Jay Singh — research hub

Stock & name index

▲ Positive

► Neutral / referenced

▼ Negative

Overall thesis

• Bracket a sale rumour between the sponsor floor and the strategic ceiling — and buy after the pop fades (SEP-20). A credible "exploring a sale" report (named banks) is the trigger; the floor is what a leveraged buyer can finance (peer-multiple EBITDA less net debt, made likelier by pre-IPO PE holders), the ceiling a synergy-paying strategic, and the base case sits between. WAY: bought at +6-7% after a +12% pop, 14× ≈ $31-33 floor, 16× ≈ $37 strategic, base "low 30s to mid-30s," likeliest buyers Optum, CVS, Veritas, Thoma Bravo.
• Accumulate a commodity producer in its weak season, with the bear case priced first (SEP-20). A de-levered producer at its lows on a seasonal scare, a double-digit FCF yield, and three cases on conservative inputs with the up/down ratio written down. EXE: 15% FCF yield, bear $80 / base $129 / bull $152 on $70 oil and $3.34-3.75 gas — "up seven down one" — a 20bp starter "and we'll continue to add to it through March."
• Price the hidden segment at its pure-play comp's multiple — and accept less upside for less volatility (SEP-20). NOK's optical unit at Ciena's ~36× forward EBITDA alone implies ~$13 against $10.68; the stable legacy business is what keeps it from falling 20-30% in an AI sell-off. Separately, ETN shows the inverse discipline: a great business with a 2028+ revenue pipeline gives you time to wait for the stated entry (under $400).
• Own a cyclical windfall as insurance on its duration, not as a long (SEP-20). Ask how many years of windfall the price assumes (tankers: one) and what one more year pays (debt gone plus 20-30% one-off dividends; legacy contracts re-pricing even if spot stalls). STNG at ~33% FCF yield on a 2% dividend, FRO, IMPP — "I think of tanker stocks as basically a hedge to the portfolio," preferred to oil because oil can be managed by SPR releases.
• When stocks are the most expensive versus bonds in 26 years, rebalance into discounted income (SEP-20). Screen CEFs that moved from NAV premiums to discounts and compare yield, leverage and duration against TLT and munis. PDO / PDI / PAX at ~5% discounts paying ~13%, plus RWT / RWTQ — "where you can earn 13%, without taking a boatload of risk" — with long-end rates judged "close to their peaks" and equity downside "maybe 10%" at ~19× forward earnings.
• When the rate call flips, hedge with the asset priced off the part of the curve that will move (SEP-13). A hike can lower the 10-year on growth fear while raising SOFR — so the hedge for a Fed he now expects to hike is not a duration trade but the borrowers priced off the policy rate. "Commercial real estate is not valued on the back end of the curve like mortgage rates for residential are. Commercial real estate is benchmarked off SOFR." Screen the exposed sector for concentration against capital, thin fee income, lagging reserves and the quality of any de-risking; then size the squeeze risk before shorting. VLY: ~$30bn CRE in a ~$51bn book, ~329% of risk-based capital, 13% fee income, short interest only ~9% — "this is a short that I can put on and not be as scared," target 20-30% down on a rate scare. The index-level version: long the S&P SmallCap 600 (earnings screen), short IWM (40-45% unprofitable, 30-45% floating-rate debt).
• Own a listed wrapper around a private asset only while the private asset is free (SEP-13). Value the operating business neutrally, subtract it from the wrapper's market cap, divide by the stake to get the implied private mark; buy near zero, sell when the implied mark reaches the most optimistic credible valuation — and pull the exit forward when the private asset's catalyst weakens. SKM bought near $19-20 as "Anthropic optionality for near free," sold at $38-40 when it priced "a neutral core telecom at 11 billion… plus Anthropic at a private mark of 1.7 trillion," the weekend the Amodei letter made an October IPO less likely — "if you need to buy it cheaper, you can always buy it cheaper."
• Buy the forced selling from a small issuer's convertible — size the hedge flow against the market cap first (SEP-13). Convert size × initial delta (0.5-0.65) = stock that convert-arb buyers must short; against a small float and a single session it is mechanical. Confirm the purpose (refinancing at a lower coupon), grade the side-news, look for a concurrent buyback, apply a management filter, and add in steps sized for the residual risk. RWT: a $150M 7% convert meant "75 million to 100 million" of shorting into a ~$500M cap, −22%; bought at 3.41-3.45 and 3.21 (average under $3.60) plus ~11% baby bonds, "you can't have too big of a position here because interest rates continue to be a risk."
• A spread that only works with borrowed money is not the house's trade (SEP-13). Convert every arb to an IRR on a realistic close, note that levered funds set the clearing spread in large deals, and require an unlevered return that pays for the break risk on its own — which pushes the book toward small, contested, two-bidder situations. ROKU/FOXA passed at 6.5% ("I'm not going to borrow money to put on a silly spread like that"); RSVR bought in the low $9s under $10.50 with a $7.50 unaffected floor and a rival bidder; "most of our merger arb exposure is in much smaller names."
• Classify a scary headline as sentiment or fundamental — and write the reclassification triggers in advance (SEP-13). Read the primary document; if it changes no numbers it is sentiment risk. Compare speakers' incentives ("Altman can afford to sound cautious because OpenAI is not trying to price an IPO next month, but Anthropic is"), and list the observable events that would make it fundamental: "an actually delayed model, a push compute deployment, lower accelerated demand, or a company guides citing lower AI spending. Until then, do not confuse uncomfortable headlines with a capex cut that nobody has actually made." Position for the second-order funding chain (labs → hyperscaler backlogs → capex → GPUs and memory) through the most exposed proxy, not through the headline.
