Penguin Solutions closed Wednesday at $72.61, up 13% on the day and about 20% across 2 sessions. Fiscal fourth-quarter results landed after Tuesday’s close: record sales of $566.7m against a $521.0m consensus, adjusted earnings of $1.00 a share against $0.77 expected, and fiscal 2027 revenue guided up about 40%. From its April 2025 low of $14.53 the stock is up 400%.

The results were genuinely good. What follows isn’t an argument that they weren’t, it’s an argument about what still has to be proven.

5 Things Worth Knowing

1. Fiscal 2026 net sales rose 26% to $1,731.5m. Integrated Memory grew 99% to $924.0m, more than accounting for the group’s entire increase, since the other 2 segments declined. Memory went from 34% of sales to 53%, and 60% of the fourth quarter.

2. Advanced Computing fell 14% to $558.8m, but not because AI demand failed. Management attributes the decline to the Penguin Edge wind-down and lower hyperscale hardware, and reports non-hyperscale AI infrastructure sales up 72% for the year and 99% in the fourth quarter. Both segments are guided to grow in fiscal 2027, Advanced Computing by about 40% and memory by about 50%.

3. The company earned $180.6m on a GAAP basis and used $151.9m of operating cash. Capital spending was $12m, under 1% of sales. This is an asset-light business with a substantial appetite for financing.

4. Working-capital discipline was better than the balance sheet alone suggests. Receivables and payables scale with gross billings of $1.57bn in the quarter, not the $567m of net sales reported, because memory logistics is accounted for on an agent basis. Days sales outstanding were 46, against 51 a year earlier.

5. On management’s own adjusted guidance of $4.45 a share the stock trades at 16.3x, not the 14x the screen suggested when I started. Those are strong-cycle earnings.

Where the Growth Came From

Contract prices for conventional DRAM rose sharply through the first half of calendar 2026, with the major forecasters publishing quarterly increases in the high double digits. Penguin’s memory revenue roughly doubled over a fiscal year sitting inside that window.

How much of that is price and how much is units matters a great deal, and the company doesn’t break it out. Management attributes growth to both, and the DDR5 replacement cycle in its enterprise and networking base is a real volume story. But industry price data can’t be mapped onto one company’s realised selling prices: the periods, product mix, customer base and revenue-recognition basis all differ. Anyone telling you they’ve decomposed it from public figures is guessing.

What can be said is that the mix has shifted hard toward a business whose economics are set by a commodity cycle, and that this is the first of 3 things worth monitoring.

On margins, I’d resist the obvious inference. Full-year adjusted gross margin fell 1.7 points to 29.3%, and the temptation is to read that as memory input costs squeezing a module assembler. Management says the opposite: the decline came primarily from a greater contribution of AI infrastructure hardware, with stronger margins inside Integrated Memory. A consolidated percentage can’t settle what happened inside a segment, and a lower percentage sat alongside record gross profit dollars of $507.9m.

The Cash Question

This is the part that deserves the most attention, and it needs stating carefully.

Receivables rose $488m and inventories $494m during the year. Those are large numbers. But payables, accruals and other liabilities rose $607m, funding most of the build, and the collection metrics moved the right way against the prior year: receivables at 46 days against 51, inventory at 48 against 51, and a cash conversion cycle of 42 days against 49. The sequential picture is less flattering, with that conversion cycle lengthening from 33 days in the third quarter.

What remains true is the headline: $180.6m of accounting profit, $151.9m of cash consumed. Management attributes the outflow primarily to working capital supporting growth, and the mechanism is clear enough. When revenue grows 26% and the fourth quarter alone grows 68%, the balance sheet scales ahead of collections, and here it scales against gross billings several times larger than reported revenue. Worth noting that GAAP profit also carried a $33.9m income-tax benefit, against a $20.1m charge the year before, so the distance between profit and cash isn’t purely a working-capital story.

The useful framing is the one the capital spending exposes. Penguin spent $12m on capital expenditure all year, under 1% of sales. It needs almost no factories. What it needs is working capital, and growth consumes it. The company ended the year with $647m of cash, having raised $750m of zero-coupon convertible notes and repaid its $100m credit facility, so the funding is in place. But a business that converts growth into cash and a business that converts growth into financing requirements are different propositions, and 1 year doesn’t tell you which this is.

What the Multiple Is, and What Could Break It

Adjusted guidance of $4.45 puts the stock at 16.3x. GAAP guidance of $3.50 puts it at 20.7x. Screen multiples closer to 14x rest on consensus estimates above the company’s own midpoint, so use the company’s. At 16x for 40% guided growth, this isn’t expensive on its face.

Showing the multiple is 16x rather than 14x proves nothing by itself. The relevant exercise is what happens if these earnings don’t persist:

• At the $4.45 guidance midpoint, 16.3x

• At $3.56, 20% below the midpoint, 20.4x

• At $2.67, 40% below, 27.2x

Those are sensitivities, not forecasts. I don’t know when the memory cycle turns, and recent industry commentary runs from tightness persisting into 2028 to a profit peak well before that. But they make the concern concrete: at this price you’re paying a reasonable multiple provided the earnings hold, and a demanding one if they normalise.

The dilution picture is milder than the financing history suggests. The fourth quarter already used 62m adjusted diluted shares against the roughly 63m assumed in fiscal 2027 guidance, so the share count is barely moving. July’s refinancing issued about 8.7m shares in exchanges, and that has happened. The new 2031 notes convert initially near $116.70, some 61% above today’s price, with capped calls lifting the point of net dilution toward $175.05. SK Telecom’s $200m preferred, struck at $32.81, is genuinely in the money, and it’s already inside those diluted counts.

The AI Business Has Substance

It’s worth being specific here, because the narrative risk cuts both ways. Penguin was selected to deploy and operate a 36,000-GPU AI factory in Norway for a neocloud provider. The customer has $10bn of contracted compute from a leading AI lab, which is the customer’s figure and not Penguin’s revenue. Contract value, margins and phasing weren’t disclosed.

That’s the third thing to monitor: whether deployments of this kind produce recurring service profits or mostly hardware revenue at hardware margins. Hardware is currently diluting the blended margin, and management describes multiyear managed-service engagements alongside it. The remaining question is how much recurring software and service profit follows the deployments. Guidance of about 40% growth in Advanced Computing says the company expects volume. Whether that volume carries services with it is what would change the quality of this business.

The Price Is Stretched

At $72.61 the stock sits roughly 80% above its 200-day average and around 35% above its 50-day, with the daily relative strength index at 74.7. That index measures the balance of recent gains against recent losses, and a reading this high says gains have substantially outweighed losses, which is a statement about momentum rather than value. Published moving averages differ slightly by source and date, so treat the exact figures as approximate.

Wednesday’s roughly 21m shares were nearly 12 times the average of the preceding 10 sessions. It was an outlier by any baseline you pick.

This stock has also already broken once in this run. It closed at $81.39 on 9 July, then closed as low as $43.70 within the same month, a fall of 46%, and spent the 3 months since between roughly $46 and $64. Nothing here moves gently.

Analysts raised their targets after the results, and the disagreement survived. Rosenblatt moved to $100 from $80, while Goldman Sachs, Needham and Stifel all went to $85, each keeping a buy. Barclays also raised, from $40 to $60, and kept an underweight rating, flagging expected pressure on gross margins. Consensus sits near $83.57. Coverage is thin, with sources listing between 7 and 9 firms.

Our Trade Plan

Read more

BMW trades at 5.2x trailing earnings. Mercedes-Benz at 7.6x. Volkswagen at 6.5x. All 3 sit below half of book value, within a few points of 52-week lows, and carry yields between 7.8% and 8.8%. On a screen, this looks like a shelf of bargains with BMW at the front.

The screen is the wrong place to start. Every one of those numbers is distorted by something, and fixing the distortions changes the ranking.

5 Things Worth Knowing

1. Trailing multiples here divide today’s price by the last 12 months, a period already overtaken by events. BMW’s forward multiple sits above its trailing one, which only happens when earnings are expected to fall.

2. BMW cut its automotive margin guidance on 16 June, from a 4% to 6% corridor down to 1% to 3%, and confirmed that range in July. Second-quarter automotive EBIT was €629m, a 2.3% margin, with tariffs costing 1.25 points and Chinese joint-venture depreciation another 1.2.

3. Mercedes grew group EBIT 22% in the second quarter, but its car division went the other way: Cars adjusted EBIT fell 26% to €909m, a 4.0% margin against 5.1% a year earlier. The group number improved on financial services, vans and consolidation items, not cars.

4. At division level across the first half, the 3 sit closer together than their multiples suggest: BMW’s automotive segment made 3.6%, Mercedes Cars about 4.0%, Volkswagen’s Brand Group Core 4.9%.

5. Returns on equity have collapsed since 2023, from 12.6% to 7.6% at BMW and 15.5% to 5.5% at Mercedes. Volkswagen’s peaked at 9.4% and is now 3.9%. That explains the discounts to book better than any multiple.

Why the Multiple Misleads

A P/E has 2 parts, and almost all the movement here has been in the bottom one, which is also the part most distorted by one-offs.

BMW’s 2022 profit of €17.9bn looks like the start of a catastrophic slide to €7.3bn. But that year included a €7.6bn pre-tax revaluation gain from consolidating the BMW Brilliance joint venture, booked in the financial result rather than earned from selling cars. The after-tax effect is smaller than that headline, so I won’t put a number on the adjustment. What matters is that 2022 isn’t a clean base.

Measure all 3 from 2023 instead, which is free of that particular distortion and is Volkswagen’s actual peak, and the order changes: BMW is down 35%, Volkswagen 56%, Mercedes 64%. These are still reported figures, and later years carry exceptional items of their own, but the company with the cheapest trailing multiple has had the shallowest reported decline of the 3.

What the Car Businesses Actually Earn

Group numbers hide more than they show, because all 3 run large financial services arms whose earnings move independently of cars.

Across the first half of 2026, BMW’s automotive segment made a 3.6% EBIT margin. Mercedes-Benz Cars ran at 4.1% in the first quarter and 4.0% in the second on an adjusted return-on-sales basis. Volkswagen’s Brand Group Core, which holds its volume car brands, made a 4.9% operating margin, within a group operating margin of 3.8%, or 4.3% before special items.

Those figures aren’t defined identically: a reported segment EBIT margin at BMW, an adjusted return on sales at Mercedes, an operating margin at Volkswagen, with different segment boundaries at each. Treat them as the same order of magnitude, not a ranking. What they establish is that all 3 car operations run in the low single digits against the 8% or better they once treated as normal, and that no group headline should be read without checking the division underneath. Volkswagen’s own finance chief called a 3.8% group margin too low.

The causes are shared. German deliveries in China fell between 30% and 41% year-on-year in the second quarter, against a market Volkswagen describes as down about 20%, so the market contracted and the German marques lost share inside it. Tariffs cost BMW 1.25 margin points in the quarter alone, and electrification meant heavy spending into demand that arrived slower than planned.

The Case Against Each

BMW. The guidance cut is real and the second quarter was poor, but the cash position is less alarming than the profit line. First-half automotive free cash flow of €1,290m was down 45%, yet full-year guidance of more than €2.5bn requires only about €1.21bn in the second half, less than the first half produced. Whether conditions allow it is another matter, and BMW has flagged restructuring charges for the period. Automotive net financial assets stood at €42.6bn at the end of June, above the entire €32.6bn market value, and down €1.7bn over the half. Deutsche Bank, still rating the stock a buy, cut its target from €90 to €78 on 29 September and to €71 within the week.

