Subscribe now

Bulls finally got a reprieve Wednesday, snapping a three-day losing streak. But the rebound came with a catch. Stocks still couldn’t push through a key resistance level, while surging energy prices and bond yields continued to weigh on the market.

The S&P 500 closed at 7,667, a gain of 0.5%. The session provided a much-needed snapback, but that was all. Professional money managers are still more likely than not to raise cash levels as they try to make sense of a war with Iran that shows no sign of ending and the highest 10-year Treasury note yields in three years.

The 10-year Treasury yield hit a high on Wednesday of 4.82%.

To make matters worse, West Texas Intermediate and Brent crude oil closed up about 1%, at $91.01 and $95.63, respectively. Traders in the energy pits bid up prices after the U.S. military said it launched more attacks on Iranian military targets. The gains reversed earlier weakness when the U.S. Secretary of Energy claimed on CNBC that 17 million barrels of oil transitioned through the Strait of Hormuz on Monday.

There is a lot of debate about the federal deficit and interest payments. This is a familiar bearish talking point, made worse by a costly war in the Middle East with no real political will on either side to tackle spending.

Politicians on the right want to cut taxes while increasing spending for the military and farm subsidies. Left-leaning politicians want to raise taxes while increasing spending for the military and social safety-net programs.

Bond traders want politicians to cut spending across the board.

Read more

Anthropic has raised more than $125 billion privately, pays its operating costs out of its own revenue, and sits on more cash than most companies already trading. It is going public in about three weeks anyway.

The Information reported that the company plans to publish its prospectus after the Labor Day holiday, hold an investor day in mid-September, and list in late September or early October.

When we covered the $2 trillion expectation three weeks ago, every version of the story pointed to October. The window has moved forward, and a prospectus has to be public at least 15 days before the roadshow, which turns Labor Day into a deadline.

What Changed In Three Weeks

The gap since our last look was unusually busy.

Bloomberg reported that annualized revenue passed $65 billion, up from the $47 billion pace that anchored most of the skeptical coverage. Second-quarter revenue came in above $11.5 billion against $787 million a year earlier, and the disclosure showed positive adjusted operating income, the first any frontier lab has shown investors.

Those figures are preliminary and unaudited, and nobody outside the company can yet see which costs the word adjusted removes.

Then came the capacity. Anthropic committed $45 billion to Nscale on August 26 and $35 billion to Lambda on August 31. That is roughly $80 billion of compute inside a single week, with the Lambda capacity landing at a Hut 8 campus in Nueces County, Texas.

Signing that much capacity in the fortnight before a filing says something about demand rather than something about spending. Anthropic sells API access at a reported gross margin above 80 percent into a queue it has repeatedly been unable to serve. For a business in that position, contracted megawatts arriving on dated schedules are the revenue forecast wearing different clothes.

Two provisions in the reporting describe how the shares themselves will move.

Anthropic is considering letting certain existing shareholders sell directly into the offering. It is separately exploring extending the lock-up for some holders beyond the usual 180 days.

Those choices look contradictory and are not. Together they produce one outcome: a specific set of early holders turns paper into cash at the offer price, while the shares that would otherwise reach the market next spring stay where they are.

A company raising money to build something maximizes the new shares it sells. A company manufacturing liquidity decides who is allowed to sell, and when.

The Money Was Always The Optional Part

We argued in July that the cash was the part of this IPO Anthropic could take or leave. The conclusion then was that a listing buys other things: liquidity for employees and early backers holding enormous paper gains, and a share currency for acquisitions and hiring. It also buys a public mark on what the owners already hold, plus audited numbers for the enterprise customers deciding what to build on.

The offering now being described is built for that list, item by item. Secondary sales for insiders, deliberate management of the float, and a calendar the company chose rather than one a cash runway imposed. Reading this listing as a fundraising event misses what the structure is doing.

What Got Cleared, And What Did Not

Two legal threads travel with the company into the window.

A federal judge ruled on August 28 that the Pentagon’s designation of Anthropic as a supply chain risk was illegal and baseless, in a 59-page opinion finding the label was retaliation. A parallel case in Washington keeps the label technically alive. Anthropic was also absent on August 31 when the Defense Department added ChatGPT and Grok to its internal AI portal, a platform already reaching 1.7 million users. The largest single AI deployment in the country currently runs without Claude on it.

The same day the judge ruled, Sony Music Publishing and Warner Chappell filed a fifth music copyright suit, naming Dario Amodei and Benjamin Mann personally and seeking up to $150,000 per work. Anthropic settled the Bartz case for $1.5 billion last September, which set a rough clearing price for this category of claim.

Neither thread threatens the business at this scale. Both become disclosed risk factors the moment the document publishes, written in the company’s own words and signed.

What The Document Settles

The prospectus converts reported numbers into audited ones. The 80 percent gross margin becomes a figure an accounting firm put its name to. The adjusted operating profit becomes GAAP, with the adjustments itemized. The compute commitments become a schedule with years attached, printed near the revenue curve they are meant to produce. Those two disclosures sitting together show whether the buildout and the forecast are the same document under two names.

Everything else in the trade gets marked against it. Data center developers, power contractors and accelerator makers have been valued on the premise that the labs will pay their bills for years, and that premise has rested on press releases and reported run rates. In about two weeks an auditor’s signature replaces both.

Subscribe now

Subscribe now

Bulls started September in a deep funk after reports that the U.S. launched several new attacks against Iran.

The S&P 500 opened below its first key support level, and it was all downhill from there. The benchmark index closed at 7,624, a decline of 0.8%. The weakness was broadly based and deliberate.

The same nervous money managers who have been chasing stocks higher all of 2025 for fear of falling behind finally decided that enough was enough. They liquidated risk assets across the board. The result was an ugly decline that saw the benchmark settle near the worst levels of the session.

The problem for bulls is rising bond yields.

West Texas Intermediate crude oil rose 5.4% to $90.39 per barrel. North Sea Brent, the international equivalent, rallied to $95.12 per barrel, up 5.1%. Both prices reached their highest levels since June, and that is driving bond yields sharply higher. On Tuesday, the yield for the 10-year Treasury note crested at 4.8%, a level not seen since January 2025. Bond yields, not the Federal Reserve, set the standard for commercial and consumer loans. Higher borrowing costs kill economic growth. Higher yields also wreak havoc with the economic models professional money managers use for risk allocation. Those models are now advising managers to sell equities.

Only four S&P 500 sectors finished in the green. They were all defensive in nature.

Read more

Subscribe now

Bulls started and ended Monday in the red after the United States launched a surprise attack on Larak Island, a tiny island in the narrowest part of the Strait of Hormuz. The escalation in the war with Iran sent energy prices and bond yields higher. The S&P 500 closed at 7,686, a decline of 0.3%. Although the loss for the benchmark was the second in as many days, for the month of August the index still climbed by 2.6%. That gain breaks a two-month losing streak and brings year-to-date gains to 12.5%. The strength of the benchmark is a real pain point for professionals.

In fairness, professionals do not want to have exposure to equities. They view the war with Iran and the rising energy prices that the conflict brings as a real headwind for stocks. The attack on Larak Island is especially salient because that landmass is a key logistics base for anti-ship missile batteries, fast-attack boats, radar systems, and infrastructure for deploying sea mines.

Hitting Larak Island represents a dramatic escalation in a war. It should bring higher energy prices, greater structural inflation, and increased bond yields. All of those metrics moved higher on Monday, with West Texas Intermediate crude oil surging 3.5% to $86.30 per barrel. Still, stock prices moved only moderately lower.

Read more

Google’s second full year of selling search ads produced about $440 million in revenue, and in 2002 that counted as blistering growth.

OpenAI beat it in about seven months.

The company said Monday that ChatGPT’s advertising business has crossed $1 billion in annualized revenue, roughly 200 days after the first ads appeared in a February pilot. A business that did not exist in January is now on a billion-dollar pace.

