Jay Singh — Weekly SSR: a hawkish Jackson Hole and a Sunday-night Iran escalation, Nvidia's record $453B day handed back on chip tariffs and circular financing, hyperscaler CDS at all-time wides — and three specials you can actually execute: the ABUS odd-lot tender, PRTH revisited, and PayPal after Stripe walked

Arbutus is a small biotech that won a patent lawsuit against Moderna over the technology used to deliver mRNA vaccines into cells. The settlement handed it $178 million in cash. Singh has owned the shares since before that verdict, "from very low levels, like $2 a share, even less" — they now trade around $5.20.

What makes this week's news a trade rather than just good fortune is what the company is doing with the money: buying back up to $230 million of its own stock through a "modified Dutch auction." That means shareholders offer to sell at any price they choose between $5.00 and $5.75, and the company then picks the single lowest price at which it can buy the full amount — everyone who offered at or below that level gets paid that same price.

Buybacks like this are usually oversubscribed, so if you offer 10,000 shares you might only get a fraction bought. That is called proration. But almost every tender offer contains a rule that exempts very small holders — anyone owning 99 shares or fewer who offers their entire holding gets bought first, in full, no proration. That rule exists to spare small investors the cost of being left with an awkward stub.

The trade is to become that small holder deliberately. Buy 99 shares — about $515 — and tender them all. If the auction settles at the midpoint, you make roughly 43 cents a share; the deck prices the full range as a 1.8% loss if it clears at the $5.00 floor and a 12.9% gain if it clears at $5.75, over about a month. And because the exemption applies per account, "if you have 10 brokers, you can own 99 shares in each" — the same trade repeated ten times over. The offer expires 29 September at 5pm Eastern and pays on 2 October. The only real risk is that the auction clears at the bottom of the range, below what you paid — a known, bounded loss.

Worth noting where the idea came from: not from a bank or a newsletter, but from the special-situations screener SSR is building and plans to release to subscribers.

Priority Technology processes card payments for merchants, automates bill-paying for businesses, and rents payments infrastructure to other companies. Its founder and chief executive, Tom Priore, owns about 57% of it and has offered $6.00 a share to buy the rest and take it private — an offer made days after a weak quarter knocked the stock down. Singh bought a starter position at $5.52 last week.

The update this week is a downward revision of his own expectations, and it is the most useful part. Two activist funds, Steamboat Capital Partners and Buckley Capital, have between them accumulated 4.1% of the shares and written to the board's independent directors arguing the company is worth $17 to $19. Singh does not believe that: "In reality, I think they'll get to seven and a half to eight." He is buying a negotiation, not a valuation argument — and he is marking the likely outcome at a 25-45% gain from his entry, not a triple.

He is also unusually candid that the sell-off was deserved. One of the three divisions, business-to-business bill payment, is growing revenue while losing money on it: higher card-network fees mean each extra dollar of payment volume carries less profit, so the company had to cut its profit guidance even as sales grew. Overheads rose 21%. Debt is about 3.8 times annual profits. "That's why the stock was down so much — not for no reason."

What protects the buyer is company law rather than the business. In Delaware, where the company is registered, a controlling shareholder buying out the minority is presumed to be conflicted unless the deal is negotiated by a committee of genuinely independent directors and approved by a majority of the non-controlling shareholders. Without both protections, the controller risks being sued after the deal closes and having a court set a higher price. That makes it cheaper for him to raise the bid than to force it through. And in the meantime his own $6.00 offer acts as a floor under the shares.

Stripe, the large private payments company, looked at buying PayPal and walked away. The takeover premium came out of the shares, and what is left is a cheap, cash-generating business that Singh treats as a range to trade rather than a company to own forever.

The numbers are the case. PayPal produces about $7 billion of surplus cash a year against a market value that puts it below eight times next year's expected earnings — a cash yield in the low teens. It holds more cash than debt. Singh's summary of what that means: it "is still generating so much cash that it can effectively buy itself in less than ten years." He thinks a fair multiple is ten times earnings rather than eight or nine.

The reason it is cheap rather than mispriced is genuine: PayPal is slowly losing checkout share to Block and others, so the business is not growing much. His own model — published live to subscribers during the call — values it at $57 a share against a Wall Street consensus range of $70-85, precisely because he discounts what the business is worth in the distant future given that share loss.

So the discipline is mechanical rather than visionary. He bought at $51.65 after the bid collapsed, sold at $54.40, and it closed near $53. "We'll buy back close to 50 again. This is one trade you could do several times." The value floor is what makes the repetition safe; the share loss is what stops it becoming a long-term holding. His own verdict on it at today's price is flat: "PayPal is not incredibly interesting here."

This remains a merger arbitrage. Paramount Skydance has agreed to buy Warner Bros. Discovery; the shares trade below the agreed price because the deal might be blocked; and that gap narrows as completion becomes more likely. Singh was buying when the gap was 20%. "It's now only 7%" — most of the money has been made.