Mercedes. First-half group EBIT fell 3.1%, net profit 6.3% and industrial free cash flow 29.9%. Second-quarter industrial free cash flow of €1.1bn included about €0.4bn from selling down part of the Daimler Truck stake; excluding those proceeds it was nearer €0.7bn, still carrying restructuring and other exceptional items. Against that, it holds €30.4bn of net industrial liquidity against a €37.9bn market value, completed a €2bn buyback in June, and is the only 1 guiding full-year group EBIT above the prior year.

Volkswagen. Reported profit is being destroyed by writedowns while the cash position held: Automotive Division net cash flow of €3.166bn in the first half, against an outflow of €1.350bn a year earlier, and net liquidity of €32.75bn at 30 June. That liquidity is close to the group’s entire €34.0bn market value. The September warning took reported 2026 operating margin to no more than 1% on about €10bn of special items, with roughly 4% indicated excluding them.

Below book is where most readers stop. It’s where the real question starts. Paying 34 cents for a euro of BMW equity, 40 at Mercedes and 19 at Volkswagen only matters if that equity can earn a return, and on net income over period-end equity all 3 have fallen hard since 2023. Volkswagen is the instructive one: its returns were modest even in the good years, peaking at 9.4%. Its discount to book is not purely cyclical.

The Volkswagen Arithmetic

Volkswagen’s 75.4% stake in listed Porsche AG is worth about €29.4bn, against a market value of just under €34.0bn for the whole group. That leaves roughly €4.5bn implied for Audi, Skoda, SEAT, Cupra, the Volkswagen brand, Bentley, Lamborghini, Ducati and Traton.

Hold that loosely. A controlling stake in a listed subsidiary would not change hands at the screen price, and the residual isn’t the market’s price for those brands either, since it carries the rest of the group’s liabilities and pension obligations. Volkswagen’s reported net liquidity also already includes Porsche’s cash, so the stake and the liquidity can’t be added together. The figure signals very low expectations. It’s not a valuation.

The Dividend Question

At these prices you’re paid to wait, and the question is whether the payment survives.

All 3 yields sit well above their own 5-year averages of 5.4% to 6.0%. That’s consistent with the market expecting lower distributions, and equally consistent with investors demanding more for the risk. It’s a reason to check the arithmetic, not evidence of a cut.

So check it. BMW’s stated policy is to pay out 30% to 40% of net income. Applied to forward consensus earnings of €8.75, that implies a dividend between €2.60 and €3.50, against €4.40 today, a reduction of 20% to 40%. That holds only if the estimate and the policy both hold, but it’s the arithmetic behind the 8.1% headline yield.

Mercedes screens better than BMW on the same test, with a €3.50 dividend against forward earnings of €6.75. Both forward figures are Yahoo Finance consensus at 5 October 2026 on an unstated basis, which limits how far either calculation can be pushed, and I haven’t run it on Volkswagen, whose forward estimate I wouldn’t rely on. Mercedes’ recent capacity to fund distributions leaned partly on an asset sale, and a completed buyback evidences past capacity rather than future commitment.

Where the Prices Sit

Over the past year BMW has ranged from €52.36 to €97.92, Mercedes €39.41 to €62.34, Volkswagen €67.02 to €109.15. Each trades 21% to 26% below its 200-day average, which establishes a broken trend and nothing more. Volkswagen is the only 1 at a genuine 5-year low; the other 2 sit around 7% above their 2022 lows. Consensus targets of €71, €56 and €104 are well above spot, and moving fast.

Where I Come Out

I prefer Mercedes of the 3, and I’m not buying it yet. Those are separate judgments.

The preference rests on 3 things: it’s the only one guiding full-year group profit higher, its dividend looks better covered than BMW’s on current forward estimates, and €30.4bn of net industrial liquidity against a €37.9bn market value is a substantial buffer, though BMW’s cushion on its own measure is larger still relative to market value. What holds me back is a car division still deteriorating, at 4.1% then 4.0%, and half-year cash generation down 30% with an asset sale inside the quarterly figure. The condition that would change it is checkable: a Cars adjusted margin that stops falling, with the third quarter the first chance to show it.

The other 2 need different things. Volkswagen needs the writedowns finished and returns on its very large equity base shown to be sustainable and capable of improving from 3.9%. At 19 cents on the euro of book, the bar isn’t a particular headline return, it’s whether the return that persists justifies the price. BMW needs the second-half cash guidance met and the dividend question settled.

Size any of them as cyclicals. The dividends are an output of earnings, and earnings are the thing in motion.

Bottom Line

A trailing multiple ranks these 3 accurately on its own terms. It’s simply a poor basis for choosing between them. Change the base year and the earnings declines reorder; look underneath the groups and 3 car operations in the low single digits look more alike than multiples of 5x, 6.5x and 7.6x suggest.

Cheapness isn’t one test here. It’s 3: whether the car business can earn again, whether the dividend is funded, and whether the assets can produce a return worth owning them for.

Mercedes comes closest on the second, which is why it’s my preference and why I’d still wait for its car margin to stop falling. Volkswagen’s long record of modest returns is reason for caution on the third, not proof it fails. And BMW is where the headline and the underlying numbers diverge most, which is usually where the work is worth doing.

This is research and commentary, not personal investment advice. The author may hold positions in names discussed.

The stock fell on the release and kept falling, ending the extended session at $32.09 and last trading at $31.54 in pre-market dealing at 06:20 New York time, down roughly 10% from Thursday’s $35.15 close. That’s its lowest since 2013, and 82% below the peak of November 2021.

A fall that size invites the obvious question. The price has collapsed, but so has the profit behind it: Nike guided this year to $1.15 to $1.35 against $3.73 in fiscal 2024. Whether the shares are cheap depends on how much of that comes back, and how long it takes.

Key Takeaways

• At $31.54 you pay about 25 times the earnings Nike guides to this year, 15 times what it earned in fiscal 2026, and 8.5 times fiscal 2024. Which number is right depends on how far profitability recovers, and how fast.

• On the revenue Nike guides to, getting back to fiscal 2026 earnings needs a 7.3% net margin and getting back to fiscal 2024 needs 13.0%, against the 6.7% it managed last year. Nike earned 12.9% as recently as fiscal 2022.

• Revenue is still falling. Guidance is a high-single-digit decline, Greater China fell 22% and Converse 28%. North America wholesale, up 9%, is the one line growing properly.

• Dividend cover has gone from 3.05 times free cash flow in fiscal 2024 to 0.91 in 2026, and the buyback is paused. The balance sheet still has room, but less of it.

Prices are Friday pre-market, early and thin, against Thursday’s $35.15 close.

The Quarter Itself Was Not the Problem

Revenue of $11,213 million was down 4% reported and 5% currency-neutral. Gross margin still rose 60 basis points, which Nike puts down to lower warehousing and logistics costs. Selling and administrative expense fell 3% to $3,910 million, with operating overhead down 6% and demand creation up 5%, so the company cut back office rather than marketing. Income before taxes was $921 million against $922 million, and on Nike’s own earnings before interest and tax, which it flags as non-GAAP, $907 million against $904 million.

Earnings per share fell a cent because the tax rate went to 22.7% from 21.1%, not because the business deteriorated. That’s the good-news case, and it’s genuine.

The Guidance, and Why Last Year Is Hard to Compare With

Nike guided fiscal 2027 revenue to decline high-single digits and adjusted earnings to $1.15 to $1.35, excluding about $0.15 of restructuring.

The cleanest way to see what that implies is to stay inside this year. Nike earned $0.48 in the quarter just reported, which leaves $0.67 to $0.87 for the remaining 9 months, about $0.26 a quarter. Nike gave no adjusted figure for the quarter and showed no separate restructuring line, so $0.48 is the reported number. If charges sat inside other expense lines the adjusted figure would be higher and the implied remainder smaller, so the comparison errs in Nike’s favor.

Comparing with last year needs more care than it’s getting. Fiscal 2026 earnings of $2.10 include a $0.52 benefit in the fourth quarter from the expected recovery of tariffs under the International Emergency Economic Powers Act, which the Supreme Court struck down in February. The benefit was $986 million and lifted that quarter’s gross margin by about 900 basis points. Nike says that across the full year the benefit largely offsets the tariff costs the year had already absorbed.

So the annual figure is roughly clean and the quarterly ones aren’t. On the annual basis the $1.25 midpoint sits 40% below $2.10, or about 45% below once both sides are adjusted for restructuring, Nike’s own definition. What you can’t do is set the guided 9 months against last year’s final 9 months, because the refund landed in a single quarter while the costs it offsets were spread across the year.

What You Are Actually Paying For

At $31.54 Nike is worth about $47 billion. That’s 25 times the $1.25 midpoint of this year’s guidance, and nearly 29 times the roughly $1.10 of reported earnings implied once restructuring is counted. On this year’s numbers alone the stock looks expensive, though a trough year is a poor basis for that judgement.

But nobody buying Nike here is buying this year. Against fiscal 2026’s $2.10 the same price is 15 times, and against fiscal 2024’s $3.73 it’s 8.5 times. Neither is a price you actually pay today. Both are prices conditional on earnings returning to a level Nike isn’t guiding to, and the argument is which of them describes the business in 3 years.

Work backwards from the revenue Nike guides to, about $43 billion. Earning $2.10 on that base needs a net margin of 7.3%, a little above the 6.7% Nike actually managed in fiscal 2026 and well above the 4.3% this year’s guidance implies. Earning $3.73 needs 13.0%, above the 11.1% of fiscal 2024. That isn’t unprecedented: Nike made 12.9% in fiscal 2022 and 12.9% in 2021, so the $3.73 case is a return to the margins of 4 years ago on a smaller revenue base. Revenue recovery would lower the bar rather than being strictly required.

The help already identified is harder to place than it looks. Pace is budgeted at roughly $2.5 billion of savings against $1.0 billion of charges, but both are cumulative through fiscal 2031 and stated before reinvestment. A cumulative total doesn’t establish annual earnings power, so it can’t be read straight into the margin for a particular year.

So the margin figures are the thing to hold on to. They tell you what has to happen; the cost program tells you only that management is working on it.

Reduced Flexibility

The balance sheet is what has to carry the company while that plays out. Nike declared $0.410 a share, so $1.64 annualised against a $1.25 adjusted midpoint: a payout ratio of 131% on guided earnings, 149% against the implied reported figure, and 121% to 143% across the guidance range.

Free cash flow was $6,617 million in fiscal 2024, $3,268 million in 2025 and $2,184 million in 2026, while dividends rose every year to $2,407 million. Cover went 3.05 times, then 1.42, then 0.91. Fiscal 2026 is the year it crossed, by $223 million.

That crossing is softer than it looks. Of the $986 million tariff benefit Nike booked, only $302 million arrived as cash by May 31. The other $684 million sat in receivables and the annual report says substantially all of it was collected after the year end, so a sum 3 times the shortfall moved out of fiscal 2026 cash flow and into fiscal 2027. It’s a real cushion, and a cushion that arrives once.

The buyback tells the same story. Repurchases went $5,480 million in fiscal 2023, then $4,250 million, $2,985 million and $146 million in 2026, with none in either of the last 2 quarters Nike has reported. That’s a pause, not an ending: $5.9 billion of the authorization is unused and the board renewed it in June 2026 with no expiry. But the diluted share count has turned, 1,484.2 million against 1,479.0 million a year ago, after 4 years of shrinking. Under the program Nike has bought 124.4 million shares at an average of $97.57. The stock is $31.54.