Run rates flatter, so use the harder number. OpenAI is guiding to $2.5 billion of ad revenue booked this calendar year, a pace Google’s ad machine needed close to four years to reach. And on the same Monday, the European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act.

No AI chatbot had ever been placed in that legal category, which was built for Google Search and Bing. The ad market and Europe’s top regulator reached the same conclusion within hours of each other: ChatGPT now runs on search economics.

The Fastest Ramp On Record

The business OpenAI described is already global. Ads run in more than 40 countries on the free tier and the entry-level Go plan, and the self-serve platform opened Monday to marketers in India, Europe, the Middle East, and North Africa. OpenAI says the ads carry labels, sit apart from how answers are generated, and give advertisers no visibility into private conversations. Reuters reports the company’s overall revenue run rate has passed $40 billion, roughly double where it ended 2025.

The timing of the disclosure is easy to read. OpenAI is working toward a public listing, and Monday’s announcement was positioned as proof of diversification ahead of it. A revenue line that goes from zero to a billion inside a year is the best diversification evidence a consumer platform can show a public investor. Subscriptions prove that power users will pay. Advertising puts a value on everyone else.

The designation was mechanical, which is what makes it valuable.

Falling under the DSA’s strictest tier requires more than 45 million monthly users in the EU. OpenAI reported an average of 159.1 million monthly European users for the six months through March, more than triple the threshold. The Commission classified ChatGPT as a hybrid service that qualifies as a search engine because it answers queries by searching the web. Compliance lands by the end of December, with fines of up to 6 percent of global revenue for failure.

A regulator with the power to compel disclosure just produced the cleanest usage number the AI industry has ever printed: 159 million monthly European users, reported by OpenAI under legal obligation.

Every other audience figure in AI comes from a marketing department. This one carries legal consequences for being wrong, and it establishes scale in the one market where scale now brings Google-grade regulatory costs. Those costs cut both ways. The compliance wall the DSA erects can only be climbed by companies already at scale, which quietly entrenches the firms that have it.

The Google 2000 Rhyme

Advertising is the layer that turned Google from an admired, cash-burning search box into one of the most profitable business models ever built. The same layer is now assembling under ChatGPT in a fraction of the time.

The reflex trade this suggests, selling Alphabet, has a poor record. The 2023 obituaries for the ten blue links preceded one of the great runs in Alphabet’s history, and nothing in Monday’s numbers shows ad budgets leaving Google. What the numbers do show is a product outside the incumbent search oligopoly monetizing query intent at scale, and doing it years faster than any predecessor. The near-term contest is for ad budgets that were never in search to begin with, the brand and discovery spending that today sits with social feeds and retail media. Where it goes after that depends on numbers neither company has printed yet.

Three Numbers To Track

The first is whether OpenAI actually books the $2.5 billion it just guided to, in its first year running a young ad system at global scale. The second is whether Google’s paid-click growth decelerates in the autumn report, the first print with a full quarter of ChatGPT ads running worldwide. The third is whether ChatGPT clears its December DSA deadline cleanly, because the designation converts European scale from a bragging right into a standing cost.

A free tier that pays for itself changes the economics of the whole industry. Every ad dollar makes free AI distribution self-funding, and self-funding distribution buys more compute, more power, and more data center capacity underneath it. The application layer got a second engine on Monday. The layers below it collect either way.

Subscribe now

Subscribe now

Bulls started Friday determined to continue the rally. The S&P 500 jumped to a weekly high of 7,771, then Federal Reserve Chairman Kevin Warsh began his Jackson Hole speech. He did not say much, but the rally reversed and the benchmark ended the session near the lows.

The S&P 500 closed at 7,712, a loss of 0.3%. For the week, bulls still added 0.5%, the third advance in four weeks. The benchmark is also on pace to close August with a solid gain of about 3%. The only question now is whether Friday was a pivot point that pushed professional buyers to the sidelines. Bulls should be worried about interest rates. Bears should fear momentum.

Treasury bond yields have been creeping higher for a month. The yield for the 10-year Treasury note climbed on Friday to 4.72%. Fixed-income traders are trying to send President Trump and Treasury Secretary Bessent a message that it is time to change the trajectory on spending.

The federal debt reached $40 trillion in August, the highest level on record. That alone is not as consequential as it might seem. The federal debt has been too high for 40 years, and investors continue to buy Treasury bonds, notes, and bills. They also have been voracious buyers of stocks.

The worry this time is that the President does not seem to care about the message being sent in the bond pits. He could change the narrative by making an effort to end the war with Iran that is stoking inflation fears. Over the weekend, President Trump claimed that he secured 65 billion barrels of oil reserves from Venezuela at no cost to U.S. taxpayers. The President, known for hyperbole, says it is the largest deal in world history and will help bring down prices at the pump.

Energy trading had not started at the time of this report, but the President’s intentions are clear. He is trying to put an end to the upward momentum in bond yields. He understands fixed-income traders are worried about inflation, and they see rising energy prices as an obstacle to reversing that trend. Stock market bears should be worried about President Trump. Policies aside, he is keenly aware of what causes them pain.

Five sectors moved higher on Friday, with consumer discretionary, communication services, energy, consumer staples, and financials posting gains of 1.7%, 1.6%, 0.6%, 0.6%, and 0.3%, respectively. Amazon was a large winner as the stock was upgraded after news the company will buy 2 million Nvidia GPUs as it seeks to build out its AI infrastructure network. Shares added 4.0%.

Read more

Subscribe now

Bulls and bears spent the first half of this week awaiting second-quarter financial results from Nvidia. On Wednesday evening, that report landed, and technology buyers were unleashed. The S&P 500 closed at 7,731, up 0.7%. Barring an unforeseen event, bulls should be able to press to new highs as professional money managers shift assets back toward risk. The reason for this shift is simple: they cannot afford to be left behind.

Their benchmark is racing higher, and sitting on the sidelines is not an option after so many professionals fell behind in 2025.

Bears want everyone to believe that the macroeconomic environment is fraught with peril, and this is undoubtedly true. Interest rates are still too high. Energy prices remain above $80 per barrel, a level that causes corporate margin compression and weaker consumer spending.

Despite all of this, bears are still losing.

More importantly, on Thursday, their hopes and dreams of an artificial intelligence bubble burst with the Nvidia 2027 forecast that revenues will soar by 70%, versus the Wall Street estimate of 44%. So, pros are holding their noses, and they are adding back risk assets. Bears are the casualty of this shift.

The demise of bears was most evident in the performance of enterprise software stocks. Salesforce reported a lukewarm second quarter alongside Nvidia on Wednesday evening. Top-line sales jumped a modest 11% year-over-year.

The sweet irony is that almost all of its bottom-line profits came via its investment in Anthropic, the AI frontier lab that was supposed to kill enterprise software. Salesforce shares jumped by 22.5% as bears were forced to cover SaaSpocalypse short positions.

Read more

Someone has written the end of Nvidia every year since 2015.

Cheaper chips would commoditise the GPU. Custom silicon inside the hyperscalers would take the workload back, an efficiency breakthrough out of China would collapse the demand for compute, and export bans would close the largest market on earth.

In every version of the story, the capital-spending cycle was always about to roll over. Markman Capital has owned the stock through all of it.

The Position Behind This Report

Nvidia has been a Markman Capital position since 2015, when the shares traded at 50 cents on a split-adjusted basis. The oldest lot still held in the Digital Creators and Consumers portfolio was opened on March 29, 2023 at a cost basis of $26.68. At the settled close of $209.66 on August 26, the session before this report landed, that position is up 686 percent. It is rated Hold. The core has never been sold.

We lead with that number because it is the part of this business that is hardest to reproduce. Every argument for selling has been available continuously for a decade, delivered with citations, and every one of them has been wrong. Holding through all of it is the position, and it is the harder half of the work.

A record quarter on its own would not be worth writing about. What makes this one different is that two of its disclosures are not about the quarter at all.