The week gave a positive and a negative. The positive is that Paramount is visibly working on remedies: it looked at selling Warner channels such as HGTV and the Food Network, decided it would not have to, and is instead considering offering to negotiate separate distribution deals with cable operators for different parts of its television business. Offering regulators a structural fix is what settlements are made of.

The negative is temperament. California's attorney general cancelled Monday's scheduled settlement meeting, accusing Paramount's leadership of leaking the substance of the talks and misrepresenting them — "demonstrating a lack of good faith." No new date has been set. And the states are reportedly preparing to demand that Paramount sell some cable channels and keep its film studio separate from Warner's, which is more than Paramount has so far offered.

For a position already 13 points into a 20-point spread, the shape of the remaining bet has changed. What is left is compensation for a narrower, more legal set of risks — and the counterparty just made the negotiation harder.

Navitas makes power-conversion chips. It has agreed to buy Claros, a company founded only in early 2025, for $232.8 million.

The problem Claros solves is physical rather than computational. An AI accelerator chip draws enormous current — thousands of amps. Traditionally the electricity is converted to the right voltage somewhere else on the circuit board and then travels across the board to the processor, and every centimetre of that journey wastes power as heat. Claros builds the conversion stage directly next to, and above, the processor. Less distance means less loss, and it lets you push more power into the same physical space.

That matters because AI data centres are now limited by power and heat more than by chip supply. Navitas says the deal more than doubles the size of the market it can sell into by 2030, to over $8 billion, and its chief executive frames it as helping "break the AI infrastructure power wall."

Singh's position is a watch rather than a purchase: the shares already rose 6% on the announcement, so "on pullbacks, this NVTS could be interesting." The pattern is one he uses repeatedly — a small company acquiring a genuine bottleneck technology is worth owning, but not at the price the announcement itself creates.

Broadcom designs custom chips for individual customers rather than selling one general-purpose product to everyone. Its biggest customer project is OpenAI's own inference chip — the processor that runs a trained AI model to answer queries, as opposed to the far more expensive process of training it in the first place.

This week OpenAI's internal testing showed that chip beating Nvidia's GB300 on two measures that decide data-centre economics: work done per watt of electricity, and how quickly it responds. At 700 watts, with deployment planned for later this year. The caveat is that it was compared against GB300 rather than Nvidia's newer Vera Rubin generation.

Singh's near-term reasoning is simpler and mechanical: Broadcom reports next week, and Nvidia has just guided to 70% revenue growth. "Based on Nvidia's outlook, you would think that Broadcom would beat and raise its outlook as well." Whatever demand is filling Nvidia's order book is filling Broadcom's too.

The offsetting fact he keeps beside it is a credit one. Broadcom is among the seven companies carrying $3.1 trillion of commitments that sit outside their published balance sheets, and Goldman's trading desk notes that Nvidia's and Broadcom's default-insurance costs have widened sharply as the market focuses on how all this construction is being financed. The equity is optimistic; the credit market is asking questions.

CrowdStrike sells cybersecurity software that runs on every computer an organisation owns and watches for attacks. It charges an annual subscription, so the number that matters is not quarterly revenue but how much new annual subscription revenue it signs — and this quarter that figure was a company record of $333 million, up 51% on a year ago.

That growth rate is the answer to the question hanging over all enterprise software: does AI make this product unnecessary? For security, the opposite is happening. Management pointed to a tailwind from AI adoption itself, because every AI agent an organisation deploys is another thing that can be attacked or misused, and it needs to be watched.

The rest of the print supported it — revenue slightly ahead, profits ahead, cash generation up a third, and full-year guidance raised. The shares rose about 20% on the week, and management called it the best quarter in the company's history.

Singh uses it as the anchor of a broader argument: the market spent this year assuming AI would eat software, and the companies that actually reported are showing the reverse. Security in particular is not a nice-to-have that AI replaces; it is the thing you have to buy more of when you deploy AI.

Salesforce sells the software companies use to track their customers — the definitive record of who bought what and who is being sold to. It was the single most obvious victim of the theory that AI would make business software unnecessary, and its shares had been priced accordingly.

Two things this week reversed that. The first was simply good results and guidance. The second was a partnership with Anthropic that lets someone working inside Claude actually perform Salesforce actions — updating a record, moving a deal along — without opening Salesforce at all. On the old theory, that should be terrifying: the AI has replaced the screen. In practice it is the opposite, because the underlying record still lives in Salesforce and is now being read and written to far more often. As Dan Niles puts it, AI agents will access software tools ten to a hundred times more often than humans do.

Salesforce is also changing how it charges to match: away from a fee per human user and toward fees tied to how much the customer actually uses the system and what it is worth to them. That directly disarms the fear that AI shrinks headcount and therefore shrinks seat licences.