I’m not forecasting a cut. Nike holds $8.4 billion of cash and short-term investments against $7.9 billion of debt, and a board can fund a payout from reserves for a long time. What the numbers show is weaker coverage and less room than the company used to have, at the moment it needs room.

Where the Damage Is

Greater China is the worst of it. Revenue fell 22%, or 26% currency-neutral, to $1,180 million, and its earnings before interest and tax fell 34%. Converse is worse in percentage terms, down 28% to $263 million with profit down 36%.

The channel split is the more interesting read. Nike Direct fell 8%, with Brand Digital down 13%, while wholesale fell only 1% and North America wholesale rose 9%. That’s the deliberate reversal of the direct-to-consumer strategy working. Wholesale carries lower gross margins but also lower costs to serve, so the effect further down the income statement is less direct than it looks. By product, footwear fell 6% and apparel rose 2%, and footwear is the franchise.

The Case Against Me

Nike isn’t a broken brand. North America, its largest market, grew. Gross margin expanded in a down quarter, and the wholesale rebuild is working, which is the thing that had to.

The bull case doesn’t need my arithmetic to be wrong, only incomplete. A $2.5 billion savings figure is what management will commit to publicly, not the limit of what a program finds once running. Revenue recovery is excluded from my sums entirely, and a brand this size reclaiming a few points of growth changes the margin math quickly, because the cost base is being rebuilt smaller.

What I keep returning to is the clock. Nike has had almost 2 years under this management, the top line is still falling, China is still falling, and the cost program runs to 2031. That doesn’t make the shares expensive. It makes them a bet on duration.

Bottom Line

Nike delivered a quarter that did what it needed to: margin up, costs down, pre-tax profit flat on 4% lower revenue. Read only the first page and the turnaround is on track. The guidance describes a company earning a third of what it made in fiscal 2024, on revenue about $8 billion lower.

So is it cheap? Every answer is conditional. A multiple of 8.5 times fiscal 2024 earnings would be a low price if those earnings came back and held, and 15 times fiscal 2026 would be fair if $2.10 is reachable. The margin figures are how to judge which: 7.3% to get back to $2.10, 12.9% to get back to $3.73, against the 6.7% Nike managed last year and the 4.3% this year’s guidance implies. The first of those looks achievable, the second asks for 2022 again. What you can’t say is that the recovery comes free. Put a 15 times multiple on normalised earnings and $31.54 implies about $2.10, which is the fiscal 2026 outcome already. No recovery at all, on the same multiple against the guided $1.25, would be nearer $19. So the fiscal 2026 case is roughly in the price and the fiscal 2024 case isn’t, which makes the question whether Nike gets past 2026 earnings rather than back to them. I’d rather see revenue stop falling before paying for that. The weekly plan I publish had Nike at Avoid before the results, and the entry never triggered.

This is research and commentary, not personal investment advice. Levels are illustrative; size positions to your own risk tolerance and time horizon. The author may hold positions in names discussed.

The market spent Friday morning arguing with itself about it. The stock opened lower at $886.62, fell to $883.10, below the September low, then reversed all of it and reached $921.70. Around midday in New York it sits near $918, up about 2.4% from Thursday’s $896.48 close and 16.1% below its May high.

Those results answer one question well and another not at all. They tell you the operating model isn’t breaking. They can’t tell you what the shares are worth, and that is what the last 4 months have been asking.

Key Takeaways

• Membership fees of $5.91 billion were 50.6% of reported operating income, which is not the same as saying half the profit is a subscription.

• The tidy explanation, that investors rotated out of expensive retailers, doesn’t survive measurement. Across 20 names ranked on what they actually cost in May, the rank correlation with what followed is minus 0.21.

• At about 44 times the year just reported, Costco is on its lowest trailing multiple since January 2024. It is still dearer than 37% of the past 5 years.

• Fee income grew 11.0% for the year but only 7.3% in the fourth quarter, because the 2024 price increase has lapped. That belongs in the bear case.

Prices below are Friday intraday unless stated. The peer and valuation history end at Thursday’s close, the last session before these results, and are labelled where they appear.

Where the Money Comes From

Costco sold $297.2 billion of merchandise last year. After $264.3 billion of buying it and $27.2 billion of selling and administrative cost, $5.78 billion was left. Membership fees were $5.91 billion, which is 50.6% of the $11.69 billion of operating income.

Be careful with what that does and doesn’t establish. Costco charges all of its selling and administrative cost against merchandise, so the $5.78 billion is a lower bound and the split is a reporting convention, not a measurement. The warehouses are part of what the fee buys. There is no separately reported subscription profit, and anyone telling you half of Costco’s profit is a subscription is reading a presentation convention as a segment disclosure.

What survives is still worth having. A fee of $5.91 billion arrives from 84.1 million paid members before anything is scanned, while the retail operation converts $297.2 billion of sales into a 1.9% margin. The renewal rate rose to 89.8%, and executive members, who pay double, grew 9.4% to 42.3 million.

Payables of $22.6 billion exceed inventories of $19.3 billion, so suppliers fund the entire stock and then some, and Costco converts 102% of net income into free cash flow. Growth is 25 net new warehouses, from 914 to 939, plus 6.6% adjusted comps.

The Explanation That Didn’t Hold

Here is a story I nearly published. Measure the big retailers from Costco’s May 19 peak to Thursday. Walmart fell 19.8%, Costco 18.1%, Kroger 16.5% and BJ’s 6.0%, while Target rose 22.7% and Dollar General 18.4%. Rank those 6 by earnings multiple and the outcomes run almost exactly opposite: investors sold the expensive ones. Tidy, and I was ready to lead with it.

The test is fussier than it looks. Ranking on today’s multiple is circular, because a stock that fell 30% is cheap now precisely because it fell, so the ranking has to use each company’s trailing earnings as actually reported by May 19. Vendor trailing figures won’t do either: Kroger’s real May multiple was 14.5 times, not the 31 its current number implies, and Albertsons was 7.9 times, not 73.

Done that way, across 20 US retailers and distributors, the rank correlation between what a stock cost in May and what happened to it since is minus 0.21. Across the 6, it’s minus 0.54, well short of the minus 0.89 the first cut suggested. Start the same test on January 2 and the correlation is plus 0.04.

The scatter is the real finding. Casey’s was on 49 times in May and fell 30%, but Albertsons was on 8 times and fell 31%. Chefs’ Warehouse was on 37 times and rose 43%, Grocery Outlet on 12 times and rose 39%. Costco’s decline fits a spring stretch in which the expensive half of this group lagged the cheap half by about 10 points, but that gap reverses if you start in January, and the ordering is too scattered to say how much of Costco’s move came from it.

So I can’t explain the de-rating, and I’d rather say that than sell a pattern that dissolves when you measure it properly. The basket is COST, WMT, TGT, KR, BJ, DG, DLTR, ACI, SFM, CASY, PSMT, SYY, USFD, PFGC, CHEF, OLLI, GO, BURL, ROST and TJX. Returns are closing prices, May 19 to September 24.

Costco’s reported operating data do not show a consumer breakdown. US adjusted comps ran at 7.2% in the quarter against 6.6% for the year, with digital up 19.8% and renewals higher, at a retailer serving 150.4 million cardholders. That is evidence of resilience rather than proof that nothing changed in how people shop.

What You Actually Pay

Near $918 Costco is worth about $407 billion. Cash of $20.2 billion plus short-term investments of $1.1 billion less debt of $6.2 billion is $15.1 billion of net cash, which takes the enterprise value to roughly $392 billion. Subtract the reported long-term operating lease liability of $2.4 billion and that net figure becomes $12.7 billion, though that adjustment covers the long-term line only, not every lease obligation. That’s 44.2 times the year just reported and 33.5 times operating income, on a free cash flow yield of 2.3%.

Now put it in its own history. For each session I divide the closing price by the 4 most recent quarterly earnings per share Costco had already published, counting a new figure only from the session after it lands, because these results arrive after the close. On that basis Costco has spent 5 years between 33.6 and 64.6 times, with a median of 47.6.

Worth being careful here. Thursday’s market was still working with $19.88 of trailing earnings, so the observed multiple into the print was 45.1 times, and only the release itself dropped it. Today’s 44.2 is the lowest reading since January 4, 2024, which is to say in nearly 3 years, and still dearer than 37% of the sessions in that 5-year window. The stock got meaningfully cheaper than it has been through its whole recent run, and it did not get cheap.

At the May peak of $1,094.32, against the $19.23 of trailing earnings reported by then, the shares were at 56.9 times. Measured against the year they turned out to deliver, that peak was 52.7 times. The multiple compressed more than the share-price decline alone suggests, because trailing earnings rose from $19.23 to $20.76 over the period.

The Case Against Me

The bull case doesn’t need the multiple to expand. It needs it to stop falling.

Costco has a long record of resilient renewals and rising executive penetration, a strong net-cash balance sheet, and capital spending of 2.2% of sales while still adding 25 warehouses a year.

Against that, 2 things. The fee line is slowing: 11.0% for the year but 7.3% in the fourth quarter, because the increase that took effect in late 2024 has now fully lapped. From here the subscription compounds at the rate members are added, not the rate prices rise, and membership economics are central to the premium investors pay. And a de-rating from 57 times to the low 40s has no obvious floor, because nothing I can measure identifies what drove it.

The tariff line shows the model working as designed. The quarter carried a $0.15 per share benefit from IEEPA tariff refunds, which Costco says it partly reinvested in lower member prices rather than keeping. That is what a company does when it treats the fee as its return and the goods as the service. It’s also why the shareholder’s claim is thinner than the quality suggests.

The Setup

Costco is in the weekly plan I publish: regime Breakdown, reclaim above $929.93, status Avoid. That plan is dated September 12 and no newer one has been issued.

1. The trigger has not been hit, so the instruction stands. Thursday closed at $896.48, under the $898.43 level the plan flagged. Near $918 the stock has recovered above its 20-day average at $911.97 but remains below the 50-day at $930.25 and the 200-day at $954.24. Thursday’s 14-day relative strength reading of 38 and trend strength reading of 18, which signals a weak trend rather than a strong one, described a drift, not a washout.

2. This morning changed the downside reference. The $883.10 low undercut the September 21 low of $885.50 and was bought back within the session, the first thing this stock has done in 4 months that isn’t simply lower. Below it, the next reference is the December low at $844.06.

3. The level that matters is the published $929.93, on the 50-day average. It is about $12 away, or 0.9 times the $13.71 average daily range, against 2.4 times on Thursday.

4. Even the trade the plan allows isn’t generous yet. Buy the reclaim at $929.93 while treating $862.00 as the invalidation and you’re paying about 1 to 1 to the June high at $997.41. That ratio is part of why the instruction is wait rather than buy the trigger.

5. Bidding below the September low is a different trade from the one I’ve published. This morning it would have worked. It isn’t what the plan says, so I won’t dress it up.

Bottom Line

Costco delivered one of its better years, and Thursday’s release gives you very little evidence of an operating breakdown.

That is narrower than it sounds, because most of the 18% decline happened before Thursday, and results published this week cannot explain 4 months of price action that preceded them. The rotation out of expensive retailers that would explain it is too weak and too unstable to carry the weight. I’d rather say so than invent a cause.

What’s left is a valuation question you have to answer on its own terms. Earnings growth and a 4-month fall have taken a genuinely exceptional business to its cheapest multiple in nearly 3 years, and 44 times trailing earnings still leaves very little room for disappointment. The results settled the case against the business. They didn’t settle the case against the price, and until $929.93 goes, the weekly plan still says wait.

This is research and commentary, not personal investment advice. Levels are illustrative; size positions to your own risk tolerance and time horizon. The author may hold positions in names discussed.