What The Quarter Actually Showed

Nvidia sold $96.2 billion of product in three months and kept 75 cents of every dollar as gross profit. Revenue more than doubled from a year ago, data-center sales came to $89.0 billion, and operating income reached $63.7 billion. Diluted earnings per share rose 128 percent.

Jensen Huang put the quarter in one line: “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.”

The rest of the report:

• Gross margin of 75.0 percent, with $26 billion returned to shareholders and $99 billion left on the buyback

• Third-quarter revenue guided to $108.0 billion, roughly $4 billion above consensus, assuming no data-center compute revenue from China at all

• Supply commitments up from $119 billion to $279 billion, primarily to lock down memory

• Inventory at $31.6 billion, up from $25.8 billion, staged for the Vera Rubin launch

Then, on the call, finance chief Colette Kress told analysts to expect fiscal 2028 revenue growth of roughly 70 percent against the 44 percent the street was carrying. Nvidia has never issued a year-ahead forecast in its history. Shares rose more than 4 percent after hours once it landed.

A company cannot credibly forecast a year out while five or six capital-spending committees control its revenue line. Nvidia just did, and the reason is sitting in a segment the company only started publishing two quarters ago.

The Bucket Nvidia Started Naming

Nvidia now splits data-center revenue in two. One line is Hyperscale, the handful of cloud giants everyone already models. The other the company calls ACIE, for AI Clouds, Industrial and Enterprise: the regional clouds, the AI natives, the sovereign programs, the banks and drugmakers and manufacturers standing up their own clusters.

Hyperscale did $48.7 billion and grew 102 percent. ACIE did $40.3 billion and grew 138 percent, which puts the customers outside the hyperscalers at 45 percent of Nvidia’s data-center business.

The longer history is the honest version of this story, and it is more interesting than a straight line. The long tail was roughly half of data-center revenue through fiscal 2025, then got squeezed to 31 percent as the hyperscalers absorbed the first wave of Blackwell. It has since gone from $16.1 billion a quarter to $40.3 billion in three quarters. The share recovered because those dollars grew two and a half times over, not because the hyperscalers slowed down.

Huang has been describing the asymmetry for a year. As he framed it on the May call, “the easiest go-to-market, of course, is the hyperscaler, because there are only five or six of them. The rest of them, the rest of the industry, represents 250,000 companies around the world.” On this quarter’s call he went further about where the weight eventually sits:

“That part of the world’s computing is likely to be larger over time than even what we are currently experiencing in the cloud.”

The bear case against Nvidia has rarely been about the chips. It has been about customer concentration: five or six buyers, one capital-spending cycle, one turn of the wheel, and semiconductors do turn. That premise is what the ACIE line erodes.

The second bucket holds a bank standing up a private cluster for risk modelling, a drugmaker installing on-premises systems for discovery work, and a government building a national AI service. None of those buyers runs on the same budget calendar as Microsoft or Meta, so a pause at one does not travel to the others.

Why A Year-Ahead Number Is Possible Now

The 70 percent figure describes what Nvidia can build rather than what customers will take. Huang told analysts the ceiling is the supply chain. His words: “Even though our demand is much greater than 70 percent, our supply allows us to confidently deliver 70 percent. We are going to continue to work with our supply chain to increase on that.”

The arithmetic underneath it is the part most coverage skipped. Nvidia now earns roughly $40 billion of revenue per gigawatt of AI data-center capacity on the Vera Rubin generation, against about $25 billion on Blackwell and $18 billion on Hopper. The same megawatt of power buys progressively more Nvidia content each cycle. A world that builds the same number of gigawatts next year still pays Nvidia substantially more for them.

That figure cross-checks against the company’s own disclosures. Nvidia has guaranteed the land, power and shell buildout for about 4.25 gigawatts at SB Energy’s PORTS-Pike campus in Ohio. The site will host Nvidia infrastructure exclusively, under twenty-year leases to OpenAI. Each generation of infrastructure deployed there runs to approximately 1.5 million GPUs, which the filing values at $150 billion to $200 billion of Nvidia revenue from a single campus.

Run the per-gigawatt figure against that site and the two disclosures line up. At $40 billion per gigawatt, 4.25 gigawatts comes to $170 billion, sitting inside the range the filing gives. Nvidia published those numbers in different documents for different reasons, and they agree.

The Plumbing For 250,000 Customers

Selling to a quarter of a million buyers is harder than selling to six, and Nvidia has spent the past year building the machinery for it rather than waiting for the demand to organise itself.

The company’s own language on the problem is unusually direct. AI clouds and model makers are “seeing extraordinary demand for AI infrastructure, yet many are growing faster than their balance sheets and long-term credit profiles can support.” So Nvidia arranges what those customers cannot arrange alone. It partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third-party capital for AI infrastructure.

The rest of the machinery is less visible. Nvidia signed fifteen-year data-center leases it intends to reassign to other parties, and it guarantees lease obligations for selected partners, with maximum exposure across every such agreement disclosed at $108.5 billion.

Kress described what Nvidia gets in return: “In this model, we get paid twice, once on the hardware sale, and again through the share of rental revenue.” That is a landlord’s economics layered on a manufacturer’s, and it exists because the long tail cannot write hyperscaler-sized cheques from its own balance sheet.

This is the same pattern that built toll roads, power plants and fibre networks. Institutional capital funds the asset, contracted revenue services the capital, and the manufacturer sells into demand that could not otherwise be financed. The constraint being attacked is access to capital, and it is the one standing between Nvidia and 250,000 customers.

The hyperscalers, meanwhile, have not slowed down. Amazon agreed to deploy an additional 2 million Nvidia GPUs across its global infrastructure between now and the second quarter of fiscal 2029.

What The Memory Squeeze Actually Does

The honest pressure point in this quarter is memory, and it shows up in the margin line.

Gross margin hit 75.0 percent in the second quarter. Nvidia guides it to 74.0 percent in the third, and Kress told analysts it bottoms at 71 to 72 percent in the fourth before settling at 72 to 73 percent in fiscal 2028. She cited “extreme pricing conditions” in memory and said the increases have run past what the company expected. That is three quarters of compression from the peak, and it is the number a skeptic should press on.

Memory is scarce because the AI buildout is consuming it, which is the same force filling Nvidia's order book. What the company did about it is the part worth sitting with. Supply and capacity commitments went from $119 billion to $279 billion in a single quarter, spread across fiscal 2027, 2028 and 2029, and the filing says the increase went primarily to memory.

That carries a second effect the margin discussion tends to miss. Memory Nvidia has contracted through fiscal 2029 is memory that is not available to anyone else. Every competing accelerator, whether it comes from AMD, from Broadcom’s custom silicon work, or from a hyperscaler’s in-house program, needs high-bandwidth memory from the same small group of suppliers. That pre-commitment protects Nvidia’s own ramp. It also sets the price and the queue position for everyone behind it.

Two balance-sheet lines moved alongside it. Days sales outstanding rose to 60 from 45, which the filing attributes to extended payment terms on large multi-quarter agreements with investment-grade customers. Inventory climbed to $31.6 billion as Nvidia stages Vera Rubin for its October launch. This is what a company looks like when it is financing a step-change in volume. It is also where trouble would surface first if demand ever softened, so both lines deserve a look on the next report.

What The Quarter Settles

The market spent this year arguing about whether hyperscaler capital spending can hold. This report says the question is going out of date. Nearly half of Nvidia’s data-center revenue now comes from customers who are not hyperscalers, that half grew 138 percent, and the company has committed three years of supply to serving it.

Two things it does not settle. Memory costs will take three or four points out of gross margin before they stop, and nobody can name the quarter that reverses. And China sits at zero in the October guide, so that $108 billion carries no credit at all for a market that was worth billions a quarter as recently as last year.

The names that benefit from this dynamic are the ones selling into the second bucket rather than the first. The power and cooling suppliers, the optical and networking companies, the landlords with contracted megawatts. A buildout that broadens from six customers to a few thousand touches far more suppliers than one that does not, and it does so on budget calendars that are not synchronised.