The shares rose 23% the next day. Singh's framing: "a strong counterpoint to the SaaS-pocalypse worries."

Okta manages who is allowed to log in to what. It sits between an organisation's employees and the dozens of applications they use, handling passwords, multi-factor prompts and permissions. The shares rose 19% on strong subscription growth and expanded enterprise contracts.

The reason it belongs with CrowdStrike rather than with Intuit is that identity is one of the few software categories that gets structurally larger as AI spreads. Every autonomous agent an organisation runs is a new identity that needs credentials, permissions and an audit trail — and unlike a human employee, it can act thousands of times a day. Somebody has to decide what it is allowed to touch.

Singh cites it in the opening minutes as already-delivered evidence going into the week's other cybersecurity results: "CrowdStrike and other cybersecurity names, Okta, etcetera, did really well last week."

IGV is a fund holding a basket of American software companies. It matters here as a measuring instrument rather than as a stock pick, because it is one half of the most crowded trade of the year: buy semiconductors, sell software, on the theory that AI enriches the chipmakers and destroys the software firms.

For most of the year that worked. The semiconductor index was up 62% while IGV was up only 4%, against 13% for the market as a whole. Then, from 22 June, the trade unwound — largely because the funds running it in size, most famously Situational Awareness, were forced to close their positions. Since that date IGV has risen 25% while the semiconductor index has fallen 22%.

That is a 47-point swing in a little over two months, and the factor data confirms it was positioning rather than fundamentals: Morgan Stanley's index measuring the gap between what momentum funds were long and what they were short fell 36%, and the concentrated technology version fell 54%.

The point Singh takes from it is not that software is cheap but that a consensus view held with borrowed money produces a violent reversal when it is unwound — and the reversal itself creates the entry point in the names on the wrong side of it.

GitLab is where software teams store their code, review each other's changes, and keep a record of every version. It reports this week, and Singh thinks it could beat expectations — while being explicit that he does not own it.

His argument is an economic one with a name. The Jevons paradox describes what happened when steam engines became more efficient: coal use went up, not down, because cheaper power meant far more machines. Singh applies it directly: "the cheaper coding becomes, the more you use it, and GitLab is a way to record it."

The fear about AI and developer tools is that if machines write the code, you need fewer programmers and therefore fewer licences. The counter-argument is that AI does not reduce the amount of code — it multiplies it. Every AI-generated change still has to be stored, reviewed, tested and traced back to whoever approved it, and the more code is produced the more essential that record becomes. "There's so much more code that's coming out in the market… people are just using GitLab more than ever."

GLD is a fund that simply holds gold bars, so owning it is owning gold without a vault. This week produced the third-largest weekly inflow into gold funds in history — $6.4 billion — taking total assets in gold funds worldwide to $615 billion, a seventh consecutive week of buying.

Alongside that, speculators in the futures market did something more dramatic: a record $22.2 billion increase in net bullish positions over three weeks, the largest build in more than a decade, split between $13.6 billion of genuinely new buying and $8.6 billion of people who had bet against gold giving up. Singh's one-line summary of what this means politically: "And Bessent wasn't able to prevent that." The Treasury Secretary has spent weeks trying to talk down long-term interest rates; the gold market is voting on whether he can.

He also flags the risk in his own trade, which is unusual and worth noting. Much of the buying came from computer-driven momentum funds, and traders paid unusually high prices for call options — contracts giving the right to buy gold higher — pushing that activity to the 93rd percentile of the past two years. Crowded positioning is how sharp reversals start, and the dollar bounced on Friday precisely because the expected dovish central banks did not materialise.

The half-joking aside that carries the underlying thesis: watching hyperscaler default-insurance costs hit record highs, "I wouldn't be surprised if a lot of these people are converting assets from vested stock into precious metals."

Nvidia reported what is, on the ordinary reading, the best quarter any listed company has ever produced: $96.2 billion of revenue, double last year's, roughly $60 billion of profit, and — the number Singh flags as the real one — customer commitments jumping from $119 billion to $279 billion in a single quarter. The chief financial officer then guided to about 70% revenue growth through the 2028 financial year, against Wall Street's expectation of 47%. The shares added $453 billion of market value in one session, the largest one-day gain in American history.

And then handed it back. Three reasons. Warsh's hawkish speech pulled every high-multiple stock down. The White House chose that moment to float tariffs on semiconductors extending beyond chips to servers, laptops and games consoles — "horrible timing." And the third is structural.