Travel booking shares fell on the news and kept falling. Expedia is down 13.1% since the launch, Airbnb 17.8%, Booking 19.3%, TripAdvisor 13.8%. Meta is up 20.6%.

Then, on September 22, Expedia announced it was joining Muse.

Key Takeaways

• At launch the 2 halves of Muse’s travel product worked differently. Flights ran through a direct integration with Duffel, which Duffel announced itself. Hotels, on Skift’s September 9 test, were shopped on consumer websites the way a person would. Meta’s own announcement names no travel partner.

• Expedia has now joined, without detailing the mechanics. Reporting says Expedia stays the merchant of record while the traveller never leaves Muse.

• That is the real question: not who processes the booking, but who controls discovery and the relationship that follows it.

• Meta’s second-quarter free cash flow was $784 million. It repurchased no stock in the first half, after $22.9 billion a year earlier, and issued $24.9 billion of net new debt.

Prices are the Wednesday September 23 close, the last completed session. Meta’s Connect event runs September 23 and 24, so Thursday is both pre-open and day 2 of the event as this is written.

Discovery, Not Transactions

Flights were the clean case at launch. Duffel, the travel infrastructure company, announced on its own site that “millions of users can now use Duffel to search, book and manage holidays on Muse”, with US users booking “from all major suppliers globally, with live pricing”. No agency in between.

Accommodation was not that. Skift’s hands-on test on September 9 found Muse shopping consumer websites, Expedia among them, operating the way a person would rather than through any integration. Meta’s own announcement is silent on partners and describes only an agent that “opens browsers, fills forms” and “checks with the person before sensitive actions”.

Expedia’s announcement sharpens the question rather than settling it. The integration is planned, not live: no launch date has been given and Expedia hasn’t explained how the connection will work. On the reported terms Expedia would remain the merchant of record, keeping the payment, the service interaction and the account, while the traveller would plan, compare and pay without ever leaving Muse. Chief executive Ariane Gorin used the Skift Global Forum to argue for being everywhere agents live, while doubting that any single agent will own the trip. Expedia already connects to Google’s AI mode, ChatGPT, Claude and Alexa.

So the threat isn’t that Expedia stops getting bookings. It’s that Expedia risks losing control of discovery: the agent decides what to surface, in what order, and the traveller’s habit forms around Muse rather than around Expedia. Whether that costs Expedia the customer depends on what people do next, and nobody knows that yet. That is a slower and less total threat than agents bypassing the agencies outright, which is the concern the share price declines raise.

What the Market Moved

Since the session before the launch Meta has added $324 billion of market value. The Nasdaq 100 rose 3.1% over the same stretch and Meta’s beta to it was 0.80 over the prior year, which accounts for roughly $38 billion. That leaves a market-adjusted move of about $286 billion.

Treat that as an illustration of scale, and not as Muse’s implied value: it isn’t one. A residual like this also carries earnings expectations, rates, positioning and analyst revisions, and the clearest example is Meta’s biggest day: September 21, up 11.3% on 48.7 million shares, 2.5 times its prior 40-session average, on a Wells Fargo target increase to $796 from $640. Travel barely moved that day.

On the other side, Booking lost $29 billion over the period, Airbnb $19 billion, Expedia $5 billion and TripAdvisor about $0.2 billion, against $241 billion those 4 are still worth. Those declines are consistent with agent risk being repriced, but they’re 4 different businesses with different exposures and the period contains more than Muse. What both sets of numbers show is breadth: the Meta move is far too large to be explained by travel, and Muse has already announced tie-ups with Shopify, PayPal and Instacart.

The Advertising Question

There is a second-order effect worth watching without overstating.

Booking spent $8,186 million on marketing last year, 30.4% of its revenue, and calls those expenses “substantially variable in nature”. Its own description of performance marketing lists search engine keyword purchases first, then affiliate programs, meta-search referrals, and “other performance-based marketing, including social media marketing” last. Expedia spent about $7.4 billion on sales and marketing; across the 4 largest agencies the figure was roughly $20 billion.

If agents take over discovery, some of that spending could be displaced rather than redirected, and Meta collects a share of it. But social is one category among several in these budgets, most performance spend goes to search, and an agency that’s losing organic customers may bid harder for the ones it can still reach. This is a question, not an offset anyone can size. Alphabet, which takes considerably more travel search advertising than Meta, is flat over the same stretch.

What Is in the Accounts

Family of Apps, which is Facebook, Instagram, WhatsApp and Messenger, is 99% of Meta’s revenue. In the June quarter it grew revenue 28.0% to $60.4 billion, and its operating income fell 6.3% to $23.4 billion. The reported operating margin went from 53.0% to 38.8%.

Be careful with that. The quarter carried $2.40 billion of legal charges and $1.18 billion of severance from a May reduction of about 8,000 people. Meta doesn’t split those between segments and says the severance hit both. If all $3.6 billion sat in Family of Apps, the margin decline would be about 830 basis points rather than 1,421, and segment operating income would have risen roughly 8% instead of falling. The truth sits between the 2. Either way, employee compensation rose 56% and other costs 75% against revenue up 28%. Reality Labs lost $4.6 billion in the quarter on $431 million of revenue.

The cash statement is less ambiguous. In the June quarter operating cash flow was $31.86 billion and free cash flow, on Meta’s own definition, was $784 million, after capital spending of $31.08 billion including finance lease principal. On the narrower measure that counts only purchases of property and equipment, the quarter produced $1.7 billion against $9.0 billion a year earlier.

How that is being funded matters too. Meta repurchased no stock in the first half, against $22.9 billion in the same period of 2025, and issued $24.9 billion of net new long-term debt. Long-term debt went from $58.7 billion at the year end to $83.7 billion. A further $10.8 billion of money market funds is now escrowed against multi-year infrastructure purchase agreements and can’t be used generally until 2028 to 2030. Full-year capital spending is guided to $130 billion to $145 billion, the low end raised twice from an original $115 billion.

The Case Against Me

This is an excellent business and the quarter says so.

Very few companies this size grow revenue 28%, and the composition is healthy: ad impressions rose 14% and average price per ad 12%. Daily active people reached 3.60 billion. Advertising is 97.6% of revenue and compounding in every region, with Rest of World up 35%. Liquidity is substantial at $90.3 billion of cash and marketable securities, though against $83.7 billion of long-term debt and $28.7 billion of lease obligations that is not a net cash position. At $744.10 the shares trade on roughly 22 times annualised first-half earnings. Management still expects full-year operating income above 2025, and has said so since January.

Muse is not a giveaway either. Meta says it is “free for most of what people need, with subscription plans for people who want to do more”, and launch coverage describes tiers at $20 and $100 a month. There is a model. What’s unproven is paid conversion and whether agent commerce earns a durable take rate.

My reservation isn’t the company. It’s that the stock was down 6.6% for the year on September 4 and is up 12.7% now, so the year arrived in 13 sessions, priced on what Muse might become.

The Setup

Read more

The argument was strong, and it was right. If you don’t ignore the app and the brand strength, SoFi is basically a consumer lender that trades at a tech company multiple. But when borrowers stop paying, SoFi’s true nature started to appear. And that’s what took the stock from USD 32.73 down to it’s USD 14.88 low.

In July, that started to change. Charge-offs fell to 2.62% from 3.03%. Deposits grew USD 5.3B in 90 days. Originations hit a record USD 14.8B, membership crossed 15.8 million, and management raised the revenue guide for the year. Despite all that, the stock closed down 8.9%.

Almost 2 months, later it now trades at USD 17.16, around 2.5% above where it sat the day before those results. The bear case that drove a 50% drawdown lost its evidence, and the stock didn’t move.

Turns out bad loans weren’t the only problem.

SoFi is growing revenue at 42% a year and keeping less of it every quarter. Net margin has gone to 16.9%, then 15.2%, then 12.9% across 3 earnings releases. In July, management lifted the revenue guide but did not change the net profit guidance (~USD 0.6 per share). That one decision says a lot.

The argument isn’t about losses any more. It’s about conversion.

Subscribe now

Key Takeaways

• Q2 delivered record net revenue of USD 1.22B (+42.5%), net income of USD 156.6M (+61%), record originations of USD 14.8B (+69%), and membership of 15.8 million (+35%). That was an 11th consecutive profitable quarter.

• Credit improved. The personal loan annualized charge-off rate fell from the prior quarter, down 21 BP year over year, with 90-day delinquencies flat against last year.

• The problem moved to the income statement. Quarterly net margin dropped from 16.9% to 15.2% to 12.9% while revenue accelerated, and the adjusted EPS guide stayed the same.

• Galileo, the technology segment, fell 23% to USD 84.5M, a third straight decline. That business carries the premium multiple, but lending is carrying the company.

• Price sits mid-range in a 6-month base between USD 14.88 and 20.13, with no trend on either timeframe. Support runs 16.50 to 16.90, repair needs a daily close above 18.05, and 15.80 breaks it.

Subscribe now

What Actually Changed Since June

SoFi runs 3 segments. Lending covers personal, student, and home loans, and generated USD 724.8M of net revenue last quarter, up 63%. Financial Services holds checking, savings, brokerage, and the credit card, at USD 466.3M, up 29%. Technology Platform is Galileo and Technisys, the infrastructure that powers other companies’ banking apps, at USD 84.5M and falling.

The bank charter is what makes the model work. Deposits now stand at USD 45.5B against USD 60.9B of total assets, which funds the loan book far more cheaply than the warehouse facilities SoFi used before. Management put the gap at 156 basis points last quarter.

Since our June coverage here, 2 things are new.

The loan marketplace, where SoFi originates for third-party buyers and collects fees without keeping the credit risk, stabilized. Third-party personal loan volume came in at USD 3.1B against USD 3.0B the prior quarter, producing USD 140.9M of platform fees. Total fee-based revenue reached USD 472.3M, or 39% of the quarter, up 22%.

The other is the stablecoin build-out. On September 22, SoFi moved its entire USD 25B card program to settle on Mastercard’s network in SoFiUSD, its own USD-pegged token. Shares popped almost 6% intraday and but it didn’t stick by market closing time.

Share

Fundamental Analysis

The headline numbers are better than the share price implies.

• Net revenue: USD 1.22B in Q2, up 42.5% year over year. Trailing 12-month revenue USD 4.31B.

• Net income: USD 156.6M, up 61%. Trailing 12-month net income USD 636.3M.

• Adjusted EBITDA: USD 357.8M, up 44%, at a 30% margin.

• Originations: USD 14.8B, up 69%, a record. Personal loans USD 10.7B, student USD 2.7B, home USD 1.4B.

• Members: 15.8 million, up 35%, with a record 1.1 million added. Products 24.4 million, up 42%.

• Deposits: USD 45.5B, up USD 5.3B in the quarter.

• Tangible book value: USD 7.34 per share, up 56%.

• Net interest margin: 5.98%, up 4 basis points from the prior quarter.

Margin on the lending book widened last quarter rather than shrank. Average asset yields did fall 32 BP year over year, but cheaper deposit funding more than absorbed it.

So where did the money go?

Into cost and tax, not credit. The provision for credit losses was USD 13.8M in the quarter, up USD 3.7M from a year earlier, which is far too small to account for it.

Revenue grew USD 118M from the March quarter while net income fell USD 10M. Every dollar of incremental revenue arrived with equivalent incremental cost behind it.

Net margin has now compressed 3 quarters in a row, from 16.9% in the December quarter to 15.2% in March to 12.9% in June, and adjusted EPS has come in at USD 0.12 for 3 straight quarters too. The full-year guide assumes an effective tax rate near 22%.