Nvidia used to depend on a handful of customers deciding to spend. It now sells into a market that is spending for reasons of its own, and it has bought the supply to serve it through 2029.

The Part Members See

You saw the Nvidia position at the top of this report. What you did not see is the part that made it a position in the first place. The entry logic. The price target we are working toward. The conditions that would change the rating.

That is Digital Creators and Consumers.

DCC is where this research is done in full. Members get the concentrated portfolio with every entry and every target, plus two deep dives a week on the companies inside it. Alerts go out when something in the book changes, and the archive runs to more than 5,000 letters going back to 2014.

Nvidia is one position among roughly two dozen.

In 2025 the portfolio closed 18 of 19 positions as winners and finished up 55 percent against 18 percent for the S&P 500.

Nine positions have closed in 2026, all of them winners, and 27 of the last 28 closed positions have been profitable.

Past results do not carry forward on their own. What carries forward is the research process behind them, unchanged in 25 years and documented in every letter we have written.

See what DCC members get

Subscribe now

Bulls and bears spent all of Wednesday battling for control as investors waited for earnings reports from Nvidia and Salesforce. The S&P 500 closed at 7,676, essentially unchanged for the session. This was not surprising. At important turning points, investors often wait for major companies to report their financial results before deciding which market narrative is correct.

Now, professional money managers must make a choice: Do they pile back into growth stocks even as the macroeconomic outlook is clouded by expensive oil and rising interest rates, or do they sell the good news and wait for greater clarity?

The numbers from Nvidia and Salesforce were a tale of expectations.

Nvidia is at the heart of the artificial intelligence boom. The company is generating profits at an extraordinary rate, and its outlook remains very strong.

CEO Jensen Huang said second-quarter sales reached $96.2 billion, up 106% from a year earlier. Profits jumped to $59.7 billion, an increase of 126% year-over-year. Nvidia expects third-quarter sales to reach around $108 billion, which would represent roughly 110% growth from the same period last year.

Despite these huge numbers, Nvidia shares fell about 0.5% after hours.

Salesforce numbers told a different story. The company reported second-quarter sales of $11.3 billion, up 11% from last year. Profits reached $3.5 billion. However, $2.6 billion of those profits came from investment gains, mainly from Salesforce’s stake in Anthropic.

Even so, Salesforce shares jumped 14% after the report. The main reason was that investors had very low expectations going into the earnings announcement because of fears about the so-called “SaaSpocalypse.”

In many ways, Nvidia and Salesforce represent different sides of the same coin.

Nvidia is the leading company associated with AI, which also means it attracts a lot of skepticism. Some investors still question whether the AI boom can live up to the enormous expectations surrounding it, despite the strong financial results.

Salesforce, on the other hand, was viewed as a potential victim of AI. Investors worried that AI agents could eventually perform much of the work currently done by enterprise software.

But that does not appear to be happening quickly.

Most Fortune 500 companies are still moving cautiously when it comes to AI. They are certainly not rushing to hand control of their businesses and customer data to AI companies such as OpenAI and Anthropic.

The SaaSpocalypse was probably never as serious as investors feared. Large companies tend to be very cautious, especially when important customer information is involved. They usually take time before making major technology changes.

Read more

Four billion dollars sounds like a lot until you put it next to the market Treasury is trying to influence.

Treasury securities traded an average of $1.2 trillion a day through July, up 12% from a year earlier. A $4 billion buyback is just three-tenths of 1% of a single day’s trading.

That is the size of the operation the Treasury Department announced on August 19, doubling its long-dated bond buybacks from $2 billion to at least $4 billion per operation. The purchases cover the 10-to-20-year and 20-to-30-year sectors and will run from September 9 through November 4.

The announcement came a day after the 30-year Treasury yield hit 5.33%, its highest level since 2007.

Yields fell within minutes, and they were back where they started inside two sessions.

Treasury describes buybacks as liquidity support, buying older and less actively traded bonds so investors can move positions more easily.

That is a real and unremarkable function.

The initial reaction was exactly what Treasury wanted. The 30-year yield fell 9 basis points to 5.196%, while the 10-year fell 6 basis points to 4.647%. By Thursday afternoon, the 30-year had climbed back to 5.251%, just a few basis points from where it was before the announcement. By Friday, the rest of the move was gone.

Then Bessent said the operations could grow beyond the announced $4 billion.

Several things were pushing yields up, not one. A global bond selloff, stalled talks to end the war with Iran, and persistent inflation readings were all in the mix. So was a new Federal Reserve chair whose reaction function nobody has seen yet, and a national debt that just passed $40 trillion. A buyback addresses none of them.

Bessent’s Own Mentor Called it Mistake

Stanley Druckenmiller hired Scott Bessent at Soros Fund Management in 1991, and the pair are still best known for the 1992 bet against the British pound. He hired Kevin Warsh at Duquesne in 2011, and Warsh named him as one of three personal mentors in his Senate confirmation testimony in April. Bessent once told the Financial Times that in macro there is “Stan and then everybody else.”

On Monday, Druckenmiller published an opinion piece in the Wall Street Journal calling the buyback expansion a mistake and describing it as price management rather than liquidity management.

His strongest point is the one getting least attention: there was no dysfunction to repair. Auctions were clearing, volatility was contained, and trading was orderly. The conditions that normally justify official intervention were absent, which leaves the timing to explain itself.

The Asset Being Spent

The credibility of the Treasury market rests on the belief that nobody is managing its price. That belief lets the largest debt market in the world absorb $40 trillion of supply without charging a risk premium for official interference. It is not a line item, and no amount of TGA cash buys it back.

Druckenmiller’s warning is about what happens once the belief goes:

“Once markets believe Treasury is defending a price, every rise in yields becomes a test of official resolve, and the operations must grow to survive the tests.”

The second cost is slower, and it lands in Congress. Long yields rise because investors want more compensation to hold thirty-year government debt, which is a message aimed straight at the people writing the budget. Muffling it buys time, and time is what makes the arithmetic worse. Druckenmiller’s phrase for this is “a subsidy to procrastination.”

What This Means For Investors

The long end of the curve is where the cost of borrowed money gets set for everyone else. A 30-year yield above 5% arrives as a real bill for every leveraged balance sheet and every business model built on cheap refinancing three years from now.

Nothing announced in August changes that arithmetic. The fiscal path is unresolved, the buyback is too small to matter, and the tool being reached for next is the government’s cash account. The reasonable working assumption is that expensive long money lasts.

That assumption sorts companies quickly. A business funding its own growth out of operating cash flow barely notices what the 30-year does. A business that needs the refinancing window open in 2029 notices a great deal, and the difference between the two stays invisible for as long as money is cheap. A 5% long bond is what makes it visible again.

Subscribe now

Subscribe now

Bulls regrouped on Tuesday after several news outlets reported that backchannel talks between Iranian officials and Pakistan’s army chief might reach a diplomatic end to the war in the Middle East. The S&P 500 closed at 7,677, a gain of 0.3%. If it seems like the benchmark index is treading water and going nowhere fast, that is an accurate assessment. Bulls and bears are trading punches ahead of an important earnings report on Wednesday from Nvidia, alongside continued concerns about inflation and Treasury bond yields.

Bears have definitely gained an edge. With the federal deficit now above $40 trillion and President Trump’s seeming unwillingness to move off hawkish rhetoric regarding the war in Iran, bears have argued that the deficit is going dramatically higher. Defense budgets are rising to replenish spent munitions, while Democrats running in the Congressional midterm elections push aggressive deficit spending. There is no pushback from their political foes on the Republican side of the aisle.

An unpopular war and rising costs for food and gasoline have pushed fiscal responsibility off the table. On Tuesday, the Conference Board said that consumer confidence edged lower in August to 89.4. A survey of economists polled by Dow Jones revealed expectations were for a reading of 90.2.