That third reason is worth understanding properly. Nvidia has invested nearly $50 billion into the AI laboratories that then buy its chips, and told investors that demand supported by its own balance sheet will be roughly a quarter of next year's business. Money leaves Nvidia, arrives at a customer, and comes back as revenue. That is not fraudulent — the chips are real and so is the demand — but it means the growth forecast depends on the credit markets continuing to fund those customers. Singh says it plainly: part of the 70% guide "is contingent on financing for Anthropic doing an IPO in October and OpenAI securing financing," and if credit tightens, investors doubt "whether that number can actually be hit." He is careful to add: "It's not the base case, but it's still a big uncertainty."

The valuation, for context, is not the problem. Raymond James notes Nvidia trades at under 15 times its expected 2027 earnings, cheaper than the S&P 500 at 18.6 times. What is being repriced is not the earnings but the confidence in who is paying for them.

Micron makes memory chips — the components that store data next to a processor. It has been one of Singh's clearest longs, because AI accelerators need enormous quantities of expensive memory and only three companies in the world make it.

This week the evidence points both ways, which is why the stance is neutral rather than positive. Against: Apple is testing memory from two Chinese manufacturers, CXMT and YMTC, for phones sold in China. A state-backed entrant flooding the commodity end of the market is the oldest way a memory cycle ends, and the shares fell 5% on the report. The report's own list of positions to reduce names Micron explicitly, on Chinese capacity and possible hardware tariffs.

For: the same research pack contains data that says the opposite. Server memory prices are up 270% on a year ago; enterprise storage up 235%; the high-bandwidth memory used in AI accelerators is forecast up 70-140% by 2027. And Nvidia's $160 billion increase in purchase commitments was described as primarily memory procurement for the next few years. Trump publicly praised Micron's new $10 billion American research investment.

The distinction that reconciles them is which memory. The contracted, specified, high-bandwidth end is sold out and repricing upward. The commodity end — ordinary phone and consumer memory — is what a Chinese entrant can flood. Micron sits in both, which is why one headline about Apple's supplier list can take 5% off a company whose contracted order book is growing.

The semiconductor index is the cleanest gauge of the AI infrastructure trade, and this week it did something genuinely strange: Nvidia guided to 70% growth against an expectation of 47%, and the index still fell 2.3% on the week.

Goldman's derivatives desk gives three reasons, and they are worth separating. First, credit investors have started asking questions equity investors have ignored — the cost of insuring Nvidia's and Broadcom's debt against default has widened sharply as attention turns to who is guaranteeing the AI build-out. Second, the conversation has begun to include inventory: are chips being bought and stockpiled rather than installed? Third, power shortages and political pushback against new data centres make it harder to know how fast, and how far, this actually gets built.

None of those is fatal, and the desk says so: "none of these kill the AI story, but the range of outcomes is clearly getting wider." A wider range of outcomes is exactly what compresses a multiple, and it has: the index's forward price-to-earnings ratio has fallen from 21-22 times to about 15.

That is the whole state of the trade in one number. Earnings are being revised up; the multiple paid for them is falling faster.

Marvell designs custom chips for specific customers — the alternative to buying Nvidia's off-the-shelf product. It raised its long-term outlook substantially, now expecting revenue of about $18 billion in the 2028 financial year, up 50% year on year, with custom chip revenue more than doubling.

The stock did not move. "The stock still didn't rally given this was priced in."

That non-reaction is the useful signal, and it belongs with the semiconductor de-rating above. In a sector where the multiple has fallen from 21-22 times earnings to about 15, a company can raise its forecasts materially and get nothing for it, because the market has already assumed the good news and is now discounting the risk that it does not persist. When upward revisions stop moving prices, the market has changed what it is worried about.

Amazon appears three times without being a recommendation. Its cloud division committed to installing two million more Nvidia chips across its data centres in 2027 and 2028 — one of the largest single orders behind Nvidia's forecast.

Second, and more awkwardly, Amazon is one of the two companies distorting the entire American earnings season. It owns a stake in Anthropic, whose valuation has risen enormously, and accounting rules require that gain to be booked as profit even though no shares have been sold and no cash received. The effect is startling: excluding Amazon alone would cut expected earnings growth for the whole consumer discretionary sector from a record 92% to 7%. Excluding Amazon and Alphabet together drops the S&P's earnings-beat rate from 29% to 10%.

Third, it is automating: Project Tetromino aims at fully automated delivery stations — the last warehouse stop before a package reaches a driver — targeting roughly two-and-a-half times the throughput of today's sites, with a 2028 pilot. Amazon disputes the reported figures.

None of this is a view on the shares. It is Singh assembling the evidence for two arguments: that headline index earnings are flattered by paper gains that could reverse, and that the AI build-out is real enough to be measured in physical orders.

Alphabet is the other half of the earnings-optics problem. Like Amazon it holds a large stake in Anthropic, and the increase in that stake's paper value flows straight through to reported profit. Headline earnings growth for the S&P 500 reads 118.5%; on an adjusted basis that strips out these one-off revaluations it is 31%. Singh's instruction is to use the second number and treat the first "with a grain of salt."