Growth that doesn’t reach the bottom line is growth the market discounts. That’s the 8.9% selloff, and the April earnings report warranted a similar reaction.

On the valuation:

• P/E (TTM): 34.8. Forward: near 29x this year’s USD 0.60 consensus and about 21x next year’s USD 0.83.

• P/S (TTM): 5.1. P/B: 2.0, or 2.3x tangible book.

• Net margin (TTM): 14.8%. ROE (TTM): 6.2%.

• Debt/equity: 0.31, with USD 3.41B of debt against USD 3.13B of cash. Net debt is roughly USD 279M.

• Consensus: mean target USD 20.34, median USD 19.00, range USD 12 to USD 30, across 22 analysts, with an overall rating of hold.

The 6.2% return on equity is the weak spot. Robinhood earns 23.6% RoE and trades at 55x trailing earnings. Ally earns 9.7% and trades at 9x. earns 6.2% and trades at 35x.

A 42% growth rate validates some of that gap, but definitely not all of it.

Trailing free cash flow is negative USD 8.80B. At a deposit-funded lender, originations held for sale run through operating cash flow while the deposits funding them sit in financing. A lender growing originations at 69% will show deeply negative operating cash flow by construction. Judge this balance sheet on capital, funding cost, and credit, all of which improved.

Fundamental conclusion: The franchise is compounding faster than almost anything in consumer finance, credit is behaving, and funding keeps getting cheaper. What hasn’t arrived is operating leverage. Until net margin stops sliding or Galileo stops shrinking, 35x trailing earnings on a 6.2% ROE asks the buyer to underwrite a profit inflection that management itself declined to guide toward. The October print is the next chance to settle it.

Subscribe now

Technical Analysis

Weekly Chart

Step back far enough and 2026 is one long repair job.

The high was 32.73, and then broke down to almost half by late March. Since then, it’s been oscillating between 14.88 and 20.13. The bottom end has been tested in 5 separate months, from March through July. The ceiling has held strongly, and there hasn’t been a weekly close above 20 since the first week of Feb 2026.

• Trend strength is absent. ADX reads 10.9, and a trend-strength gauge below 20 says range-bound chop rather than directional trend. Breakout entries fail repeatedly in this regime.

• Price is threading the long averages. The 20-week EMA, roughly 5 months of price, sits at 17.81. The 50-week, about a year, is 18.71. Both sit just overhead and both have flattened. The 200-week, close to 4 years of price, is 15.78 and rising into the range from below.

• Momentum is improving from a deep hole. The weekly MACD line at minus 0.49 has climbed above its signal at minus 0.64, with the histogram positive at 0.15 and expanding, meaning downside momentum is fading even though the reading remains below zero.

• The ceiling is higher than the range top. The weekly Ichimoku cloud, a shaded band that acts as trend-following support and resistance, sits at 20.67 to 22.36. Price hasn’t traded inside it since February.

• RSI 45.5 is neutral, with no divergence either way. Weekly volume last week ran at about half its 20-week average, which is what apathy looks like.

A 6-month base holding above a rising 200-week average is good, but the weekly close pattern is not so good.

Daily Chart

The daily picture is tighter and currently tilted against buyers.

Price at 17.16 sits below all 4 daily moving averages, and they’ve bunched into a band barely 7% wide: the 20-day at 17.46, the 50-day at 17.62, the 100-day at 17.90, the 200-day at 18.72. Compression like that resolves eventually. It’s just hard to see which way it’ll resolve.

• The August rally has rolled over. From the July 29 low of 14.88, shares ran to 19.50 by August 28, then made lower highs at 18.71, 18.35, and 17.96. The September lows have held a shelf at 16.53 to 16.83.

• Momentum has turned down. The MACD line at minus 0.27 has slipped below its signal at minus 0.18, with the histogram negative at minus 0.09, so short-term momentum is softening rather than building.

• The trend gauge is flat on its back. ADX 10.7 confirms the weekly read. No trend to trade, so mean reversion inside the range has better odds than chasing either edge.

• Buyers haven’t actually left. Directional indicators sit at +DI 23.8 versus -DI 18.7, meaning buying pressure still holds a slight edge over selling despite the drift lower.

• Short-term momentum has reset. StochRSI at 26 on the fast line is near the oversold floor and curling up, which is where bounces inside a range usually start.

• Fibonacci confluence sits just below. Measured from the 14.88 low to the 20.13 high, the 0.618 retracement is 16.89 and the 0.786 is 16.00. The September base is sitting on the first of those.

Then there’s yesterday, September 22, which was quite an informative session.

SoFi announced that its entire USD 25B card program had gone live settling in stablecoin on Mastercard’s rails. That is a real catalyst. The stock opened at 17.93, hit 17.96, and closed at 17.16 on 98.9 million shares, roughly 2.3 times its 20-day average volume.

A genuine headline, heavy volume, and nearly the whole move sold by the close.

Technical conclusion: Both timeframes say pretty much the same thing. SoFi is moving in a range, and not trending either way really. The base held on 5 separaste occasions, so the floor is being defended. But the ceiling is also not caving in. The setup favors buying into weakness in the lower half of the range, and chasing breakouts might be beneficial under certain conditions where the price move above the major EMAs.

Subscribe now

Read more

First, it’s a marketplace generating USD 5.86 billion of quarterly profit. The second company is an AI buildout that burned through USD 9.97 billion of capital in 3 months and produced a loss in the division meant to sell the output.

The first is paying for the second, and the math stopped working this year.

BABA closed Monday at USD 115.75, 39% below where it was last October and 62% below its 2020 peak. Analysts have a mean target of USD 185.90 on it, 61% above here.

Subscribe now

Key Takeaways

• The buildout is bigger than the cash flow funding it. Capex of USD 9.97 billion (up 75% year over year) vs USD 3.38 billion of operating cash flow, with free cash flow of negative USD 6.58 billion.

• China e-commerce is still strong at 41% of revenue, but it fell 8%. Quick commerce and AI cloud both grew 45%. The headline 9% blends those, and the half that's shrinking is the biggest and highest-margin part of the company.

• The AI division loses money. AI Cloud earned USD 0.83 billion. AI Labs lost USD 2.04 billion, which is USD 4.15 of loss per dollar of its own revenue.

• Buyback was cut by 80% to USD 162 million from USD 815 million.

• Adjusted EBITA fell 30% and operating income fell 57% on revenue that grew 9%.

USD figures use 6.785 RMB/USD, the rate in the company’s own release.

runs China’s largest online marketplaces, its largest public cloud, and a set of AI models that compete with other global players. Gartner ranks it fourth worldwide in infrastructure-as-a-service and first in Asia Pacific. Omdia put its share of China’s AI cloud market at 38.1% for 2025, up from 35.8%.

There’s also an ongoing lawsuit to consider. A securities fraud class action is pending that alleges Alibaba misstated its status under the 2025 defence authorisation act as a company tied to China’s industry ministry.

Subscribe now

Fundamental Overview

Capital allocation

Alibaba spent USD 9.97 billion on PPE last quarter, while it generated USD 3.38 billion of cash from operations, about 1/3 less than it spent. CAPEX was also around 25.2% of revenue. 1 year ago, CAPEX was about USD 5.70 billion. Spending grew 75% while the cash paying for it grew 11%.

FCF also came out at negative USD 6.58 billion.

Something had to bridge that gap. The buyback did, falling to USD 162 million from USD 815 million (an 80% cut). Shareholders used to have the priority claim on this cash, but have since taken a back seat to AI spending.

Building the fourth largest cloud on earth is expensive, and those expenditures are starting to materialize. T-Head Semis had shipped over 560k AI chips by April, there’s around 100k Zhenwu accelerators that run on Alibaba’s own cloud, and external cloud revenue grew to 45%.

All this changes what it means to own . A year ago Alibaba was a cash-returning franchise with a cloud option attached. Today it’s a capital project with a side retail business.

Revenue segmentation

• China e-commerce: USD 16.34 billion, down 8%. Still 41.2% of the company and the single biggest line.

• China quick commerce: USD 7.85 billion, up 45%. Instant retail, a land grab with real prize money and negative unit economics nearly everywhere it’s been tried.

• AI Cloud: USD 7.14 billion, up 45%, with AI products inside it at USD 1.82 billion, about 26% of the segment.

• International: USD 4.09 billion, down 1%. Global wholesale: USD 2.05 billion, up 7%.

The e-commerce group grew revenue 4% and its adjusted EBITA fell 1%.

Growth that shrinks profit is a mix problem. The shrinking piece is the high-margin piece and the growing piece is instant retail, so nothing holding up the 19.3% margin is the part expanding.

What AI Labs costs

AI Labs and Applications made USD 0.49 billion of revenue and lost USD 2.04 billion at the adjusted EBITA level. A year earlier it lost USD 0.48 billion. Revenue grew 16% while loss grew more than 4x.

Alibaba reports AI in 2 pieces, and every report quotes the profitable one. AI Cloud made USD 0.83 billion. Put the 2 together and the AI division lost USD 1.21 billion in the quarter, with the applications loss running at ~2.5 times the cloud profit.

Qwen is part of that segment, and the models are good. Qwen3.8-Max runs to 2.4 trillion parameters and Omni-Flash takes text, image, audio and video across a million-token context. But that doesn’t make the AI business profitable.

Relative valuation

At USD 287.7 billion, what you’re paying depends entirely on which earnings you’re looking at.

• Trailing revenue of USD 154.0 billion gives 1.9x

• The June quarter’s non-GAAP earnings annualised gives 23.6x.

• Trailing GAAP earnings, 26.5x.

• The June quarter’s GAAP earnings annualised, 46.3x.

The flattering one is trailing earnings, and it’s the most quoted metric. It contains quarters this company no longer produces. Adjusted EBITA is down 30%, operating income down 57%.

Annualise the outflow against the market value and the free cash flow yield is negative 9.2%.

Fundamental conclusion: the spending is defensible and the balance sheet can carry it. Debt to equity 0.25, current ratio 1.36, and cloud adjusted EBITA up 133%, so operating leverage is arriving where Alibaba is spending. What’s changed is the claim order. The core marketplace is shrinking, the growth inside the profitable division is margin-dilutive, and the AI division loses money once you count both halves.

Share

Technical Overview

Weekly chart

The long picture is a broken uptrend that hasn’t repaired yet.

The week ended September 18 closed at USD 113.24, below all 3 weekly averages: the 20-week at USD 119.04, the 50-week at USD 125.63 and the 200-week at USD 120.36. It was the 5th consecutive week closing under the 20-week.

One thing still positive is that the 50-week is above the 200-week, so the multi-year structure hasn’t inverted.

Weekly RSI is 44.2. Weekly ADX, which measures how hard a trend is running, is 11.4. Under 15 means no trend at all on this timeframe, so this is more of a range than a downtrend.

The weekly Kijun, a midpoint of the last 26 weeks that often marks the floor of a trend, sits at USD 118.76.

Daily chart

Friday ran 4.33% to close at USD 113.24, reclaiming the 20-day. Monday added 2.22% to close at USD 115.75. Price now sits USD 2.36 above the 20-day of USD 113.39.

• RSI is 52.7, back above 50 for the first time since end of August.

• ADX is 20.9, and the directional lines have crossed. +DI at 32.8 now sits above -DI at 31.6. Buyers have the edge for the first time since August 21.

• MACD has crossed above its signal, -2.27 against -2.60. It’s still negative in absolute terms, so this means momentum is turning, but not good yet.