Semiconductors were stronger in the session after Raymond James upgraded Advanced Micro Devices from Outperform to Strong Buy. The investment broker believes shares are headed to $641, versus its previous target of $565. Analysts at Raymond James believe AMD has the strongest combination of earnings leverage and data center share gains, especially in CPU servers, a market segment it sees reaching $201 billion by 2030. This view sparked several rounds of short covering across the semiconductor group. AMD added 4.9%, Marvell gained 4.8%, and Micron Technology was up 2.5%. Nvidia was the laggard, rallying by 2.2%.

Read more

For decades, IBM's mainframe has been one of the most closed environments in enterprise computing. Its software runs on IBM's architecture, its customers build around it, and the systems are designed to stay in service for years.

The company unveiled a dual-architecture processor at the Hot Chips Conference on Monday. Its cores run IBM's own mainframe instruction set and Arm's on the same silicon, switching between them in nanoseconds. Nothing is being emulated or translated.

It is the first dual-architecture mainframe processor anyone has built.

Two Instruction Sets, One Core

A processor core normally understands one language. IBM’s next one understands two.

Built on a 2 nanometer process, the chip carries 11 cores running above 5.7 GHz, with inference accelerators designed to catch fraud inside a transaction while it clears. The fraud check happens on the same chip that moves the money, rather than somewhere downstream after it has already gone.

What it removes is a choice that has frustrated large institutions for a decade.

A bank running its core deposit systems on a mainframe could adopt modern AI tools, or it could leave its transaction engine alone. Doing both meant rewriting code written over thirty years, or buying a second set of hardware to stand beside the first. Now Arm software runs beside the mainframe workload inside the same box.

Mainframes have been declared obsolete by every wave of computing since the personal computer.

Roughly 71% of Fortune 500 firms still run them and close to 97% of banks depend on them, with enterprises spending something like $330 billion a year keeping those systems alive. Most of the world’s large banks still run their daily business on this hardware, and they are budgeting to keep doing it.

That spending usually gets described as a burden. It reads better as a measure of how immovable the work is. This code clears payments, prices policies and settles trades, and the cost of getting it wrong is counted in regulatory findings.

So the opening was always going to come from somewhere other than replacement. Modern software could reach these machines by learning their language, or by teaching them everyone else’s. IBM and Arm chose the second.

The Deal Behind The Chip

Most coverage is running this as an IBM chip story. The detail that reframes it sits in IBM’s own release: the processor is the first milestone out of an IBM and Arm collaboration established in April 2026.

A company quietly adopting a rival’s instruction set to stay current is one story. Two companies spending months building a core that runs both, then presenting it together, is a different one. Arm’s executive vice president for cloud AI, Mohamed Awad, said what the partnership is for:

“Bringing Arm compute and its software ecosystem to these platforms will extend that momentum into mission-critical enterprise infrastructure to give organizations greater choice in how they deploy AI.”

The ecosystem is the asset being imported, and IBM is the importer. Arm counts more than 22 million developers, which it puts at over 80% of the world’s total, along with the framework libraries most AI work is already built on. A mainframe that speaks Arm inherits all of it without anyone porting anything.

Arm’s Other Business

Arm’s loud business right now is selling finished server chips straight to hyperscalers, moving from roughly ten dollars of royalty per device toward silicon with a four-figure price. That shift has driven most of the attention on the company this year.

This is the quieter one, and it works nothing like it. Arm sells no chip here, builds no hardware, and takes on no manufacturing. IBM designs the processor, fabricates it and sells the systems, and Arm’s architecture ends up running inside them. Terms were not disclosed, so the economics are not public. The strategic effect is legible anyway: the architecture reaches a class of machine it has never run on, inside institutions that will keep those machines for another decade.

Arm is on phones, is taking share in servers and PCs, and now has a route into the systems that clear the world’s payments. IBM gets a hardware refresh cycle, which is real, bounded, and something it has captured many times before.

The Payoff Is Years Out

No ship date, first revenue years away, and a product this distant can be delayed or quietly rescoped. Nobody should reprice anything on Monday’s news.

What the announcement settles is direction. The most conservative buyers in computing have the most to lose from change. They have now signalled that their next platform will speak the language of modern AI software, instead of asking that software to speak theirs. Companies positioned as architecture owners rather than component vendors are the ones that collect when that pattern repeats, and Arm is the clearest example most readers already know.

The chip itself will settle very little. Watch whether other closed platforms make the same choice, and how quickly the institutions carrying that $330 billion in maintenance start moving AI work onto hardware they already trust.

Subscribe now

Subscribe now

Bulls lost more ground on Monday after semiconductors and artificial intelligence infrastructure stocks moved broadly lower. The S&P 500 closed at 7,653, a decline of 0.3%. The weakness for the benchmark index masks even more aggressive selling in shares of key businesses tied to the AI data center buildout. Bears are building a narrative that these centers simply will not be built as political opposition grows throughout the United States and most of the West. Some of this is politics as usual in an election year, but professional money managers are trimming positions as shares lose upside momentum.

Speaking of politics, the prospect of a new trade war with Canada is not helping bulls. Over the weekend, the promised trade deal with Canada collapsed as the two sides could not agree on the final details of the pact. Prime Minister Mark Carney promised to match new tariffs on Canadian goods set by the U.S., dollar for dollar. President Donald Trump said that he plans to increase tariffs to 50% on all cars, trucks, automotive parts, and steel imported from Canada. The caveat is these new tariffs will start January 1, 2027, conveniently after the Congressional midterm elections in November.

Treasury Secretary Scott Bessent was busy on Monday trying to shore up the fixed-income market, even as he outlined new crippling sanctions on Iran. CNBC reported that the Treasury Department plans to use the General Account to fund $4 trillion in longer-term Treasury buybacks. The 10-year Treasury note yield fell back below 4.7%, and the 30-year Treasury bond yield fell to 5.21% after trading as high as 5.3% last week.

Separately, Bessent’s Operation Economic Outcast will attempt to cut off Iran from the global economy. The plan enforces secondary sanctions for third-party nations, foreign banks, and businesses that trade with and process revenues with Tehran. Oil prices gave up the gains made on Friday. West Texas Intermediate crude oil slipped 2.5% to $84.87 per barrel.

S&P 500 sectors were mostly higher on Monday. Eight of the 11 sectors finished in the green. Consumer staples, financials, utilities, and communication services all posted gains of at least 1.0%. Consumer staples rallied 1.8% on the back of snapback rallies for Walmart, Costco, and Philip Morris. On the flip side, technology was the biggest loser, falling by 1.6%. Politico ran a report Monday that showed data centers now face bipartisan opposition. The new bearish narrative is these centers will not be built, especially if Democrats take control of Congress in November.

Read more

Subscribe now

NVIDIA took a frontier AI model that could solve just 30% of a notoriously difficult benchmark and pushed it to a perfect score.

The model was Anthropic’s Claude Opus 5, and nothing inside it changed: no retraining, no fine-tuning, not one adjusted weight. Everything that improved sat outside the model, in a software system Nvidia calls AVO that manages the model’s memory, plans its next moves, and watches for mistakes. Better software pulled three times more performance out of a model that already exists.

What AVO Actually Is

The software layer between a model and a task is called a harness, and Nvidia’s researchers describe its job in one clean passage:

A frontier language model is only one component of an AI agent. The surrounding agent system—often called a harness—determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and sustains progress over long-running tasks.

AVO, short for Agentic Variation Operators, is Nvidia’s harness, and it has four working parts:

• A persistent memory carries forward everything the agent has tried, learned, and reasoned through, so it picks up where it left off instead of rebuilding its understanding from scratch.

• A supervisor watches the main agent the way a manager watches a new hire, stepping in when progress stalls and redirecting it toward a different strategy.

• An iterative loop drives the work: inspect the situation, plan, act, evaluate the result, repeat.

• Swappable tools adapt the system to whatever job it is given.

The swappable tools are what make AVO general rather than a one-benchmark trick. Nvidia first built the system to optimize its own GPU code, then replaced the code tools with game controls and pointed it at a completely different job, where it performed at the frontier again.