• Strip the theme out of a cheap thematic name — buy the company, short the ETF (SEP-07). The cleanest new method of the autumn, and the one that makes the whole "add event-driven names that are less volatile" mandate executable. When a stock is cheap because its sector sold off, you are being asked to take two bets at once; separate them. Establish that the company's own results moved up while the price moved down (a beat plus a decline is a flow event; a miss plus a decline is information), identify the thematic index it is being sold with, and short a liquid ETF tracking it — sized to the beta you want removed, not to the dollar value of the long. What remains is the only question worth answering: does this company outperform its own theme? HIMX is the case: revenue $227M, +14% q/q at a 33.1% gross margin exceeding guidance, four quarters of declining growth snapped, Q3 guided to $243-250M (+23.8% y/y at the midpoint) with margin to 34% and flat inventories — "its last quarter beat and its stock is down simply because of the AI sell off… the market doesn't seem to be rewarding this strong beat" — and the idiosyncratic residual is nameable: 40% global share of automotive display driver ICs, an auto cycle that "has bottomed," and "all of the Meta Ray-Bans use Himax" through the WiseEye edge sensor. The caveat is built into the method: only run it where the company-specific case is genuinely idiosyncratic, because hedging out the beta of a name whose only merit was the beta leaves you holding nothing.
• Distinguish a guided timing dip from a guided growth impairment — and buy the first (SEP-07). Read the stated reason for a guide-down word for word: "customer spending timing," "lumpiness" and "pull-forward" are calendar language; "pricing pressure," "share loss" and "demand softness" are not. Then check the reason against the segment's structure (a high-margin business selling to a few large clients on short contracts will be lumpy — a known property, not new news), find where consensus has been reset to, build your own full-year number off the segments, and bracket the multiple between what the growth half would fetch standalone and what the declining half would. EVC: management guided ATS revenue down sequentially for Q3 on "normal customer spending timing," "this pause in EBITDA growth resulted in a big sell off in the stock," and consensus went to $18-25M a quarter — against a segment that had just done +230% revenue to $182M and +673% operating profit to $40M. On his own $100-115M of 2026 adjusted EBITDA the stock is 7.5-8.5× forward versus 12-12.4× trailing, and the bracket is explicit: "at eight times, EVC is priced at a discount compared to pure play ad tech platforms which often trade above 12 to 15 times… but at a premium to legacy radio and broadcasters which trade at only four to six times." The catalyst must be mechanical and dated, not a re-rating you wait for — here, political advertising through Spanish-language stations in the states that decide the midterms, landing in H2 of an election year at very high margin.
• Grade an arbitrage against the unaffected price, and say out loud when it is finished (SEP-07). Find the last price before any leak — that is what the business is worth with no buyer, and it is the denominator for everything. Estimate the plausible deal price from the reporting and the rejected offers rather than from where the shares trade (which already contains the market's guess), then write the ratio: dollars of upside against dollars back to the unaffected price. Four spreads at four stages this week make the point. VRNS: Thoma Bravo (via Proofpoint) in advanced talks, unaffected 41, high 48, settled 46, plausible bid $50 — "from the lows, this was trading at 20 back in March… there's probably not much upside from here." LPSN: the opposite end — "a microcap spread which we talked about in the Discord, which we thought was an easy layup… bought in the 280s… rallied to 316… almost like a two-day trade," wide because funds cannot size a $3 stock, not because the deal was risky. ATAI: pure carry into a September 9 close. WBD: finished — "the largest spread right now, given the compression in Warner Brothers-Paramount, is NSC/UNP." The book rotates to where the ratio is, not where the conviction is.
• When the input-cost regime reverses, sell the sector — and date the reversal off the political calendar (SEP-07). A "cyclical bottom" trade is a bet on the spread between input cost and output price widening, not on volumes and not on the company. When the raw material reprices faster than the finished product, the thesis is invalidated however well the business is run — and where the input price is set by geopolitics rather than supply and demand, the timeline is an election calendar. LYB, in Q&A: "we were effectively timing a cyclical bottom. Since then, chemical names rallied quite a bit. And because of rising energy prices, I am not really that positive on the sector anymore. So what I would do for LYB is just trim or sell." The catalyst test is applied and fails on a date: "I actually think that we could see an escalation of this Iran war before it gets better after the midterms. There's no immediate catalyst for peace." And the conclusion is generalised deliberately — "names that are sensitive to oil prices and gas prices, I'm just not as excited about" — with the mirror image acknowledged in the same call: record $106/bbl diesel cracks are "very good for refiners right now." Note also the discipline of saying the action word: "trim or sell" is a decision; "less constructive" is a mood.
• Take the side of the Fed trade with the political constraint — and price the path, not the level (SEP-07). Read the prediction market as the clearing probability you are being paid to disagree with, collect the sell-side distribution rather than one house's call, and note why each changed (a forecast revised on one speech and one payroll print is a weaker input than one revised on a trend). Then look for the constraint the economists' models do not carry. Kalshi 52% hike / 48% hold into September 16, with UBS wanting September and December hikes, Barclays two, Morgan Stanley a hold on core inflation falling to 3.1% — and Singh under all of them: "we think that probability should be lower for a hike… I'm less hawkish than UBS obviously," reading the chair as managing a room rather than committing to act ("Warsh is just playing this game because there are 12 voting members and he doesn't want to look like he's clashing against them"). The hawkish payroll input is discounted at source — ~77% of the 162,000 jobs were leisure, government and healthcare, "not as strong as the numbers seemed." The settling datapoint is named and the conclusion withheld until August CPI. And then the path, which is what equities actually trade: "if there's just one hike, the S&P should do reasonably well after a couple weeks of uncertainty. But if there's more than one hike, things can get a lot worse in risk assets."