Running alongside that is a genuine bull case from Wolfe Research, which named Alphabet its top pick for 2027 with a $460 target. The argument is about the chips Google designs itself. Wolfe models more than $105 billion of revenue next year from selling access to those processors at roughly 30% margins — about $32 billion of operating profit — and $300 billion cumulatively through 2028, with Google's electrical capacity scaling to 21 gigawatts next year. On those numbers the cloud division would eventually generate up to 40% of the company's operating profit.

So Alphabet is simultaneously the reason to distrust reported index earnings and one of the few companies with a credible path to justifying its AI spending with actual revenue. Singh does not resolve the tension; he notes both.

The single most important fact about Meta this week is not about Meta. It is reportedly spending up to $10 billion a year buying Anthropic's models — which makes it one of the largest external sources of the revenue that Anthropic is about to float on. When one private AI company's $2 trillion valuation rests partly on payments from a public one, the two are linked in a way that neither balance sheet shows.

Meta is doing this while building competing models internally, and its own frontier model has reportedly slipped to at least October. That is the honest read on where the technology stands: even a company spending tens of billions on its own laboratories finds it cheaper to rent someone else's best model.

Separately it agreed an $18 billion settlement with US states over child safety — $17 billion committed to youth online-safety programmes over ten years plus $2.2 billion to California, with under-18 usage capped at two hours a day, blocked overnight, and an independent auditor for five years. About $10 billion of legal expense hits the third quarter, though the company says its financial guidance is unchanged.

And it heads the list of seven companies carrying $3.1 trillion of commitments that do not appear on their balance sheets — the number behind this week's record-wide credit-insurance costs.

Microsoft appears only as one of seven names in a single number, and that number is the point of the section. Between Meta, Google, Microsoft, Amazon, Oracle, Nvidia and Broadcom there are $3.1 trillion of "off-balance-sheet arrangements" — long-term obligations, leases and commitments that are disclosed in footnotes rather than counted as debt on the face of the accounts.

They are not all guarantees, and Singh says so. But the credit market has started pricing them as if some are: the cost of insuring these companies' debt against default hit all-time highs this week. His view on whether that recedes is blunt — spreads "could calm down if they stop issuing more and more debt, but I don't think that that is the goal."

This is scaffolding for an argument rather than a view on the shares. The argument is that the AI build-out has grown large enough that the way it is financed, not the demand for it, is now the thing that decides whether the forecasts get hit.

Oracle sits inside the same $3.1 trillion of off-balance-sheet commitments, and inside the same logic. Building AI data centres now costs more than these companies generate in spare cash, so the gap is funded with borrowing and with contractual obligations that sit outside the headline debt figures.

Singh's general formulation is the one to keep: elevated AI capital spending "continues to erode free cash flow for the hyperscalers, which could lead to continued debt and equity issuance for the group in the future." More issuance means wider credit spreads, and wider credit spreads eventually mean a higher cost of building — which is how a demand story becomes a financing story.

Apple is not the subject here; it is the cause. It is testing memory chips from two Chinese manufacturers — CXMT for the working memory and YMTC for the storage — for devices sold in China, pending a decision from Washington on whether that is permitted, possibly after a Trump-Xi meeting in September.

That single sourcing decision knocked 6% off SanDisk and 5% off Micron in a week. It is a good illustration of how much of a supplier's value rests on the assumption that it faces no credible new competitor: Apple has not switched, has not committed to anything, and is merely evaluating — and the market repriced the incumbents anyway.

The reason it is more than a scare is that Chinese memory capacity is state-supported and does not need to earn a normal return. The counter-argument, which Singh's own deck supplies, is that the Chinese entrants are in commodity memory rather than the high-bandwidth kind that AI accelerators need, and at a materially higher cost per bit.

Singh's entire verdict on the sell side's enthusiasm here is four words: JPMorgan "reiterated its overweight rating on SpaceX with a 240 target. I don't know how."

The bank's case is that the company is becoming an AI business as well as a rocket business — it completed the acquisition of Cursor, the AI coding tool, which brings roughly $4 billion of annual recurring revenue, about three-quarters of it from businesses, and whose data is already improving the Grok models.

The genuinely striking item is the compute plan: SpaceX intends to put Nvidia's newest rack-scale systems into orbit aboard a first-generation satellite, adapting them for the power, cooling, bandwidth and reliability constraints of space. Singh notes this is "much earlier than anticipated" — the idea being that orbit offers free cooling and uninterrupted solar power, which are precisely the two constraints throttling data centres on the ground.

SpaceX also functions as the yardstick everyone else markets against: Anthropic's pitch to IPO investors is a $30 trillion addressable market topping SpaceX's $28 trillion, and it is trying to raise more money than SpaceX did. When two private companies are competing on the size of the claim rather than the size of the revenue, that is itself information.