The model supplies the intelligence. The harness keeps that intelligence pointed at the job long enough to finish it.

The benchmark comes from the ARC Prize team, and it is built to resist what AI models are usually good at. ARC-AGI-3 drops an agent into game-like worlds with no instructions, no explicit rules, and no stated goal.

The agent has to figure out what each game even wants through trial and error, the way a person handed an unlabeled puzzle would. Memorized knowledge does not help, because the agent has never seen these games before. What gets measured is learning on the fly, the skill long business tasks actually require.

The public set holds 25 environments of six to ten levels each, 183 levels in all, and scoring combines completion with action efficiency, benchmarked against humans playing the same games for the first time. AVO, running Claude Opus 5 underneath, completed every level in 6,624 actions, about 12 percent fewer than the previous most efficient system. The bare model completes about 30 percent.

The Pilot Problem

Enterprises do not buy benchmark scores. They buy work that gets finished, and that is where AI adoption has been stuck.

MIT’s NANDA initiative found that only about 5 percent of the enterprise AI initiatives it examined produced measurable business value, despite an estimated $30 billion to $40 billion in spending. The failures were rarely about model quality. Companies could not get AI to fit into existing work, remember what it had learned, or improve with feedback. An AI system that is expensive to run and unreliable is a business liability, not a productivity tool.

A harness attacks that failure mode directly. Memory means the agent stops repeating yesterday’s mistakes. Supervision means it stops burning money on dead ends. Fewer wasted actions lower the cost of every completed task while raising the quality of the result.

Companies do not need to wait for a new generation of models to expand what AI does for them. The models already in production carry far more capability than current software extracts, and drawing it out costs a fraction of what training a new frontier model does.

None of this shrinks the compute story. An agent that works a task for hours consumes far more inference than a chatbot answering a question, so dependable agents grow the compute bill. Each big drop in the cost of computing has expanded how much of it gets used. Reliability works the same way.

Where The Value Lands

The investor question is who profits when AI moves from impressive demos to dependable work. Companies like Microsoft, ServiceNow, Salesforce, and Palantir already sit inside the workflows enterprises want agents to run, which puts them exactly where the harness meets real work. The harness needs the context those platforms own: the documents, the tickets, the customer records, the code.

Then there is the company that ran the experiment. Nvidia built AVO in-house, an agent system it also uses to optimize its own GPU code. The same week the benchmark result landed, Nvidia reportedly agreed to pay $6 billion to license the model-building software of the coding startup Poolside. The company that sells the chips is assembling the software layers above them, one deliberate move at a time.

Enterprises stalled on AI because the jobs rarely got finished. The software that finishes them has started to arrive, and the companies supplying it are the ones positioned to collect.

Subscribe now

Subscribe now

Almost everything went wrong for bulls last week, and they back-stepped to a large loss over the course of the five sessions. The S&P 500 fell to 7,674, a decline of 1.4% on the week. This was masked somewhat by a small rally on Friday, when the benchmark index gained 0.4%. Professional money managers are being pushed into a corner. Higher energy prices and bond yields are being goosed by the war with Iran. That conflict shows no signs of ending anytime soon. Bears are using higher bond yields to revive their long-time narrative that the federal deficit is too high and a major debt crisis is unfolding before our eyes.

Bears have been talking about a debt crisis for the past 40 years. We are not making the argument that government spending is irrelevant. Governments globally are spending too much and purposefully deflating the value of their currencies. However, they have been doing so for 40 years, and as fixed-income traders have bought bonds, stock prices have risen. Crises only evolve when bond traders demonstrably want governments to rein in the spending. In this case, fixed-income traders want President Trump to end the war with Iran and get some control over the direction of energy prices.

We know this is true because Treasury bond yields collapse the moment the President hints the war is in the diplomatic phase. We can also see that high-risk corporate debt is still being bought. The iShares High Yield Corporate Bond exchange-traded fund is pennies from its record high set five days ago.

The problem for Bulls is that President Trump refuses to tone down the rhetoric with Iran. To make matters worse, he is ramping up trade deal tensions. A deal with Canada blew up over the weekend, and the President said he would impose new tariffs on Colombia for refusing to accept deportees. The 10-year Treasury note ended Friday at 4.73%, the highest level since January 2025. High bond yields force professional money managers to re-evaluate asset allocation models.

Read more

The verdict at the time was that the internet had finally arrived for the warehouse club. What actually happened is that Costco’s earnings per share grew 14% that fiscal year while the multiple investors were willing to pay for those earnings fell from 30.8 times to 26.0. The business went one way and the price went the other. Costco regained its pre-announcement price on November 30, 2017, one hundred and sixty-seven days after the panic began.

Anyone who bought at that July low has since made 521%.

We bought our position in Compound & Conquer in March 2019, at $219, after the dust had settled.

Costco is having another bad year. The stock closed Thursday at $933.51, which is 14.7% below the record it set in May and roughly 6% below where it traded twelve months ago. Over that same twelve months the S&P 500 returned about 21%.

Inside the company, meanwhile, the numbers went the other way. Trailing earnings per share grew from $18.21 to $19.88, and operating income through the first three quarters of this fiscal year is up 12%. Membership renewals are still above 92% in the United States and Canada.

So this looks like 2017 again, and the temptation is to tell you it is.

The two episodes are arguing about different things. In 2017 the market was arguing about the business, and it was wrong. Today it is arguing about the price, and that question deserves an actual answer rather than a reassuring anecdote.

Answering it means taking the last ten and twenty years of Costco returns apart and separating the portion the business earned from the portion the market lent. Those two behave very differently from here. The business half is the best story in retail. The lent half is the reason a great company can be a mediocre investment for a while.

Table Of Contents

• The Business: How Costco Actually Makes Money

• The Engine: The Membership Machine

• The Moat: The Price Cap That Built The Moat

• The Numbers: A Decade Of Compounding In Two Layers

• Where The Stock’s Return Really Came From

• The Arithmetic Of A De-Rating

• Killing The Thesis: Five Ways Costco Can Disappoint You

• Conclusion: Great Business, Expensive Stock

The Business: How Costco Actually Makes Money

Costco describes itself in its own annual report with a sentence that has barely changed in thirty years:

“We operate membership warehouses and e-commerce sites based on the concept that offering low prices on a limited selection of nationally-branded and private-label products in a wide range of merchandise categories will produce high sales volumes and rapid inventory turnover.”

Every part of that sentence is load-bearing. Low prices bring the members. Limited selection is what makes the low prices possible, because buying one pallet of one brand of olive oil instead of eleven varieties gives Costco a volume position no supermarket can match. High sales volumes and rapid inventory turnover are the payoff, and they are the reason a company earning eleven cents of gross margin on the dollar can operate at all.

At the end of its 2025 fiscal year Costco ran 914 warehouses, 629 of them in the United States and Puerto Rico and 285 abroad, and it added 24 net new locations during the year. Those 914 buildings produced $269.9 billion of net sales, which works out to roughly $295 million per warehouse. That number is the whole company in a single statistic. It is what allows a business with almost no gross margin to cover its costs, and it is the number every competitor has to answer.

The merchandise itself is only part of what happens inside. Costco’s filings split the revenue three ways. There are the core categories, meaning foods and sundries, non-foods and fresh foods. There are the warehouse ancillary businesses: gasoline, pharmacy, optical, the food court, hearing aids and tire installation. Then there are the other businesses, which cover e-commerce, business centers and travel.

The ancillary operations sit inside or next to the warehouse for a reason the company states plainly: they encourage more frequent shopping. The gas station exists to bring you back.

Behind the buildings is a distribution system built around not warehousing things. Costco buys most merchandise directly from suppliers and routes it to cross-docking consolidation points it calls depots, which receive large shipments and ship them straight out to warehouses. Goods move through rather than sit. That is what makes the inventory turn fast enough for the model to work.