• ATR is USD 3.16, about 2.7% of price, modest for BABA, so a stop sits further away than the chart makes it look.

Also the 21st straight session below the 200-day at USD 125.90, so nothing structural is fixed.

As for the levels, the 61.8% retracement of the run from June’s low to the USD 145.53 swing high sits at USD 112.44. Alibaba closed below it for 7 sessions straight and bottomed at USD 106.64.

Monday’s high was USD 116.20, the gate, tagged to the cent. The close was 45 cents underneath at USD 115.75. Every rally since July has died below that line, and this one walked up to it and stopped.

Since the June low, 32 of 60 sessions have traded down to the 20-day at some point.

Technical conclusion: a repair in progress, and 2 sessions further into it than last week. Weekly ADX at 11.4 still says no trend on the higher timeframe and price sits below every weekly average. The daily has reclaimed 1 moving average out of 4, though RSI, ADX and MACD all turned at once. What makes it tradeable is the reclaim of the 61.8% retracement at USD 112.44. What stops it being a buy here is the USD 116.20 gate, which Monday touched and failed to close through, on less volume than an average day.

Subscribe now

The Trade Plan

Alibaba hasn’t set a date for the September quarter yet. It reported the same quarter on November 25 last year, so we can expect next earnings to be in around 2 months.

• Entry zone: USD 112.00 to USD 113.50. Filled Monday morning, then vacated.

  • Hovers around the 20-day and the 61.8% retracement. This is a level to wait for.

• Stop: a daily close below USD 106.50.

  • Just below the September 16 low of USD 106.64. From the middle of the zone that risks USD 6.25, or 1.98 ADR. This stock fell 52% between last October and June, so size for that.

• Targets:

  • T1 USD 118.70, where the daily Kijun at USD 118.63, the weekly Kijun and the 50% retracement all cluster.

  • T2 USD 125.50, where the 200-day at USD 125.99 meets the 38.2% retracement at USD 125.08.

  • T3 USD 145.53, the swing high.

The level to watch is USD 116.20, and Monday touched it. The high was USD 116.20 to the cent and the close was USD 115.75, so price came to the line and got turned away on below-average volume. A daily close through it turns this from a bounce into something you can hold. That close is the trigger.

So there are 2 honest ways in from here, and buying Tuesday’s open isn’t either of them. Wait for a daily close above USD 116.20 and take the breakout with the 50-day underneath you. Or wait for price to come back to the zone, which 32 of the last 60 sessions have done at some point.

What the fundamentals change here. With free cash flow negative there’s no cash yield underneath to cushion a disappointment, and no buyback bidding for the stock the way there was 2 years ago. Smaller size, and take T1 and T2 when they come.

Subscribe now

Conclusion

The cloud business is working. Alibaba has a defensible claim to being the AI infrastructure of the world’s second largest economy.

Right now the marketplace is paying for that, which shrank 8%, as well as shareholders who gave up 80% of their buyback. Free cash flow runs near negative USD 26 billion a year, and the applications arm is losing faster each quarter.

At 26.5 times trailing earnings none of that is priced as a disaster, and it shouldn’t be. It also isn’t priced as the largest capital program in the company’s history with nothing to show for it yet.

2 levels decide this, and Monday touched both. On the bottom side, 112.44, lost for 7 sessions this month and taken back on Friday. Monday’s gap fill retested it and held. On the upper side, 116.20, which has capped every rally since July. Monday’s high was exactly that, and it closed below it.

All content provided is for informational and educational purposes only and does not constitute investment advice, a recommendation, or an offer to buy or sell any security. The trade plans, levels, and scenarios discussed are illustrative frameworks based on market structure and are not guarantees of performance. Markets involve risk, and losses are possible. Past performance is not indicative of future results. Each reader is responsible for their own investment decisions, position sizing, and risk management.

As of 7:30am in New York on Friday, before the open, Oracle was indicated about 6% higher and Adobe about 4% lower.

Plenty went into that split. Oracle beat on earnings, on backlog and on a cash burn smaller than feared. Adobe carries a recurring-revenue deceleration, a freemium transition, an AI-competition overhang and a chief executive handover. What follows isn’t a claim about which moved the price, but an argument that the most durable difference between the 2 sits on the cash flow statement.

Key Takeaways

• Oracle spent $28.5 billion on capital projects against $19.3 billion of revenue, and free cash flow was negative $5.4 billion. Adobe spent $85 million and generated $2.4 billion.

• Of Oracle’s record $23.1 billion of operating cash flow, $15.4 billion was deferred revenue, and $11.4 billion of that is prepayment the filing describes as carrying a significant financing component.

• Oracle’s gross margin on cloud and software fell 936 basis points. Its operating margin rose 612, helped by cutting sales and research spending.

• Adobe guided fiscal 2026 ending recurring revenue growth to 10.2%, below the 11.2% it just reported. It fell 4% on a quarterly revenue guide 0.4% light.

Prices are the Thursday September 10 close, the last completed session, hours before both sets of results. Friday is pre-open as this is written.

The Same Week, Opposite Capital Models

Oracle spent more on capital projects during the quarter than it took in revenue. Revenue was $19.3 billion, capital spending $28.5 billion, 147% of revenue. Free cash flow was negative $5.4 billion in the quarter, after Oracle burned $23.7 billion across fiscal 2026.

Adobe was almost asset-light by comparison. Revenue was $6.76 billion, capital spending $85 million, 1.3% of revenue. Free cash flow was $2.4 billion and it returned $2.2 billion through buybacks, 92% of it.

The financing lines complete the picture. Oracle sold $19.9 billion of stock through an at-the-market program, repaid $4.2 billion of debt, and its diluted share count rose from 2,909 million to 3,000 million. Adobe repurchased about 9.5 million shares.

One company is issuing equity to fund a capital buildout. The other is returning almost all of its quarterly free cash flow through buybacks.

Where Oracle’s Record Cash Came From

Operating cash flow of $23.1 billion, up 184%, was the headline number. It deserves reading closely.

$11.4 billion of it sits on a line that didn’t exist a year ago: “increase in deferred revenues from customer prepayments with significant financing component”. That phrase is precise. Economically, part of the arrangement functions as customer financing: Oracle receives cash well before it delivers the capacity, and the standard requires it to account for the time value of that money.

Another $4.0 billion came from other deferred revenue. Together, $15.4 billion, or 66% of operating cash flow, came from increases in deferred revenue. Excluding that increase, operating cash flow was $7.7 billion, against $28.5 billion of spending.

This isn’t a criticism of the model. Being paid in advance by counterparties of that size is a real advantage. As of the June quarter Oracle put the prepaid and customer-supplied hardware portion of its large AI contracts at $75 billion, a figure it didn’t update this time. It guides to $90 billion to $95 billion of reported capital spending this year against net cash spending of no more than $70 billion, a gap of $20 billion to $25 billion it attributes to customer prepayments and financing arrangements.

But it does change what the cash flow statement means. A record operating cash number that is 66% a build in deferred revenue measures demand and payment terms as much as profit.

The Margin Nobody Mentioned

Oracle’s operating margin expanded from 28.7% to 34.8%, which sounds like scale arriving. Underneath, the opposite happened.

Cloud and software revenue rose 33% to $17.2 billion. The cost of delivering it rose 77% to $6.4 billion. Gross margin on cloud and software fell from 72.1% to 62.7%, a 936 basis point decline, consistent with a mix shift toward infrastructure as software revenue declines. Software revenue itself fell 3%.

So where did the expansion come from? Sales and marketing fell $252 million, research and development $90 million, amortization $218 million and restructuring $321 million: $881 million across 4 lines, 36% of the entire increase in operating income.

Research and development went from 16.7% of revenue to 12.4%, sales and marketing from 13.8% to 9.4%, in the year Oracle remade itself into an infrastructure company selling AI compute. That may be efficiency. It’s worth watching whether it’s durable.

The depreciation question sits behind it. Property and equipment went from $100.0 billion to $127.8 billion in 3 months, while depreciation, though it more than doubled from a year ago, was still $3.2 billion. Much of that new asset base hasn’t reached the income statement. Capital spending doesn’t hit operating profit the day it’s incurred; depreciation begins as assets are placed in service, then runs for years.

What Oracle Actually Delivered

The demand isn’t in doubt, and it’s worth stating plainly because the rest of this piece is skeptical.

Remaining performance obligations reached $664 billion against a revenue guide of at least $90 billion. Oracle booked more than $30 billion of new AI cloud contracts, delivered 850 megawatts of capacity and more than 300,000 GPUs, almost triple the previous quarter. Cloud infrastructure revenue grew 121% to $7.4 billion.

The sequential additions are lumpy rather than linear. Oracle’s releases put remaining obligations at $523 billion after November, $553 billion after February, $638 billion after May and $664 billion now, so the quarterly additions run $68 billion, $30 billion, $85 billion, $26 billion.

This quarter’s $26 billion is in line with February’s $30 billion, and May’s $85 billion was the outlier. It’s also a net figure: obligations fall as revenue is recognized against them, which is why Oracle can sign more than $30 billion of new contracts and still show a $26 billion increase. Worth watching, not yet a trend.

Adobe’s Problem Is Not This Quarter

Adobe’s quarter was fine. Revenue beat the $6.69 billion expected, non-GAAP earnings beat about $6.08, subscription revenue grew 14%, the business and consumer group 16%, and the full-year targets went up.

2 numbers best capture the concern.

The first is the fourth-quarter revenue guide of $6.80 billion to $6.85 billion, a midpoint 0.4% below consensus. A rounding error in most contexts, a 4% move here.

The second matters more. Annualized recurring revenue ended the quarter at $27.50 billion, up 11.2%, and the company guided full-year ending growth to 10.2%. In a subscription business the recurring base leads reported revenue, so reporting 13% growth while guiding that base to 10.2% suggests some of today’s growth is the echo of a faster-growing past. That gap, not the guide, is the argument.

Set against that, Adobe converts 97% of operating cash flow into free cash flow and trades at 10.2 times the fiscal 2026 earnings it has one quarter left to finish. Oracle trades at 18.9 times a fiscal 2027 guide with 3 quarters still to run, so the horizons aren’t the same and the gap is wider than it looks.

The market isn’t confused about which is cheaper. It’s judging which still has a future, and it’s judged the whole category. Measured from their own record closes, Adobe is 64% below its 2021 peak, Intuit 61% below its 2025 peak, ServiceNow 44% and Salesforce about a third. This is not a verdict on Adobe.

The Case Against Me

The bear case on Oracle assumes the capital cycle disappoints, and there’s no evidence of that yet. Contracts are signed faster than capacity can be built, customers are prepaying at a scale that materially reduces Oracle’s external funding requirement, and Oracle repaid debt rather than adding it. Net debt is $88.3 billion against $303 billion of assets. Equity funding at these prices is dilutive, but it’s also the conservative choice.

And the bull case on Adobe is that a business growing 13%, with a billion monthly users and 92% of free cash flow going to buybacks, is priced as though generative AI has already taken it, when the evidence is a 1 point deceleration in recurring revenue. Its non-GAAP operating margin was 43.9% and its GAAP operating margin 34.8%. Adobe’s 34.8% GAAP operating margin was the same as Oracle’s, even though Oracle spent roughly 117 times as much capital per dollar of revenue this quarter.

Both could ultimately work. But today’s valuations imply radically different judgments about where the economics of this capital cycle end up.