• The 40-year bond bull market is over — and "a return to normal" is the wrong frame (SEP-07). The structural call that now sits under every duration decision in the book. Yields "are back to levels we haven't seen since the early 2000s, but we can't call those 2000 level yields normal given 2000's yields were well below the 1990s and certainly well below the 1980s." Each half-decade for 45 years delivered lower yields than the prior one — "until now" — and the 2020 low of 0.5% (from a 15% apex in 1981) "marked the bottom and the end," made official by "the breakout of the prior downtrend in 2022." What that regime was quietly paying for is the useful part: from 1981 to 2020 "yields were lower at the end of the seven years than at the beginning 90% of the time," so every refinancing lowered the cost of capital — which "allowed for many mega trends," including financial engineering, the private equity boom, rising multiples, and "an era of minimal consequence for governments to keep increasing public debt." The bill is arriving: debt service at an annualised $1.2 trillion and 19-20% of tax revenue ("once you get above 30%, I think people will start panicking… once you get to 50%, you have to really depreciate the currency dramatically"), a four-year PE distribution drought that is also "why private credit is struggling," and the subtlest observation of the call — $300-400B of interest thrown off old deals with no new LBO supply to absorb it, "which is one of the reasons why this cash flow can be reinvested in all these new AI bonds, because there's nowhere else to put it." The ceiling is stated too: "bonds in the US are effectively reasonable here… we don't expect yields to go much higher without inflation going higher."
• Size a credit scare before acting on it — and separate the liquidity problem from the solvency one (SEP-07). The discipline is applied against his own bearish datapoint. CCC spreads have broken out for four months to 13% (JPM) to 15% (BofA) yields — and then: "remember that's only 10% of high yield. So if high yield's 1.5 trillion, this is 150 billion. And if you see defaults even at 10% of 150 billion, you're only going to get 15 billion, which is nothing… high yield defaults meant a lot more in 2008 versus now, because the equity market's just so much bigger than the high yield debt market." Gating is graded separately, because a redemption cap says nothing about credit quality and everything about the mismatch between illiquid assets and liquid promises: BCRED (a 5% cap on a $77.2bn fund after 10% of requests), Cliffwater (16%), Blue Owl (20%), BlackRock (9.3%), Morgan Stanley (10.9%), plus Ares and Apollo — "so the private credit risk has not gone," with the cause traced upstream to the PE distribution drought and the alarming counterpoint that the strategy still raised $99 billion in Q2. The household metric that does worry him is separate from all of it: subprime auto 60-day delinquencies at a record 5.2%, more than double four years ago. And the equity trigger is a level, not a mood: "JP Morgan says a 5% 10-year treasury yield could be the level that starts to pressure stocks. So we're only 20 bips from there right now."
• Insider filings as a two-sided screen — and a conflict is a research task, not an average (SEP-07). The origination tooling is widening from corporate events to insider behaviour: "we started to track in our tracker Form 4s and we're working on prioritizing which ones are good and which ones are bad. That screener is supposed to be out in October." Weight a purchase by role and by size relative to that person's own history — VST's CEO James Burke bought 2,000 shares at $135 ($270k), "his eighth largest trade out of 13" — and weight a sale by the proportion of the stake rather than the dollar amount: CAT's CEO "sold 48% of his CAT stock… for people looking at shorting CAT as a high valuation cyclical name, the fact that the CEO is selling half his shares might mean the company might have peaked in the short term." Flag unusual mechanisms explicitly (an Okta director's purchase "flagged as an unusual 10B5-1 buy signal"). And where the signals conflict, refuse to net them: TEM drew a Cantor overweight with an $80 target in the same week it appeared among the largest insider sales — "I have to look at that."
• Read a big assembler's earnings call as a survey of everybody else's supply chain (SEP-07). Companies that buy from everyone are a census of the industry, not a statement about themselves. DELL beat massively ($47B vs $44.9B, +58% y/y; EPS $7.00 vs $4.90, corroborated by HPE at +34%), but the value is in the call: constraints "remain the same across DRAM, NAND, spotty CPU shortages, disk drives, power integrated circuits, microcontrollers, drivers, ABF substrate, T-glass and optical" — nine unrelated categories, which is a systemic capacity problem no single fab fixes, and therefore also a timing warning on everyone's revenue schedule. Map each constraint to its beneficiary and test whether the scarcity is contracted or spot: MU, where Samsung has now locked 70% of capacity through 2031 into long-term contracts (after SK Hynix and Micron) while HBM3E 36GB spot runs at 4-5× contract pricing and DRAM export prices are +36.6% May-to-July even as volumes fell 13%; and AVGO, guiding AI revenue $58B → $115B → $230B by FY28 on 20 gigawatts of committed compute. The binding constraint is upstream of all of it: "AI growth is now being constrained by power… a potential 15 gigawatt shortfall just by next year alone," which is the frame under the VST insider buy.
• Buy operating leverage at a discount to private-market value — then date the trigger and admit if you are early (SEP-07). Quantify the leverage as a unit sensitivity rather than a narrative, value the assets against private transactions rather than listed comps, check that the asset base is being monetised in ways that do not need the cycle to turn, and read the payout policy as part of the option. WY: "a $10 MBF change in lumber prices adds about 50 million in annual EBITDA," $10/MSF of OSB adds $30M, "single family home construction requires roughly three times more lumber per unit than multi-family," the stock is at a 30-40% discount to private land NAV, land-solutions EBITDA guidance is raised to $450M (solar, wind, carbon capture, higher-and-better-use sales), and a 3.5% base dividend plus a variable one targeting 75-80% of free cash flow means "the dividend could go up dramatically by this formula if the EBITDA were to inflect." The discipline is what makes it a method rather than a pitch: he tests the trigger against his own macro — mortgage rates "could easily go above 7%," private construction outside data centres declining — and concludes against his own enthusiasm: "I'm not as bullish on this one… we're early on Weyerhaeuser."