Then there is the pricing rule, which is the most quoted fact about the company and the most important. Costco’s long-stated internal policy caps the markup on any branded item at 14% over cost, and on any item under its Kirkland Signature house brand at 15%. Most retailers treat their markup as the output of the business. At Costco it is a constraint set in advance.

The result is a company that keeps very little of what passes through it. In fiscal 2025:

• Gross margin was 11.12% of net sales.

• Operating margin was 3.77%.

• Net margin was 2.94%.

Costco moved $275 billion of revenue last year and kept under three cents on the dollar.

That three cents explains both why the business is so hard to attack and why the stock is so sensitive to small operational wobbles.

The Engine: The Membership Machine

Ask anyone who follows Costco how the company makes money and you will hear the same sentence: it sells goods at cost and makes its profit on membership fees. It is the single most repeated claim about the business. It is also, as stated, wrong, and it is getting more wrong every year.

Membership fees were 51.3% of operating income in fiscal 2025, a bit over half rather than all of it, and that share has fallen five percentage points in two years. The reason it is falling is the interesting part: merchandise profit is growing faster than fee income. Gross margin has widened from 10.57% to 11.12% over the same three years. The retail operation earns real money in its own right, and it is earning more of it than it used to.

The fee side is still extraordinary, and it works because of renewals. The complete series, as filed, is rarely assembled in one place:

The worldwide rate peaked in fiscal 2024 and both series gave ground in fiscal 2025. The US and Canada rate gave back about 60 basis points, and the worldwide rate is back below 90% for the first time since fiscal 2021. Bears have noticed.

What they have generally not done is read Costco’s explanation, which sits in the same paragraph of the same filing that discloses the decline. Renewal rates were, in the company’s words, “negatively impacted by a higher number of memberships sold online, including through digital promotions, entering the renewal rate calculation.” Cheap sign-ups renew at lower rates than people who walked into a building and paid at the counter. The measure is also a trailing calculation covering renewals from seven to eighteen months before the reporting date. The number that slipped is therefore reporting on members Costco signed up cheaply as long as two years ago, well before anyone renewed or failed to.

Meanwhile the valuable end of the membership base is growing much faster than the base itself. Executive members pay an extra $65 a year, earn 2% back on purchases, and now make up 47.8% of all paid members, up from 45.5% two years ago. In the most recent quarter paid members grew about 4% while executive members grew nearly 10%.

That mix shift matters because of what happened on September 1, 2024, when Costco raised its fees for the first time since 2017:

• The annual Gold Star and Business fee, the basic tiers, went from $60 to $65.

• The Executive membership went from $120 to $130.

• The cap on the 2% Executive reward went from $1,000 to $1,250.

Costco recognises membership revenue “ratably over the one-year membership period,” which means a fee increase does not land in one quarter. It seeps into reported revenue across roughly two fiscal years.

You can see it seeping right now. Through the first three quarters of fiscal 2026, membership fee revenue grew 12.7% while net sales grew 9.7% and paid membership grew around 4%. That gap between 12.7% and 4% is the price increase working its way through deferred revenue rather than a wave of new members, and it has a finish line.

The fiscal 2025 filing also splits comparable-sales growth into its two parts, and that split is the best test of whether the value proposition is intact. Comparable sales grew 6%. Of that, shopping frequency contributed about 5% and average ticket about 1%. The growth came from members visiting more often. The average basket barely moved.

The Price Cap That Built The Moat

The standard question about a business this profitable is why nobody has copied it. The warehouse club format is not patented. Sam’s Club opened in 1983, the same year Costco did, and Walmart has spent four decades and enormous resources on the format.

The answer is that the format was always the easy part. What costs money is keeping the promise underneath it, every year, for four decades.

A 14% markup cap is a commitment device. It removes the lever every other retailer reaches for when a quarter goes badly. A supermarket with a soft month raises prices on the items shoppers do not track closely. That works, briefly, and it costs the retailer a little trust that nobody measures.

Costco has given that lever up on purpose. When costs rise, the margin absorbs it. When costs fall, the price comes down.

Ron Vachris described the policy on the most recent earnings call, saying Costco is “the first to come down and the last to go up.”

You can watch the company choosing that trade in real time.

Gross margin in the third quarter of fiscal 2026 fell 21 basis points against the prior year, and management pointed at deliberate price cuts on Kirkland Signature items. Margin went down, trips went up, membership income compounded. That is the design, and it only works if members believe the price is always right, which is a belief that takes decades to build and one bad year to spend.

The same logic runs through the payroll. Costco raised its US and Canadian starting wage to at least $20.00 an hour in March 2025 and lifted the top of its scales by a dollar. That brought the average US hourly rate to roughly $32.00. In the same year about 7,500 hourly employees completed a six-week management development course. The current chief executive, Ron Vachris, joined Price Club, the warehouse chain Costco later merged with, as a part-time forklift driver in 1982 and is the third CEO in the company’s history.

Jim Sinegal, who founded the company with Jeff Brotman and ran it until 2011, explained the economics of that decision in a 2013 interview more clearly than any consultant has since:

“Our feeling has always been, ‘You get what you pay for.’ If you’re going to pay low wages, you’re not going to get the productivity that you’re going to get if you hire really good people.”

High wages inside an 11% gross margin look like a contradiction until you notice that Costco’s sales per warehouse are roughly $295 million against something closer to $171 million for a typical Sam’s Club location. Costco pays more per hour and moves far more revenue through every building it pays those wages in. The second number is what makes the first affordable.

Sinegal also gave the cleanest available description of why this business and the market that prices it are so often out of step:

“We have been in the business of trying to build a company that’s here for years and years and years. Wall Street is generally in the business of trying to make money between now and next Tuesday, so there is that difference.”

He added that this was not a knock on Wall Street, and that the system had been very good to Costco.

The Numbers: A Decade Of Compounding In Two Layers

The operating record comes first, because it has been magnificent.

Over nine years revenue grew at 9.8% a year, operating income at 12.2%, and earnings per share at 14.6%. Each of those numbers is larger than the one before it, which is the signature of operating leverage. Costco sold 2.3 times as much and earned 3.4 times as much per share, because operating margin climbed from 3.09% to 3.77% while selling, general and administrative expense fell from 10.17% of revenue to 9.07%.

Costco got more profitable by getting bigger, which is exactly what the model promises and almost nothing in retail delivers.

The current year is not a break in the pattern. Through three quarters of fiscal 2026, revenue is up 9.7%, membership fees up 12.7%, and operating income up 12.0%, with operating margin higher year over year in every single quarter.

Free cash flow in fiscal 2025 was $7.8 billion on $13.3 billion of operating cash flow, with capital spending of $5.5 billion, up from $2.8 billion five years ago as the build-out accelerates. The company is paying for its own expansion out of the till and still has billions left over.

The diluted share count has drifted slightly up, from 443.2 million to 443.5 million over the past year. Costco pays a dividend, raises it, and occasionally pays a large special one, including $15.00 a share in January 2024 and $7.00 in 2017. It does not meaningfully shrink the share count. Whatever has driven the stock, buybacks are not on the list.

So the operating record is not in question, and nothing in it explains the last twelve months. The explanation sits in the second layer.

Where The Stock’s Return Really Came From

Read more

Subscribe now

Bulls continued to follow bond prices on Thursday. The result was another large decline and more resolve among bears. The S&P 500 closed at 7,641, a decline of 0.9%, leaving the benchmark index below its 20-day moving average for the first time since July 30. The date is significant because shortly thereafter, the White House embarked on a series of steps to bolster bond prices, and a significant stock market rally ensued.

Professional money managers are very much aware of the politics driving financial markets, and they are reluctant to move too far into the bear camp. However, pros are now being pushed to their logical limits because bearish narratives are ramping up.

The chief bearish storyline is about the surging federal deficit. The key marker crossed $40 trillion earlier in the week, the highest level on record. The United States will spend $1.8 trillion in 2026 on interest payments alone. American deficits are not usually a headwind for financial markets; however, President Trump is currently locked into an expensive war with Iran, and he shows no sign of backing away from the fight. In fact, he ramped up the rhetoric on Wednesday when he announced an economic doomsday for Iran unless the country completely surrenders. That message led to even higher energy prices and more weakness for Treasury bonds.