IREN was a Bitcoin miner. It owned buildings full of specialised computers and, crucially, contracts for large amounts of electricity — which turns out to be the scarcest input in AI. It has been converting those sites to rent computing capacity to AI companies instead, and this week reported $1 billion of annualised revenue already running and $2.8 billion of new borrowing secured against the graphics chips themselves.

"The stock still didn't really outperform," and the accounts show why the market hesitated. The quarter's profit came in at less than half what was expected, and the full year carried a $702 million loss including $638 million of write-offs for scrapping the old mining hardware. Conversion is expensive and the cost lands now.

The other side is contractual and lands later: AI cloud revenue up roughly eightfold, $4 billion of contracted revenue for 2026 capacity that is already largely sold out, capacity rising from 0.3 gigawatts next year to 0.8 in 2027, $14 billion of cash and committed finance, and recent three-year contracts priced 125% higher than before.

This is the general shape of every miner-turned-AI-landlord: a real, contracted, multi-year revenue stream sitting behind an income statement that currently shows the demolition costs.

SoftBank is raising about $6.3 billion by selling seven-year bonds to Japanese retail savers at a coupon of 4.3-4.9% — the largest retail bond offering ever by a Japanese company. Singh's read on the purpose is direct: "basically to put more money, I would think, into OpenAI," where its cumulative commitments are expected to pass $60 billion.

Two details make this more than a funding story. It has sold its entire Nvidia stake, and it has increased borrowing secured against its holding in Arm. So the money going into the AI laboratories is coming from selling one AI asset, mortgaging another, and borrowing from Japanese households.

That is the same crowding-out mechanism Singh has been tracing all year, seen from the supply side. The AI build-out now needs more capital than its participants generate, so it is reaching further out — to retail bond buyers, to insurers, to anyone who will lend long. Each step makes the whole structure more sensitive to the price of credit rather than to the demand for computing.

Strategy holds roughly 4% of all the Bitcoin that exists, financed largely by issuing shares and preferred stock. The news this week is defensive rather than acquisitive: it has built up about $5.1 billion of dollar reserves and added a further $1.59 billion of cash, and now runs at close to zero net leverage.

The reason this matters to Singh is not the common stock but the preferred shares he owns. Preferred stock pays a fixed dividend ahead of the ordinary shares, and the whole question with these is whether the company can keep paying that dividend through a crypto downturn without being forced to sell Bitcoin. A large designated cash pile — explicitly earmarked, per the disclosure, for paying preferred dividends and interest, repurchasing stock and repaying convertible bonds — is exactly the coverage those holders need to see.

In other words, the company has spent this quarter converting a leveraged bet into a funded one. That is bad for anyone who owned it for maximum Bitcoin exposure and good for anyone who lent it money.

Paramount is the buyer in the Warner Bros. Discovery deal, and this week its own conduct became the obstacle. California's attorney general cancelled the scheduled settlement meeting, accusing the company's leadership of leaking the substance of the talks to the press and then misrepresenting them — "demonstrating a lack of good faith." No new date has been set.

On the substance, Paramount is moving. It ran the numbers on selling Warner's cable channels such as HGTV and the Food Network, concluded it would not have to, and is instead considering offering to negotiate separate distribution agreements with cable operators for different parts of its television business — a structural remedy that addresses the concern without the sale.

The states want more: divesting some cable channels outright, and a binding commitment to keep the film studio separate from Warner's. The gap between what Paramount is offering and what the states are asking is now the deal's remaining risk — and it just became harder to negotiate across.

Two Harbors' merger with CrossCountry Mortgage received its final regulatory approval and closed on 25 August, converting each share into $12.00 in cash. The common shares delist, the preferred shares are redeemed at $25, and the senior notes are bought back at a small premium.

The detail Singh singles out is a small mechanical one worth remembering: holders of record also received a stub-period dividend of about 20 cents, and that dividend "will be paid with the merger consideration and will not reduce or otherwise affect the merger consideration." In many deals a dividend paid before closing is deducted from the price. Here it is additive — free carry for anyone holding the spread through the final weeks. Checking which of the two applies is part of pricing any arbitrage that pays income while you wait.

Caesars is an existing arbitrage position and it now has dates. The definitive proxy was filed, the shareholder vote is set for 22 September, and the antitrust waiting period expires on 14 September.

Those are the last two procedural gates. The interesting part of this deal happened earlier: the proxy revealed that Carl Icahn had bid $34 a share in cash in July and that the bid collapsed on financing, debt covenants and the founding family's refusal to roll their equity into the new deal. Singh had hoped for a higher price and concluded it would not come. What remains is simply waiting for a vote he expects to pass.

Kenvue — the consumer-health business spun out of Johnson & Johnson — is being acquired by Kimberly-Clark, and two regulatory legs moved this week in opposite directions.