• Harvest the odd-lot exemption — the one structural edge sized for a small account (AUG-30). The most directly usable idea the product has produced. Screen issuer self-tenders rather than third-party M&A (no antitrust risk, no financing condition, no vote — the only variable is price); require a modified Dutch auction with a stated collar, which turns the trade into a bounded distribution you can price before entering; read the document for the odd-lot priority provision exempting holders of 99 shares or fewer who tender in full from proration; verify the cash is already received rather than projected; and replicate the position across brokerage accounts, since the exemption is per holder per broker. ABUS: a $230M modified Dutch auction at $5.00-$5.75 funded by the $178M Moderna patent settlement, in a name he is long "from very low levels, like $2 a share" — "shareholders owning 99 shares in every brokerage… who tender their entire position are exempt from proration," worth 43¢ a share at a $5.50 clear off a $5.07 entry, or −1.8% to +12.9% across the collar, expiring September 29 and settling October 2. "The final clearing price is unfixed… but at least you know what your downside is." Found by the in-house screener, now slipping to an end-September / early-October launch with valuation metrics, real-time SEC EDGAR descriptions and value drivers next.
• Buy the activist's setup, but publish your own number — and make it much lower than theirs (AUG-30). The week's most disciplined act was marking down his own upside. On PRTH, Steamboat and Buckley (4.1% combined) are pressing the special committee toward a published $17-19 against Tom Priore's $6.00 bid at ~5.5× cash flow — and Singh's answer is "in reality, I think they'll get to 7 and a half to 8." He underwrites the fall rather than dismissing it (B2B payables revenue +22% while gross profit fell 10% on interchange and mix, SG&A +21%, EBITDA guidance cut, 3.8× net leverage — "that's why the stock was down so much, not for no reason"), and locates the leverage in company law rather than in the argument: under Delaware's MFW framework, a controller's bid lacking a strict majority-of-the-minority condition leaves the special committee as the only defence, so raising the price is cheaper than post-closing fiduciary litigation. The general rule: read an activist letter for its method, not its output — the published number is a negotiating anchor, not a forecast.
• Credit is now pricing what equity ignored — and a widening spread on a company nobody expects to default is an information event (AUG-30). "Hyperscaler CDS spreads hit all-time wides… hyperscalers have 3.1 trillion of off-balance-sheet arrangements, not all of which are actually guaranteed" across Meta, Google, Microsoft, Amazon, Oracle, Nvidia and Broadcom — and it does not mean-revert, because "these spreads could calm down if they stop issuing more and more debt, but I don't think that that is the goal." Goldman's Delta One desk supplies the transmission: credit "is starting to ask questions that equity largely ignored," plus inventory builds in the semi chain and power bottlenecks and policy pushback on data centres — "none of these kill the AI story, but the range of outcomes is clearly getting wider = lower multiples (SOX 24-month forward P/E from 21-22 to 15ish)." The proof: Nvidia guided 70% against 47% consensus and the semiconductor index still fell 2.3% on the week. The companion discipline is to write down the financing contingency inside a supplier's guidance — part of Nvidia's guide "is contingent on financing for Anthropic doing an IPO in October and OpenAI securing financing," with Nvidia's own CFO putting balance-sheet-supported AI-lab demand at ~25% of next year's business. His qualifier is the model: "it's not the base case, but it's still a big uncertainty."
• Strip the unrealised marks before comparing an earnings season to history — and prefer free cash flow when the two diverge (AUG-30). Headline GAAP EPS growth of 118.5% against adjusted growth of 31%, because of "how big Anthropic's mark-to-market was for Amazon and Google's P&L": excluding Amazon alone takes consumer-discretionary EPS growth from a record 92% to 7%, and excluding Amazon and Alphabet takes the index beat rate from 29% to 10%. Breadth is nonetheless real — ex-Mag 7 +20.7%, median company +14% EPS / +6% revenue, 10 of 11 sectors growing, record 16.9% margins. The cycle marker to track is non-operating income as a share of core operating income (~9% heading to low double digits, versus over 25% in the 1980s) — "a metric that bears watching, as it has historically risen in later-cycle environments." And the load-bearing sentence: AI capex accounted for ~70% of S&P earnings growth over three years, hyperscaler capex is on track for $700B+ this year (+90%) against consensus of only +30% in 2027, and "while earnings growth is going up, free cash flow growth is going down — and free cash flow, I think, is more important than earnings." The index is "likely over-earning today relative to historical levels."
