finding a floor in the US, the rising-inflation probability has COLLAPSED, and liquidity remains an outlier support. That is historically a friendly cocktail for risk assets and an unfriendly one for the dollar.. the only thing missing is a Fed that acknowledges its own data.

In energy, we remain patient rather than trying to be heroic on oil. The Houston premium tells you everything about the panic bid for US barrels that never showed up this round, and the spread sits back around a dollar against $7 in March. Even if we have two straits worth of headlines, the physical market still refuses to panic, which is precisely why we neither chase crude longs nor dare short it into headline roulette.

Our energy risk stays where the structural story is: long European natural gas into a winter with the thinnest storage cushion in 15 years, and long the electricity build-out complex.

And a short word on the book itself: we are still decently UP on the year as of Thursday’s close (and it looks like we will get even more through Friday’s session). Staying green through the worst momentum month in recorded history, while holding a few of the exact names that were being force-liquidated, is a result I will happily take.

So, structured by asset class, our best ideas from here..

In FX: short USD remains the anchor, expressed against the EUR on the inflation-divergence story and against the JPY on positioning and an ever-steeper domestic curve. I have gotten less worried about a new energy spill-over to JPY and EUR here.

In Energy/Commodities: long TTF natural gas into the storage winter, long gold and gold miners (Barrick the latest addition) as the debasement hedge, and flat-to-agnostic crude while the two straits keep us busy.

In Rates: fade the real-rate hysteria.. we want to be received in the US front end on a Fed that will not hike into collapsing inflation nowcasts, with a steepener bias further out given the benign issuance outlook (no cliff-edge from Bessent in August, the TGA has behaved).

And in Equities: patiently accumulate the Leopold overlap, i.e. memory and the hardware bottleneck complex, own Alphabet as the quality scaler (and Apple as the “non CapEx compounder), keep the power-semi lottery tickets for the 2027 leg (e.g. Wolfspeed), and let Citadel do the heavy lifting on the entry point. None of this needs to be executed today. The whole point of surviving July is that we get to be picky in August.

No changes this week in my portfolio, and thankfully I didn’t fall for the panic earlier this week (as a lot of other people did)... We will see how things develop into next week, but I think most of the fog dissipated this week.

In the end, it was all just Leopold.. I wish him a great wedding this weekend. It has to be a nice change of topic for him, at least.

横向滑动查看完整图表