New Zealand cleared the deal, but only on condition that Kenvue's feminine-hygiene business is sold. That is the more informative event: a conditional clearance in a small jurisdiction usually previews what larger regulators will ask for, because the competitive overlap that troubled New Zealand exists everywhere the two companies both sell.

Meanwhile the European Commission formally opened its review, with a first-phase deadline of 29 September. That date is now the deal's next binary: clearance at phase one means it proceeds, while a referral to a second-phase investigation would add many months and widen the spread.

Victory Capital has agreed to buy First Eagle, another asset manager, for about $7 billion — $4.4 billion in cash and $2 billion in newly issued Victory shares, plus assuming $575 million of First Eagle's existing debt.

The structure is the part worth reading. First Eagle's private-equity owner, Genstar, ends up with about 14.6% of the combined company but only 4.9% of the votes, with the rest held in non-voting preferred stock that converts later, and the whole holding locked up for three years. That arrangement exists to keep a seller from being treated as a controlling shareholder and to stop it dumping stock into the market on day one.

It is also a good illustration of the private-equity exit problem Singh keeps returning to. With flotations scarce, a sponsor takes payment partly in the buyer's paper, gives up its voting rights, and agrees not to sell for three years — accepting illiquidity as the price of getting out at all.

Boston Scientific makes medical devices — stents, pacemakers, catheters. It disclosed a cybersecurity incident that has caused global disruption to its operations, and the shares fell 3%.

What makes this more serious than a typical breach headline is what the company said the disruption affects: "the ability to process and ship customer orders." That is not data theft, it is a halt in the mechanism by which the company turns products into revenue. And two admissions in the filing define the risk — "the timeline for a full restoration is not yet known," and the company "has not yet determined whether the incident is reasonably likely to have a material impact."

An interruption to revenue recognition with no stated end date and no quantified impact is an unusually open-ended situation for a large-cap. Singh logs it as an event to watch rather than a position — the sort of dislocation that becomes interesting only once the company can say when it ends.

Atlassian makes the tools software teams use to organise work — Jira for tracking tasks, Confluence for documentation. It was squarely in the group of companies the market had written off as AI casualties: if machines write the code, why pay per human seat?

On 6 August it reported solid earnings and outlook and the shares rose 35% the next day.

Singh cites it as the first of three exhibits — with Workday and Salesforce — that forced Dan Niles to question his own framework. Niles has historically liked only three kinds of software: systems of record, security, and gaming. He now "wonders whether the fundamental implications of Atlassian, Workday, and Salesforce are supportive of the technical reactions in the software stocks as a group." The honest version of that thought is that three separate companies delivering results the bear case said were impossible is evidence, not noise.

Workday sells the software companies use to manage employees — payroll, hiring, records. Another obvious AI victim on paper, since the value is supposedly tied to headcount.

It rose 18% on 13 August, not on results but on reports that Silver Lake, a large technology private-equity firm, might buy it. Singh's reasoning about why that mattered is the transferable part: "private equity has higher bars to clear given their use of leverage and holding period than the typical investor." A buyout firm borrowing money to own a business for five to seven years has to believe the cash flows will still be there at the end of that period and will service the debt throughout. A fund manager adding on a good quarter has made a much weaker statement.

So a credible leveraged bid functions as an independent underwriting of terminal value — which is exactly what AI fear had removed from the sector. It "probably put a temporary floor underneath software." When Workday then reported on 27 August, results were "good enough," but Singh notes that the acquisition prospect, not the numbers, drove the 6% move.

Anthropic is preparing to float at a target valuation of $2 trillion, raising up to $100 billion — more than SpaceX raised — and marketing to investors a total addressable market of $30 trillion. Singh's reaction to that last figure is not analytical, it is arithmetical: "which is bananas because US GDP is only $32 trillion."

The revenue underneath is nonetheless extraordinary. Annualised revenue across Anthropic and OpenAI went from $29 billion at the start of the year to $105 billion seven months later, with most of the growth at Anthropic; the company reported $11.6 billion of revenue in a single quarter. Meta alone is reportedly paying up to $10 billion a year, and Salesforce has built the ability to execute its own software's actions from inside Claude.

The reason a private company matters this much to a public-markets call is the dependency running the other way. Part of Nvidia's 70% growth forecast, Singh says explicitly, "is contingent on financing for Anthropic doing an IPO in October." And the paper gains on Amazon's and Alphabet's stakes in Anthropic are what inflate reported American index earnings from 31% growth to 118.5%.

So one unlisted company's flotation now sits underneath the largest listed company's forecast and the index's headline earnings at the same time. That is the concentration risk of this cycle stated in a single sentence.

Three OpenAI items this week, pointing in different directions.