Bottom Line

Oracle closed at a record $328.33 on September 10, 2025, the session after it first revealed a $455 billion backlog. At Thursday’s close, exactly one year later, it was 53.4% below that peak. Over roughly the same period the backlog grew another 46%, to $664 billion, the capacity is arriving and revenue is growing 30%. The market has spent that year deciding the terms of the deal matter as much as the size.

The terms are now visible. Oracle is funding an infrastructure business with customer prepayments carrying a financing component and with new shares, while the margin on what it delivers compresses and the reported margin rises anyway, helped materially by lower sales, research, amortization and restructuring expense. Adobe is funding buybacks with its own cash while the recurring base that drives its revenue slows by a point a year.

One of those is a growth company taking financing risk. The other is a cash machine taking terminal risk. The market paid 19 times for the first and 10 times for the second, which is a coherent view. It is not an obviously correct one.

This is research and commentary, not personal investment advice. Levels are illustrative; size positions to your own risk tolerance and time horizon. The author may hold positions in names discussed.

Split the quarter into the 2 segments the company reports and it stops looking like one business having a good quarter.

Autonomous Systems, 72% of revenue, grew 21.3% to $346.0 million and earned $62.3 million of segment adjusted EBITDA, an 18.0% margin. Space, Cyber and Directed Energy, the other 28%, shrank 20.6% to $134.5 million and lost $8.9 million.

Consolidated revenue rose 5.7%, the average of a business compounding and a business contracting, describing neither.

Key Takeaways

• Segment adjusted EBITDA fell 5.6% even as revenue rose. Autonomous Systems added $9.5 million; the other gave back $12.7 million.

• Space, Cyber and Directed Energy has lost money in 3 of the 5 quarters since it arrived, and lost $2.6 million across fiscal 2026 on $618.8 million of revenue.

• Guidance implies GAAP earnings of $0.21 to $0.53 a share against adjusted earnings of $3.02 to $3.34. At $140.80 those are 381 times and 44 times.

• The auditor issued an adverse opinion on internal control over financial reporting at April 30, while giving the accounts a clean one.

Prices are the Wednesday September 9 close, the last completed session, hours before the results. Thursday is pre-open as this is written, with the stock indicated roughly 5% to 7% higher.

Which Segment Is Which

It matters what sits in each segment, and the answer isn’t tidy.

Autonomous Systems holds what AeroVironment built, Switchblade and Puma and JUMP 20, but also the air and missile defense, electronic warfare and undersea units that came with BlueHalo, plus ESAero, bought for $177.9 million in March. Its growth isn’t purely organic and the company doesn’t say how much isn’t. Within it, uncrewed aircraft did the work, up 71% to $120 million, while precision strike managed 8%.

Space, Cyber and Directed Energy is different: every dollar was acquired, and AeroVironment had essentially none of it before May 2025. Both halves fell, space and directed energy down 28% to $51 million, cyber and mission solutions down 16% to $83 million.

What That Segment Has Delivered

BlueHalo closed on May 1, 2025. Announced at $4.1 billion in November 2024, the stock fell before completion, so the consideration recorded was $3.53 billion, or $3.48 billion net of cash acquired: 17,425,849 shares at $151.52 plus settlement of BlueHalo’s debt. That purchase was split across both segments, so the price doesn’t attach to the shrinking one alone.

Since then it has posted adjusted EBITDA margins of 2.2%, then -3.8%, -1.3%, 1.0% and -6.6%, and lost $2.6 million across the full fiscal year on $618.8 million of revenue. In the April quarter, seasonally the strongest, it turned $149.2 million into $1.4 million while Autonomous Systems was converting at 28.2%.

Much of the decline is mechanical rather than collapsing demand. In January the government issued a stop-work order, later a termination for convenience, on BADGER antenna work under the Space Force’s Satellite Communication Augmentation Resource program, and that revenue stopped.

Then the accounting. AeroVironment wrote down goodwill in its Space unit and reported a third-quarter net loss of $156.6 million. On June 22 it filed an amended report restating that loss to $243.8 million: the carrying value used in the impairment test had omitted goodwill arising from acquired deferred tax assets and liabilities, understating the write-down by $89.4 million. The restated impairment was $240.7 million, and every dollar sits in this segment.

The amendment said why. The error “originated from a material weakness” in controls over the goodwill impairment analysis: the company “did not have a properly designed control requiring preparation and review of a reconciliation of goodwill by reporting unit.” A second one came in with the acquisition: BlueHalo “did not design and maintain effective information technology general controls” for systems feeding its financial reporting.

Deloitte’s report dated June 29 “expressed an adverse opinion on the Company’s internal control over financial reporting because of material weaknesses”, while giving the accounts an unqualified opinion. In the quarterly report filed September 10, management concluded disclosure controls “were not effective as of August 1, 2026”, remediation ongoing. Nobody says the numbers are wrong. They say the machinery producing them has already failed once.

And Then It Won the Laser

Here’s what stops this being a simple story.

On September 2 the Army awarded AeroVironment $464.8 million to produce LOCUST X3 laser systems under the Enduring High Energy Laser program. The Army calls it its first ever production contract for a high-energy laser weapon, and AeroVironment quotes engagement costs below $5 a shot. It belongs to the segment that just lost money, follows a $52 million first international LOCUST order booked in the quarter, and landed after the quarter ended, so the record $1.5 billion backlog excludes it.

So the bear case and the bull case are the same segment: terminated space work, a $240.7 million write-down and the control failures on one side, the largest award of its kind on the other.

The 66% Fall Wasn’t Just About BlueHalo

AeroVironment closed Wednesday 65.6% below its highest close of the past year, which is tempting to read as the market’s judgment on the acquisition.

The peer group makes that hard to sustain. Kratos fell 64.2% over the same period without making a comparable purchase, and Red Cat 53.2%, while Lockheed fell 22.5%, Northrop 32.9% and the defence sector fund only 13.3%. On daily closing returns since June 1, AeroVironment correlates 0.79 with Kratos and 0.42 with Lockheed.

That doesn’t prove the acquisition contributed nothing. It does show a large common factor moving the expensive end of defense, so buying this because it’s “66% off” is betting on a sector dislocation more than a company-specific one.

2 Sets of Books

At $140.80 the equity is worth $7.16 billion. Add $730.1 million of long-term debt and $120.8 million of leases, which I’m including in enterprise value, take off $580.2 million of cash and short-term investments, and you get $7.43 billion, or 3.4 times guided revenue and 23.6 times guided adjusted EBITDA.

The earnings multiple depends which earnings you accept. Guidance is net income of $10 million to $27 million, $0.21 to $0.53 a diluted share, against adjusted earnings of $3.02 to $3.34. At the midpoints that’s 381 times GAAP and 44 times adjusted.

The adjusted number strips out $0.69 a share of intangible amortization and purchase accounting, plus $0.04 of acquisition costs. It asks investors to look past recurring acquisition accounting while the acquired businesses inside Space, Cyber and Directed Energy are shrinking.

The balance sheet frames the same thing. Goodwill and intangibles are $3.38 billion of $5.73 billion of assets, 59% of everything the company owns, and tangible equity is $1.02 billion. After the impairment, retained earnings stand at $4.1 million against $4.40 billion of paid-in capital.

The Ramp Is Not Hidden

Guidance was maintained, and the shape it implies is steep, but management says so out loud: 45% of revenue to the first half and 55% to the second, adjusted EBITDA roughly 33% and 67%, adjusted earnings 30% and 70%.

So the arithmetic is the company’s own. With $480.5 million banked, revenue of $2.125 billion to $2.225 billion needs $548 million to $582 million each remaining quarter, and adjusted EBITDA $84 million to $91 million against the $53.4 million just delivered.

That shape is normal here: last year’s April quarter produced $140.1 million of adjusted EBITDA against $56.6 million in the July quarter. The question isn’t whether a ramp is required, because management says it is. It’s whether the segment running at negative margin can turn LOCUST volume into the second-half profitability the guidance assumes. One quiet detail: the $3.18 adjusted midpoint sits under the $3.22 the street carried into the print.

The Case Against Me

The drone franchise is genuinely good: 21% growth, 18% margins, uncrewed aircraft up 71%, bookings of $0.7 billion at a 1.4 book-to-bill.

The stronger argument is that the weak segment is absorbing a terminated program while directed energy ramps behind it. On that reading the margin trough is the setup rather than the warning, and the $464.8 million award is exactly what it predicts. Excluding leases, debt less cash and investments is about $150 million, so there’s time to find out. Going into the print, all 20 analysts covering the stock rated it hold or better, with a mean target of $225.77.

My reservation is that we’re 5 quarters past the close, the segment’s revenue is lower than when it arrived, its margin is the worst it has been, and the acquisition-related control weaknesses still aren’t fixed.

The Setup

Read more

Leading a group that moved together isn’t the same as being repriced. On daily close-to-close returns over the 3 months to September 8, IREN correlated 0.85 with Core Scientific, 0.83 with Applied Digital and 0.78 with both TeraWulf and Cipher Digital. When the basket moves, IREN moves.

Tuesday shows what that costs in information. IREN announced that its 2 gigawatt Sweetwater Hub had been conditionally classified as Base Load in the ERCOT process, a genuine milestone on its largest development. The stock rose 5.0% and finished mid-pack. TeraWulf rose 8.2%, Nebius 7.7% on a Palantir partnership, Applied Digital 7.4%.

Underneath sits the fact every write-up reached for. In the June quarter IREN’s AI cloud revenue passed its bitcoin mining revenue for the first time, $70.5 million against $66.7 million. The crossover is real. It’s also not unique, and why it isn’t is the story.

Key Takeaways

• AI cloud revenue was $70.5 million in the June quarter against $66.7 million of mining. Across the full year bitcoin was still 82% of revenue.

• TeraWulf and Core Scientific crossed earlier. What separates IREN is that it buys and runs the hardware while they lease space and power.

• Deferred revenue grew $1.84 billion, 88% of the $2.10 billion of reported operating cash flow. Of the $1,842.6 million closing balance, 97.5% is non-current.

• Contracted obligations were $16.6 billion at June 30, before July’s $2.8 billion and an August deal of undisclosed value.

Prices are the Tuesday September 8 close, the last completed session. Wednesday is pre-open as this is written.

The Crossover Is Real, and It Isn’t Rare

For the year to June, IREN did $707.0 million of revenue, up 41%. Bitcoin mining was $578.2 million and AI cloud $128.8 million, so the old business was still 82% of it. The June quarter inverts: $137.2 million total, AI cloud $70.5 million, bitcoin $66.7 million. The AI share went from 18.2% across the year to 51.4% in the final 3 months.

The annual rate hides the cost. Quarterly revenue peaked at $240.3 million in September 2025 and fell to $137.2 million by June, a 43% decline, because bitcoin collapsed faster than AI arrived. The run rate nearly halved while the full year grew 41%. IREN lost $702.6 million in fiscal 2026, $638.8 million of it non-cash impairment from decommissioning mining hardware.

Now the part that undercuts the easy version. TeraWulf got there first: HPC lease revenue was 62% of its total in March, then $31.9 million against $12.8 million of digital assets in June, 71%. Core Scientific is further along, with June colocation revenue of $136.7 million against $21.5 million of self-mining, roughly 84%. Cipher renamed itself Cipher Digital in February to make the same point.

IREN didn’t cross first and isn’t crossing alone. Anyone selling the crossover as the differentiator hasn’t read the peer filings.

What Actually Separates Them

The difference isn’t that these companies left bitcoin. It’s what they became.

TeraWulf, Core Scientific and Cipher Digital are landlords. They lease space, power and cooling to somebody else who brings the compute. TeraWulf’s Anthropic arrangement runs 20 years over 401 megawatts; Cipher’s capacity goes to AWS and to Fluidstack for Google. Landlord is shorthand: these companies still design, finance and operate complex data centres. The customer owns the compute risk.