West Texas Intermediate crude oil closed Thursday at $86.15 per barrel, up 2.1%. International Brent crude oil finished at $93.00 per barrel, up 1.5%. The 10-year Treasury note yield surged back above the important 4.7% level.

On Wednesday, Treasury Secretary Scott Bessent tried to reassure bond traders that his department would provide liquidity via an aggressive new repurchase program. That initiative caused some reflexive buying, and yields did briefly fall below 4.7%. However, the underlying problem is the war with Iran. Without some resolution to that conflict, bond traders are likely to keep selling U.S. Treasuries, sending yields higher. Bessent can talk about his preference for lower bond yields, but he will need either the Federal Reserve or his boss in the White House to change policies to make a meaningful difference. In the interim, fixed-income traders are likely to keep selling, pushing yields up.

Read more

Three Months From Market Selloff To The Most Expensive New Car Ever.

How We Got Here:

Our Luce Sell-Out Prediction · Our Ferrari ($RACE) Deep Dive

Subscribe now

This is our third report on the Ferrari Luce, the car investors sold over a photograph. When Ferrari unveiled its first electric car in Rome on May 25, the reaction was instant: the Milan shares fell about 8 percent the next day and the New York listing closed at $329.91.

We made the case at the time, in Controversial New Ferrari is a Gift to Investors, that the selloff was the mistake and the car would find its buyers.

Two weeks later we published full Deep Dive on Ferrari, one of the seventy names in our Compound & Conquer compounders list, for anyone who wants the complete case.

Ferrari’s New York shares closed Tuesday at $412.44, up about 25 percent from that May 26 close.

The stock chart is the slow verdict on that argument. The fast one landed on Saturday.

At RM Sotheby’s Monterey Car Week sale, the first production Luce became the most expensive new car ever sold at auction at $40 million. The estimate was $1.1 million. The bidding ran 36 times past it.

The $40 Million Verdict

The car is Chassis 0, the first Luce built. Ferrari finished it in a one-off Tailor Made specification: pearl semi-gloss paint mixed with a custom pigment that shifts from green to violet as the light moves, laid over metallic Perla Le Mans leather. After the sale it returns to Maranello, then reaches its unnamed buyer in the first quarter of 2027. Every dollar of the $40 million funds the Ferrari Foundation’s educational programs.

The charity setting gave bidders a public-minded reason to stretch, but a good cause alone cannot manufacture demand at this scale. A charity lot still needs people in the room fighting to own the thing, and this one had several. The bidding is the signal. The wealthiest people in Ferrari’s orbit wanted this car badly enough to compete for it in public, at a price no new car has ever reached.

What The Sellers Missed

The case for selling in May rested entirely on the product: the styling, the $640,000 starting price, and the idea that an electric drivetrain betrayed everything the badge stood for. The Luce carried four electric motors, 1,035 horsepower, five seats, and the first car ever designed by Jony Ive and Marc Newson’s LoveFrom studio. Investors looked at all of that and treated it as a product problem.

Ferrari’s business runs on scarcity, status, and provenance. It builds fewer cars than it could sell, on purpose.

In 2024, roughly 81 percent of new Ferraris went to existing owners, and nearly half went to people who already owned more than one. Customers that loyal are buying standing in Maranello along with the machine. Our June piece put the whole situation in one line:

Ferrari rarely goes on sale. It is on sale now, and the reason is a picture.

Within three weeks of the reveal, the car the market called a mistake was sold out through the end of 2027.

The Collectible Logic

Seen from the collector’s side, the electric drivetrain added to the case. The Luce opened a new category inside Ferrari’s universe, and there will only ever be one first electric Ferrari. Add the first car Jony Ive ever designed, and Chassis 0 carries two layers of provenance no other object on the road can claim. Collectors were never being asked to accept a substitute for a proper Ferrari. They were offered the opening entry in a new chapter of the marque.

The auction confirmed that Ferrari’s most valuable customers understood this before the stock market did.

Not every dollar of the stock’s recovery traces to the Luce, and the honest reading credits a rising market for part of it. The May drop is a different matter. That decline was built entirely on a misreading of the product, and the shareholders who sold into it handed their shares to buyers who understood the company better.

The lesson is bigger than one car company. When a headline knocks a stock down, the useful question is whether anything inside the business moved. In May the answer was no. It took the market three months to admit it, and Ferrari’s customers a few minutes of bidding to settle it. The gap between reading a headline and understanding a business was worth $40 million on Saturday.

Subscribe now

Subscribe now

Bulls recouped some of their early-week losses on Wednesday after the Treasury Department intervened to halt rising Treasury bond yields. The S&P 500 rallied to 7,708, an advance of 0.2%. Although the gains were modest, the advance ended an important decline. It also revealed yet again that the White House is watching financial markets and fighting furiously to end a narrative it does not like. At this stage, professional money managers are captive to momentum. They cannot afford to have a repeat of 2025, when many of them failed to keep pace with the S&P 500 benchmark.

The session began when officials at the Treasury Department said it would double the size of its government bond repurchases. Global bond prices have been under pressure as fixed-income traders worry about the war with Iran and systemic inflation. Add this to ongoing fear about the sustainability of U.S. government debt, and you get the recipe for a credit crisis. The U.S. fiscal deficit jumped to $432.3 billion in July, its highest monthly total since March 2021.

The Federal Government spent $1.8 trillion in 2026 to finance the debt. The repurchase arrangement provides liquidity to nervous bond traders who are less willing to finance increasing fiscal deficits at current rates. The goal is to send a message that the Treasury Department is watching and prefers lower rates.

This is not quantitative easing or money printing. That distinction is important and was reflected in late-session selling. The Federal Reserve is not creating reserve dollars. The Treasury Department is merely swapping one form of debt for another. Ultimately, the 30-year Treasury bond yield slipped from its 19-year high of 5.33% on Tuesday to 5.183%. The 10-year Treasury note yield fell modestly to 4.637%. Both metrics are lower than on Tuesday; however, yields remain too high for the models that most professional money managers use to set asset allocations.

Read more

Subscribe now

Bulls retreated further on Tuesday as persistent hawkish rhetoric surrounding the war with Iran sent energy prices higher and risk assets lower. The S&P 500 closed at 7,692, down 0.7%. The loss was the biggest decline in a month, and leaves the benchmark index within easy striking distance of a key moving average. Theoretically, a decline to that level should entice professional money managers back into equities, with the back-up in interest rates and rising energy prices providing the primary caveat.

The Iran conflict moved closer to renewed escalation on Tuesday after a 60-day memorandum of understanding expired without a permanent ceasefire or follow-on negotiations. President Trump said no talks with Tehran were scheduled and he maintained the naval blockade. While the President insisted that the Strait of Hormuz remained open, Iran disputed that account, saying the waterway would stay closed until the U.S. lifted the blockade, sanctions, and other military pressures.

Separately, the United Arab Emirates reported detecting two ballistic missiles launched from Iran, both of which fell into the sea. With Iran reportedly preparing to adopt a more offensive posture and maritime traffic still severely disrupted, oil prices moved higher. West Texas Intermediate gained 0.52% to $84.94 per barrel, the highest level since July 24.

Worries about energy prices and the war are sending bond yields higher, causing even more anxiety among professional money managers. Although they have been buyers of dips, they spent all of Tuesday unwinding risk assets.

Semiconductors led Tuesday’s sell-off, with the Philadelphia Semiconductor Index falling roughly 5.5%. Investors rotated away from AI and memory stocks. The rise in long-term borrowing costs has an adverse impact on professional economic models because it raises the discount rate applied to distant earnings. Higher rates also increase concern that data center investment may take longer to generate returns. Nvidia fell about 2.3%, Micron lost 7.0%, while SanDisk and Western Digital dropped roughly 9.0% and 7.4%, respectively.

Read more