• Read a central banker's hawkishness from what was removed, not what was promised (AUG-30) — and price bonds off the marginal change in nominal growth. Warsh gave no forward guidance, so the message was the framework: "in July, Warsh kind of hinted that we are waiting for a new inflation measure. He took that off the table for now by being strict about the PCE. That was probably to satisfy the committee, and that is what the market takes as hawkish." Odds went to 60-70% September / 90% December; Singh dissents on politics and seasonality (September is the poorest seasonal month, worse in midterm years) and reads the sequencing — "Warsh is buying himself time until the new committees… are in place," so a hike "would only be to satisfy the groupthink at the Fed now, to be able to deliver on the productivity side later." On the bond side the anchor is nominal GDP, not geopolitics: "nominal GDP growth has not been this high since the mid-2000s, which is also the last time the 10-year was this high. We do not view this as a coincidence." So Q2's 8% nominal against 1.5% real is a peak, and "on a marginal basis, if earnings growth is peaking this quarter and GDP growth is peaking this quarter, I'm less worried about rates going higher." The expression is 10-to-20-year municipal bonds in high-tax states on a taxable-equivalent basis, cushioned by 10-year futures positioning "about as net short as they've ever been in history… including people like Druckenmiller" — with the stated risk that "another string of hot inflation data, more likely given the rise of gasoline and diesel prices in August, could send rates even higher." Meanwhile Bessent's doubled buyback and the floated $950B TGA drawdown were dismissed as "either ineffectual… or unlikely," and it is "not really QE… issuing bills at the very short duration and buying bonds, long duration, which is effectively increasing the interest rate risk at the Treasury."
• The rate-peak call is withdrawn — trade the gap between the stated pain threshold and the firepower behind it (AUG-23). The week's controlling framework, and a reversal of the AUG-17 "temporary peak in the 10-year." Size an announced defence in the units of the market it must absorb: $4B per operation, maybe $15-30B across the window, against a market that swallowed $742B of Treasury sales in a single week, $97B of net long-bond issuance and a TBAC-flagged ~$1.45T funding gap across fiscal 27-28. JPMorgan calls it "a band-aid on a bullet hole"; Jefferies' Zervos calls it a Treasury-run Operation Twist that is not QE (bills swapped for bonds create no reserves) but still buys fiscal room, and notes the 10-year has held a 90bp range for 22 months. Singh's read is the BoJ analogue — escalate until you own half the market and let the currency be the release valve — "which is positive for gold, positive for silver, positive for real estate and agricultural land, and also positive for crypto." The tape agreed: GLD 370 → 420, bullion ~$4,700, silver ~$69, copper ~$6.60 ("copper stocks like HBM… which we are holders of"), and a three-sigma crypto move. Institutional overlay: Jeffrey Lacker's "Revisiting the Treasury-Fed Accord" testimony against Bessent and Warsh, arguing the Fed's third statutory mandate (moderate long-term rates) is served by killing inflation risk, not by buying bonds, and that any new accord must force a bills-only Fed — "which is what resulted in the market to second guess the ten year."
• Buy the controlling shareholder's lowball bid — the offer is the floor, the special committee is the option (AUG-23). The origination screen that produced the week's only new long. Look for a non-binding take-private from a holder who already controls the company (no rival bidder is possible, so it is a negotiation, not an arb), time-stamp the bid against the tape (a bid filed right after a self-inflicted print, anchored between the pre-drop price and the post-drop low, is priced by the buyer's timing), value the cash (levered FCF yield, then net debt to EBITDA), and count the objectors' arithmetic. PRTH at $5.52: Priore owns 55-58% and bid $6.00-6.15 at ~5.5× cash flow days after the stock fell $7 → $4.87; the business earns $91M of FCF on a ~$460M cap (~20% levered FCF yield) against ~$920M of net debt (~4× EBITDA); Buckley Capital published a $15-20 sum-of-the-parts and Steamboat joined — so downside is a small discount to a bid that "acts as a structural price floor" and "you could easily see a bump to a $7 or $8 offer" (25-50%). The same structure appeared twice in one week: JBS N.V. (82%) bidding for the rest of PPC.
• The consumer finally printed the slowdown, and the crowding-out loop explains the long end (AUG-23). Two facts that now bound the equity book. Consumer: Walmart −8%, its worst day since May 2022, on the slowest US comps in six years (2.6% vs 3.7%), with Goldman cutting second-half real spending growth to 1-1.5% and gasoline held high by crack spreads (global refinery runs −7 mb/d on Middle East and Russian outages while margins sit near records). Long end: the big-tech bond paradox — hyperscaler capex outgrew free cash flow, so it is funded with debt (Google's 100-year bond, Meta's and Oracle's 6-8% paper at 200-300bp over Treasuries), and "if you are a credit investor and don't think that Meta and Google are going out of business, why would you not buy their debt over US government debt?" The growth meant to solve the debt is financed by competing with it; the one clean circuit-breaker he names is ending the Iran war (lower oil → lower inflation expectations → Treasuries bid), and instead the US opened a Section 338 trade war with Canada (50% tariffs on ~$20B, no USMCA carve-outs, retaliation September 8) with "even more trade wars" promised after the midterms.
• Own the bottleneck, and read a pass-through as a power map (AUG-23). Nvidia raising AI-server prices 15%+ because it cannot absorb memory inflation is the cleanest statement yet of where the pricing power sits: "the inability of the industry's most dominant company to hold the line on prices… shows how much leverage makers of memory chips like Samsung, SK Hynix and Micron have," with NAND and DRAM "parabolic" — so a 75% gross margin is "likely a peak margin" and the hikes "add complexity" to a build-out already fighting power, permitting and labour. The suppliers confirm it with capital: SK hynix's surprise 30 billion buyback (JPMorgan overweight) and Micron's $10B Boise labs inside $250B of planned US spend. The model layer runs the other way: open weights now match the frontier on 70-80% of routine enterprise tasks (Zhipu AI's OX Alpha, likely GLM-6, possibly "another deep seek moment" if it can find the compute), and Anthropic's Fable 5 has plateaued at ~11% of its own corporate spend — "capability leadership" no longer converts automatically into revenue, with the October IPO about to price that question.