The first is a real technical threat to Nvidia. OpenAI's own chip, designed with Broadcom, beat Nvidia's GB300 on performance per watt and on response speed in internal testing. It is built only for inference — running a finished model — not for training, which remains Nvidia's fortress. The caveat is that the comparison was against GB300 rather than Nvidia's newer generation. But the largest single buyer of AI chips now has a credible alternative for the larger half of the workload.

The second is scale: OpenAI is generating 30 petabytes of data a day, roughly 30 trillion events, and that alone is driving the growth of the database company ClickHouse.

The third is people and money. Its data-centre head left last week, weeks after the chief revenue officer was replaced — turnover at the top of the function responsible for building out capacity. And SoftBank is raising $6.3 billion from Japanese retail bond buyers to put more money in, with cumulative commitments past $60 billion.

Put together: a company whose engineering is strong enough to challenge Nvidia's product, whose costs are large enough to need continual outside financing, and whose ability to secure that financing is one of the stated conditions on Nvidia's own forecast.

Intuit makes TurboTax and QuickBooks. It reported a good quarter — revenue and profits both ahead, a dividend increase, large buybacks — and the shares fell, because it guided next year's earnings to roughly $23 a share against a $27 expectation. A $5 shortfall on a $27 number is not a rounding error; it is a change in what management thinks the business can earn.

The reason matters more than the number. Tax preparation is a self-contained task with a clear right answer, performed once a year by a consumer following prompts. That is precisely the shape of work a language model does well. Singh's phrasing: "the risk of AI disruption in tax seems to be a bigger issue."

Set it beside Salesforce in the same 48 hours and you have the whole framework. Salesforce holds the definitive record of a company's customers, so an AI agent has to go through Salesforce to do anything useful — and it re-rated upward on an alliance that makes exactly that happen. Intuit sells a task, and the task can be done elsewhere.

The report's list of positions to reduce generalises it: de-risk point-solution software providers that lack embedded workflows, platform scale or a clear distribution alliance. Intuit is the named example.

SanDisk makes flash storage — the chips that hold data when the power is off. It fell 6% on the news that Apple is evaluating NAND flash from YMTC, a Chinese manufacturer, for products sold in China, on top of the threat of American chip tariffs. "Memory underperformed AI overall."

The report's list of positions to reduce names it explicitly, alongside Micron, on margin compression from the Chinese capacity ramp.

The distinction worth drawing is which end of the memory market a company sits in. The high-bandwidth memory that goes next to an AI accelerator is specified, contracted years ahead, and currently sold out — Nvidia's $160 billion increase in purchase commitments this quarter was mostly memory procurement. Ordinary storage for phones and laptops is a commodity: buyers compare price per gigabyte, switching costs are low, and a state-supported entrant that does not need a normal return on capital can take share simply by pricing below the incumbents.

SanDisk sits closer to that second market than Micron does, which is why it fell further on the same headline.

IWM tracks the Russell 2000, the index of small American companies. It fell about 1.4% on the week while large-cap value rose, and the mechanism is one Singh restates every time rates move: "the Russell has a lot — half the companies are unprofitable. So if interest rates go up and they have debt, that affects them more than large-cap companies."

Two things happen to a small unprofitable company when yields rise. Its existing debt costs more to refinance, which comes straight out of whatever cash it has. And because its value rests almost entirely on profits expected years from now, a higher discount rate reduces what those distant profits are worth today by more than it does for a company earning money now.

The specific move this week was a "bear flattening" — short-term yields rising faster than long-term ones, which is what happens when the market suddenly believes the central bank will raise rates. That is the most direct possible hit to a leveraged, unprofitable borrower.

The report's instruction is to trim the exposure. It is the cleanest way to express the view that Warsh has made a hike more likely, without needing to be right about when.

DICK'S bought Foot Locker, and this quarter the acquisition is what broke. Group revenue rose 53% purely because Foot Locker is now consolidated, but adjusted earnings per share fell 19% and missed. The split tells the story: DICK'S own stores grew sales 4.9%, while Foot Locker's comparable sales fell 3.6%.

The full-year guidance cut is severe. Earnings guidance went to $11-12 a share against a $14.24 expectation, and Foot Locker's segment profit was re-guided from a positive $110-150 million to a loss of $40-80 million — a swing of roughly $200 million in one quarter's thinking. Operating margin fell almost five percentage points and inventory is up 63%.

Ed Stack's explanation names the cause precisely: the athletic footwear market "became increasingly promotional," and Foot Locker is disproportionately exposed because it depends on limited-edition launches and retro models rather than everyday product. When those launches are fewer and sell below expectations, the inventory has to be discounted, and the margin goes with it.

It belongs beside Walmart's worst comparable-sales miss in years, a cautious Lowe's and a more upbeat Target as this week's mixed read on the American consumer — with the specific warning that a struggling acquisition is where a soft consumer shows up first.