• Label every position as a holding or a trade — and say which out loud (AUG-17). The clearest structural tell in the public interview: the same buy list splits into two books with different exit rules. Long-term holdings get a valuation case and a multi-year horizon — Uber (UBER) at 68 on 12× forward with $10B FCF this year going to $13B in 2027 and $15B in 2028, AppLovin (APP) held underwater from a low-300s average cost for "easily back at 500 over the next 2 years," Shopify (SHOP), a little Netflix (NFLX), the gold miners, Redwood (RWT) for the 15% yield. Trades get a dislocation trigger and a scheduled exit: memory at four times earnings after the Korean leveraged-ETF unwind (SK Hynix, the DRAM ETF, Nanya) and the neoclouds off rising GPU lease rates trackable on Bloomberg (CRWV, NBIS — trimmed into a +34% print) were bought "just for a trade, not for long-term," with the cut-off already dated to November. The discipline is in the labelling: a trade that becomes a holding because it went up is how books blow up.
• The rally is rented, and the rent is due in November (AUG-17). Two of the three kill-switches are temporarily off — the 10-year (Oman/Iran side deal; Hormuz strikes paused before the midterms) and inflation (CPI in line, PPI final demand 0% m/m, September hike odds >60% → <40%) — leaving a "short-term Goldilocks" window he is renting, not owning: TLT bought for the first time in years, most of the QQQ hedge covered after the weak August 7 payrolls, gold and mortgage exposure added because "real rates have temporarily peaked." The third switch, a big war escalation, is the only thing he says would put the 10-year back up. And the plan is pre-committed: "I do plan to take some chips off the table ahead of November."
• Never use margin — ~20% cash, amplify with call spreads (AUG-17, the risk rule under everything). Stated as house policy at Special Situations Research, taken from Seth Klarman's Margin of Safety: "we almost never use leverage… about 20% cash at all times," with extra exposure taken through call spreads rather than borrowed money, because margin "puts all the power in the broker's hands — they can sell whatever they want." Even 1× leverage gets you carried out in a Japan flash crash, a VIX blow-up or a COVID gap; he was up 300-400% through COVID with none. The counter-example is Situational Awareness — Leopold Aschenbrenner's ~4×-levered $45B long-AI-infrastructure / short-software book (CoreWeave, SK Hynix, SanDisk vs Adobe) that gave back a 400%+ gain and sold its public portfolio to Citadel: "if he only leveraged 50%… he would have still been running billions today." That forced unwind is also what launched his own software longs.
• Buy the forced seller, and buy what the crowded trade must short (AUG-17 — the two origination screens). Forced selling: a small-cap index rebalance plus rate fear dumped Redwood Trust (RWT) regardless of price; he bought at "a three handle" on a ~17% yield after speaking with management (jumbo origination and securitization for 750+ FICO non-W2 borrowers) — ~$4.77 within weeks, still 15%. Hedging flow: because AI longs dwarf individual software names, funds hedging them generated overwhelming short interest and systematic selling in software — so he bought DDOG (cybersecurity his biggest overweight), WIX, NOW and SNOW from March at ~10× cash flow. Both screens find price dislocations with no fundamental cause; the catalyst in each case is a flow event, not an earnings event.
• Strip the mark-ups, then the index isn't expensive — if the earnings hold (AUG-17). S&P Q2 earnings growth reads +50% year-to-date, but "the real number is closer to 30%" once Alphabet's mark-to-market on its Anthropic stake is removed — still extraordinary against the 23% expected on June 30. Breadth is real (88% reported, 86% EPS beats, 76% revenue beats, all 11 sectors growing revenue, index revenue +15% off a strong base) while the magnitude is concentrated in ~5 names. Because earnings re-rated up faster than price, the forward multiple fell 23× → 20×: "the S&P is actually not even that expensive — if you believe these metrics are sustainable." The conditional is the whole call.
• AI capex: ~$1T to be raised, and the pivot is from chips to power (AUG-17). Ex-Microsoft (which rose 13% purely for guiding capex flat) "all of the hyperscalers are now burning cash flow" — Alphabet's $45B quarterly capex took its non-GAAP quarterly cash flow to −$5.85B on +24% revenue. Roughly $1 trillion gets raised for AI projects over the next year, Nvidia alone $500B to finance purchases of its own GPUs, Morgan Stanley models $1.2T of 2027 AI capex, and the backlogs are circular (OpenAI/Anthropic money raised in the same capital markets) — so "the numbers will continue to beat for the next couple quarters. At some point next year, the market will say, how long is this sustainable?" He still likes GOOGL/MSFT/AMZN, but they are no longer capital-light. The durable side is the recurring input: "they may not need to buy chips every single year, but they will need to buy power every year" → VST and co-location, plus HBM, grid connection, liquid cooling and Broadcom demand — against "a symmetric risk of the cost of overbuilding" with data centres running one-to-two-year delays.
• Put a date on the de-risking, and publish it before you act (AUG-16). The method that now frames the whole book: separate the reason a risk is quiet from the risk itself. Iran is quiet because of "a lack of munitions" and because the administration "doesn't want the war to escalate before the midterm elections" — a constraint with an expiry date, not a resolution. So the schedule is stated in advance: fully invested through the second half of August, September and October, then "after November, I plan to take some risk lower," with the tail named ("especially if he loses the House… there could be some retaliation"). The base rates size it as a hedging problem rather than a selling one: since 1990, end-July to Nov 9 in midterm years returns a median +0.9% with a median peak loss of 6.2%, versus +2.7% and −3.5% otherwise — so "add some hedges back on further market gains and get more selective." Everything else on the AUG-16 call sits inside that window: the soft-data trifecta (CPI and PPI below expectations, retail sales −0.6% vs +0.1%) put September hike odds under 40%, and the new money went to U (accumulate 35-40, base 55, upside 71), RIOT at $19.85 on a $9.1B / 109 MW Anthropic contract worth ~$15/share that does not start until December 2027, APP in the low $300s, and BTBT on WYFI's beat — while half of BW came off into a 40% pre-market gap and two new shorts (GBFH, DFNS) were opened. The income spine keeps winning the comparison it sets itself: plain-vanilla fixed income −1% to +2% YTD against a pref book up mid-to-high single digits, stress-tested by issuer quality (a 3-5% drawdown in the riskiest prefs at a 5.5% 30-year, less for utility names "like Duke Energy") on the view that Bessent "would effectively force Trump to pull back."
• The IPO wave was a top signal — and the SpaceX pair trade is now closed (JUL-26). SpaceX listed at ~$2.1T on a tiny float (a textbook "sell the news" that crushed the listed pre-IPO/space complex — VCX all-time lows, SPCE/PWRL unwound). The expression was owning the story cheap via EchoStar (SATS) at a 40–45% NAV discount hedged with a ~30% SPCX short into the unlock. SpaceX then fell further than EchoStar, the spread compressed as designed, and he took the trade off: "take your short-covering profits on SpaceX and sell Echo — you can just get out of the whole position." (Alphabet's $94B SpaceX stake, $80B of it sale-restricted, is now mainly interesting as the source of 87% of Google's record net income.)
• Circular AI financing → cascade risk, and now the pricing power is cracking. FCF is no longer real (SPV debt, capex past $1T): the $35B Broadcom-backstopped "Big Sky" Anthropic SPV routes its proceeds to Nvidia, lenders shown no financials; AI debt heads to ~$400B → $1T+ by 2028. The new tell: Microsoft moving Copilot to usage-based pricing and adding self-hosted DeepSeek v4 (~⅓ the input / ~1/7 the output cost of Claude) — a "watershed" stripping OpenAI/Anthropic of pricing leverage → short the pre-IPO wrappers (VCX, DXYZ) holding their secondaries.
• Data-center delays are confirmed — a sellers' market for live capacity. Jefferies/JPMorgan: ~50% of 2026 and ~80% of 2027 announced builds haven't broken ground (SemiAnalysis pushes back but agrees delays are large). The constraint is physics, not capital — transformer lead times 12–18→30–36 mo, interconnect queues 3–5 yrs, capacity-clearing prices +1,037%. So operators with secured power command premium lease rates (CRWV, NBIS, DGXX); the bubble has "at least a year" left.
• Private credit is the quiet fault line. Redemption gates spreading (Cliffwater 17%, B-CRED 10%, Partners Group, Blue Owl OTIC 40%) and $30B+ of Collateralized Fund Obligations (CFOs — PE-equity-backed bonds, no organic coupon, levered ~60%, stuffed into life insurers) echo 2006–07. Fresh marks: Thoma Bravo −$5.1B, Blackstone −$2B, take-back paper ~40¢; defaults ~6% (Fitch) and rising. The bag-holder is the annuity/policyholder (cf. SHIP at ~15¢).
• Turning positive on gold — and two new special situations. Gold −25% from the January highs with bearish positioning the highest since 2017 into a relentless central-bank bid → adding spot (PHYS, GLD), miners (GDX, SIL) and Alamos Gold (AGI), whose 19% earthquake plunge (at a minor mine; Island Gold, 60%+ of NAV, untouched) overshot a ~4–6% warranted hit (~65% upside to $50). Plus Five Point (FPH) — a California land bank mispriced as a homebuilder (EV ~$750M vs ~$2.5B real-estate NAV → a potential 3x) — and ServiceNow (NOW), FCF +36% while the multiple fell 60% (7% FCF yield, ~90% DCF upside, an AI enabler not a casualty).
• The Mag7 / hyperscaler capex-burn de-rate — and the memory cartel as its mirror. ~$5.3T of mega-cap value gone since October as the market punishes capex with unprovable ROI: it's all multiple compression + falling FCF, not weak earnings (Mag7 32×→22×; hyperscaler FCF ~⅓-trillion→~zero; only GOOGL's raises land, META gets hammered on raise rumors, ORCL's CDS/layoffs flag it worst). Add MSFT/GOOGL/AMZN gradually into the de-rate, avoid META/ORCL. Conversely own the bottleneck: TSM (sole Vera Rubin fab) and Micron (MU) — which out-earned every Mag7 name this quarter (85% margins) and signed 16 five-year take-or-pay SCAs with $22B of deposits and floor pricing above any prior peak, structurally breaking the 47-year commodity cycle (bear case: Chinese funds say memory has "no moat"). Open-source Chinese AI taking enterprise share (~15%→~45% in a year via "model routing") is the threat to the whole monetization case.
• Merger-arb & special sits. Latest (JUN-28): WBD–Paramount/Skydance (EU set to clear with remedies — Sky exits the Universal JV), VRNS ("selling itself" on PE interest), PRA (closing Friday, ~7% IRR), the S&P/Mobility (MGGL) spin (completed JUL-05), GME/EBAY (shorted eBay into the takeover hype, covered), CAR (a ~$650M Section-16(b) windfall from Pentwater), HHH (low-80s JPM SOTP). Prior book: CZR (Fertitta), MGM, ROKU/FOXA, LION, WBS, HUN/OLN, NSC/UNP, TWO/UWMC, the HONA spin; housekeeping across OGN/CCRN/VAL/ESPR/ZIM.
• The ballast. ~25–30% in 700 high-yield preferreds/baby bonds (10%+ yields, a third floating; PFLA the latest add) — plus a cash margin of safety, the long bond (TLT at a cyclical-peak ~4.5%/5%), a high-dividend mortgage/agency-REIT basket still being scaled (DX, RWT under $5, the RWTS ~10% baby bond, AGNC, and NLY after its raise to $0.75 = 13.4% — "if you believe rates have peaked, all of them are pretty much adds here") on the rate-peak view, hated value (CVS, Merck, Moody's/S&P, Datadog) and small AI tangents (T1 Energy, DGXX).
• "Bad news is good news" — until unemployment breaks 5% (AUG-09 — the regime the Fed risk resolved into). The rates channel that broke in July repaired itself on one data point. His stated band: soft data suppresses hike risk and lifts equities "unless it's terrible, or until unemployment goes above 5%. But right now, unemployment is only around 4.1%." The print delivered it — −23k payrolls vs +80k, hourly earnings +0.1% vs +0.3%, and 103,000 of revisions stripped out of May and June — taking September hike odds 60% → 40%. The action that follows is on the hedges, not the longs: cover ~25% of the short book into the repricing, because the longs were already added during the geopolitical fear that preceded it. Two corollaries he now runs on: read the two-month revisions as the real print (confirmed by JOLTS openings −178,000 to 7.36M), and treat the market as range-bound until the November midterms, after which 2027 earnings become the story.
• The quarter was flattered by mark-ups, and the payback has a date (AUG-09). Before believing an earnings season, strip out non-operating unrealized gains — value changes on private stakes that never sold a share. Alphabet's $9.11 of Q2 EPS was $2.88 operating and $6.23 mark-ups on Anthropic ($183B → $380B → $965B) and SpaceX ($800B → $1.25T → a $1.77T IPO, now −40%); do the same at Amazon and S&P Q2 growth falls 47.4% → 28.0%. The forward test is repeatability: Anthropic must raise again for more gains to be booked and SpaceX has already fallen, so GOOGL and AMZN GAAP EPS can decline in 1Q/2Q 2027 against an unrevised $405/share (+13%) consensus — "a dangerous dead weight for 2027 growth statistics." The real half of the growth is just as concentrated: MU alone is ~25% of the index's earnings growth ($25 of index EPS, 850% 2026 growth, 2027 capacity sold out with customers on 60-70% allocation), and it sits inside a credit-funded ring — NVDA's $250B OpenAI backstop → 40% of MSFT's cloud backlog → 22% of NVDA's revenue → 17% of MU's — where only Microsoft still generates free cash flow to equity and Oracle's CDS is the first thing to crack. Positioning follows the split: own the disciplined spender (MSFT), the bottleneck (MU), the AI-insulated software (DDOG), and watch Meta's ~400,000 spare GPUs as the one event that breaks the neo-cloud rental case (NBIS, CRWV).
• The Fed, not the war, is the market's real risk (AUG-02 — the framework update). The rates channel finally broke, and it broke on an institution, not a headline: five months of missiles never closed the VIX consistently above 17.5, but one FOMC meeting under presidential pressure took it to 21, the Dow −1,153, the 30-yr to 5.21% — with the dollar down against nearly every G10 currency while yields surged and gold up on the day, the signature of "investors charging a sovereign more while trusting it less." A war shock has an organ (oil/freight/rates vol); a credibility shock does not — "the Fed's credibility is not a sector, it is a denominator," so it prices as correlation. Practical rules: ~2% off the index per 10 bps on the 10-yr, accelerating above 5%; 4.90-5.00% is the risk-off trigger; oil below $70 is the only easy release valve; and watch the Japanese 30-yr above 4% (possible Fed swap lines to the BoJ). The falsification he'll accept: restored guidance, a hawk-owned hike, or visible Fed-vs-White-House daylight, followed by the 30-yr retracing and the dollar firming. Underneath it, the transmission mechanism itself is broken — the top 10% of earners do ~50% of consumer spending and are rate-immune (fixed pandemic mortgages, portfolio-driven spending, higher income on cash), so hikes squeeze only the borrowing 90% and the rate-sensitive sectors.
• Capex discipline is now the sorting variable inside the Mag7 (AUG-02). The cohort stopped trading as a cohort the moment the guides diverged: hold next year's capex and still print free cash flow with an external proof-point (Azure +43% past $100B run-rate, >30M Copilot seats) and you get +13% (MSFT); accelerate a fifth straight quarter in the profit engine and even a $220B guide and −$7.6B trailing FCF is forgiven (AMZN, +15%); raise the guide and guide revenue down and have no cloud backlog and you lose 10% (META, $784M of FCF). The cohort arithmetic still drags: ~20% index weight at −10% YTD subtracts 2% while the other 494 names return +12%. Own the disciplined spenders and the equal-weight; the bottleneck names (memory, power, neo-cloud capacity) stay the cleaner expression.
• Macro (prior framing). "Rates drive the world and the war drives rates." The June-18 Iran ceasefire was violated (IRGC hit a Singaporean tanker → US strikes on radar/missile sites → Iranian retaliation → fresh re-talks as futures opened) — a "wobbly" peace that keeps timing the market's close/open. PCE came in a touch below feared (4.1% y/y / 3.4% core); NY Fed's Williams sees it edging down (tariffs played out, Hormuz transitory, housing slowing) with the Fed holding 3.5-3.75% — but still far above the 2% target. Europe weakening (UK services PMI 48.7, France stalling), Chinese tech (KWEB) in a bear market, and private-credit redemption gates (APO/ARES at 5% caps on 14-17% requests) deepening. Buy gold/silver and quality on dips.