IREN is doing something structurally different. It buys the NVIDIA hardware, operates it, and sells AI cloud services directly. Customer prepayments, the company says, cover 45% to 55% of the associated GPU capital cost.

That choice changes everything downstream. IREN carries the hardware, so it absorbs obsolescence risk a landlord never touches. Its portfolio averages roughly 4 years against 10 to 20-year base terms across much of the colocation cohort, so it re-prices sooner, either way. It needs more capital per megawatt. And if it works, it keeps the margin between renting a building and selling compute, which is the reason to take the risk.

That’s a real fork in the road, and the market isn’t pricing it. These 6 names moved within 11 percentage points of each other over those 4 sessions while running materially different businesses.

What the Contracts Say

Microsoft signed on November 3, 2025: 5 years, roughly $9.7 billion of contract value, a 20% prepayment, and an expected $1.94 billion of annualised run-rate revenue once commissioned. NVIDIA followed in May 2026 with about $3.4 billion. In July, Prometheus, Perplexity and Together AI added about $2.8 billion, and in August IREN signed a frontier lab whose value it hasn’t disclosed. Horizon 1, the first of 4 liquid-cooled 50 megawatt Microsoft deployments, arrived that month.

The company’s own measure beats adding up press releases. Unsatisfied performance obligations plus contracted lease value totalled $16.6 billion at June 30, before the July and August additions.

Hold that against the accounts, and mind the dates. The AI business billed $70.5 million in the June quarter, annualising to $282 million of recognised revenue. Operating run-rate was $1 billion on August 26. Contracted run-rate is $4 billion on capacity targeted to be operational by December 31. Those are 3 clocks, and IREN warns run-rate can materially exceed recognised revenue. The distance between $282 million and $4 billion is where the equity value lives.

Management intends bitcoin mining to be effectively decommissioned by the end of December. This is not a hedged pivot.

The Same $20.2 Billion, Divided 4 Ways

At $46.93 the equity is worth $18.5 billion. Add $7.59 billion of borrowings, take off $5.90 billion of cash, and enterprise value is $20.2 billion.

Against the $707 million reported for the year, that’s 28.6 times. Against the AI quarter annualised, 71.6 times. Against the $1 billion operating run-rate, 20.2 times. Against the $4 billion contracted, 5.0 times.

All 4 are defensible and they differ by a factor of 14. Anyone quoting a single multiple has chosen a conclusion first. What the range tells you is that the valuation already assigns substantial value to capacity that hasn’t yet produced reported revenue.

Where the Cash Comes From

Operating cash flow was $2.10 billion. On $707 million of revenue, that should stop you.

It isn’t earnings. Deferred revenue grew $1.84 billion during the year, 88% of the reported figure. That’s customer money, the Microsoft prepayment and its equivalents, sitting as a liability until capacity is delivered. Strip it out as an analytical adjustment and the underlying figure is about $259 million. The closing balance was $1,842.6 million, 97.5% of it non-current, so IREN doesn’t expect to recognise it within 12 months.

Against $4.33 billion of capital spending, free cash flow, on my calculation rather than a reported line, was negative $2.23 billion.

The gap closed in the capital markets. Inflows included $4.74 billion from ordinary shares, $6.30 billion from convertible notes and $938 million from financing facilities. After $1.62 billion of induced conversion payments, capped calls, fees and repayments, net financing cash flow was $9.68 billion. Cash went from $564.5 million to $5.90 billion against $7.59 billion of borrowings, leaving net debt near $1.7 billion, the one comfortable number.

The stack is more layered than it looks. Customers fund a large piece up front, asset-backed lenders finance much of the GPU balance, and the capital markets fund the infrastructure around it. IREN says its Microsoft GPU financing plus prepayments covers 96% of the associated GPU cost, and a newer $2.8 billion of GPU financings covers 90% before any prepayment. Revenue arrives after the capacity does.

The Pattern Worth Noticing

IREN’s intraday high was $76.87 on November 5, 2025, 2 days after the Microsoft contract. The stock is 38.9% below.

Oracle traced the same shape, peaking the session after it revealed the backlog that made it an AI infrastructure company, and sits about 53% below. The announcement was the top; the construction has been the drawdown. If that holds, the re-rating comes not from another contract but from the AI line climbing from $70.5 million a quarter toward the run-rate the company publishes.

The Case Against Me

The bull case deserves stating properly. 2026 capacity is largely sold out, the counterparties are Microsoft and NVIDIA rather than start-ups, Horizon 1 is delivered, and mining is being switched off rather than milked. Roughly 26% of the shares were short in mid-July, so delivery into that can move the price a long way.

The ERCOT ruling is more load-bearing than its reaction suggested, though it’s provisional and a later commitment gate still has to be cleared. If power is the constraint, a ruling on 2 gigawatts beats another contract IREN can’t yet serve.

The model argument cuts both ways too. Owning the hardware is what makes IREN more than a landlord, and it’s also what could ruin it if utilisation or GPU pricing disappoints.

The Setup

Read more

On Thursday evening, after the close, Oracle and Adobe report.

Oracle set its 52-week high on September 10, 2025, the session after a first quarter that showed the market a backlog since grown to $638 billion. The stock rose about 36% that day. One year on, to the day, it reports again 54% below that high, though it has rallied 12% in the 3 sessions just gone.

Adobe, whose problem is not demand but whether it stays the layer where creative work gets monetised, is down 28% from its high. It fell 6.7% on Friday alone, after naming a new chief executive.

That ranking is the thesis. The company that proved AI demand exists has been punished roughly twice as hard as the one whose product AI is supposed to be replacing. The market would rather own contested relevance that generates cash than uncontested demand that consumes it.

Key Takeaways

• Oracle’s backlog is $638 billion, up 363% and 7.1 times guided fiscal 2027 revenue, though only about 12% converts inside 12 months. Some $75 billion is customer prepayment or hardware the customer supplied.

• Serving it costs about $70 billion of net cash capital outlay in fiscal 2027, 78% of that year’s guided revenue, with gross spending $20 billion to $25 billion higher. Fiscal 2026 free cash flow was negative $23.7 billion; net borrowings are $97.6 billion.

• Adobe’s AI-first recurring revenue passed $500 million and tripled. A freemium and pricing trade-off sized on its last call at about the same $500 million runs the other way.

• Adobe converts 40.8% of revenue into free cash flow and trades near 10.3 times it. It reports with a lame-duck chief executive, an interim finance chief and its creative head leaving.

Prices are the Friday September 4 close, the last completed session. Both report after the close on Thursday September 10. United States markets are shut Monday for Labor Day.

The Same Question, Opposite Answers

Every large software company is valued on 1 question: does AI feed it or eat it?

Oracle answered feed, turning from a database company that also sold cloud into a landlord for AI compute. Adobe is the test case for eat: it sells creative software by the seat to people whose work generative models can increasingly do, and has spent 2 years arguing it can sell the models instead.

Both answers are defensible. The risks they create are not symmetrical.

Oracle: The Backlog Is Real, and So Is the Bill

Remaining performance obligations ended May at $638 billion, up 363% and up $85 billion in a quarter. That RPO equals 7.1 times fiscal 2027 guided revenue of $90 billion, and it converts slowly: Oracle expects only about 12% recognised over the following 12 months, roughly $77 billion.

Oracle told you how it was built. Most of the last 2 quarters’ increase came from large AI contracts where the customer either prepaid for the graphics processors or supplied them. Those portions total $75 billion, close to 12% of the backlog.

That matters, but not for the reason it first appears. Management says those structures carry no degradation in margin, so this isn’t about bad contract economics. What the $75 billion does is cut Oracle’s own funding requirement: the customer has already paid for, or supplied, the most expensive component. Gross margins are still guided down through fiscal 2027, but Oracle attributes that to data-centre ramp timing and mix, not these contracts.

The business underneath is strong. Cloud infrastructure revenue grew 93% to $5.8 billion and total cloud revenue rose 47%. The Multicloud AI Database, which puts Oracle’s database inside other companies’ clouds, grew 404%. That matters most strategically, because it’s what rivals can’t replicate: the installed base of mission-critical databases is Oracle’s, and letting customers run it anywhere defends it. The applications half is weak at 10%, a drag carried by infrastructure at a company guiding 18% earnings growth.

What Oracle Has to Fund

Capital spending goes from $21.2 billion in fiscal 2025 to $55.7 billion in fiscal 2026, and fiscal 2027 is guided to about $70 billion of net cash outlay. Read that word: gross reported capital spending could run $20 billion to $25 billion higher, with customer prepayments and timing absorbing the difference. Even on the net figure, Oracle lays out 78 cents for every dollar of the $90 billion it expects to sell.

It funds that externally. Fiscal 2026 operating cash flow was $32.0 billion against $55.7 billion of spending, so free cash flow was negative $23.7 billion. Borrowings ended the year at $129.5 billion against $31.9 billion of cash and securities, so net borrowings of $97.6 billion, with leases on top. It raised $43 billion of debt and $5 billion of equity last year and expects roughly $40 billion more this year, including a $20 billion at-the-market equity programme, and says it expects no further debt issuance in calendar 2026.

None of it is hidden or fatal. Oracle still guides to $8.05 of non-GAAP earnings, under 20 times at $158.78. But the character of the asset has changed. Buying Oracle used to mean buying a toll on enterprise data. It now means underwriting a construction programme whose payback depends on customers honouring multi-year commitments. The reported $300 billion OpenAI contract alone is equivalent to roughly 47% of today’s entire backlog, which illustrates the scale of Oracle’s exposure to a counterparty still expected to consume enormous amounts of cash.

Be fair about the other side. Oracle says global utilisation of its graphics processors runs at 97.5%, and its finance chief puts steady-state return on capital in infrastructure in the high 20s. The bear case isn’t that Oracle signed bad contracts. On management’s own numbers it didn’t. It’s that an enormous amount of capital must arrive, and be spent, before those economics can prove themselves.

Adobe: The Cash Flow Is Real, the Relevance Is Disputed

Adobe’s problem is the opposite shape. Nobody doubts it can collect money. They doubt what for.

The AI progress is real. AI-first recurring revenue passed $500 million last quarter and tripled, and Firefly is approaching $300 million. More than 80% of its experience and content management customers use the agentic features built into those products, and it has pushed its agents into ChatGPT and Claude.

Then the base. Total recurring revenue exiting the May quarter was $27.10 billion, so the AI-first line is 1.8% of it. Tripling from a number that small is a leading indicator, not yet a financial event.

Now the number that reframes it. Adobe’s full-year target of 10.2% recurring revenue growth carries roughly $480 million of acquired book from Semrush, and absorbs a deliberate trade-off sized on the second-quarter call at about $500 million. Adobe didn’t guide to that figure but didn’t dispute the arithmetic, and Shantanu Narayen put roughly half on deferring creative pricing and half on widening the free funnel across Acrobat, Express and Firefly. So the AI revenue Adobe is proudly adding is close to the recurring revenue it’s choosing to give up. One is the future, the other the price of buying users to sell it to.

There is a pipeline underneath: remaining performance obligations of $22.27 billion with 67% due inside 12 months, roughly $14.9 billion contracted for the coming year.

What Adobe has that Oracle doesn’t is cash. Over 4 quarters it produced $10.28 billion of free cash flow on $25.2 billion of revenue, a 40.8% margin, against net debt of $1.4 billion. At $105.9 billion of market value that’s 10.3 times free cash flow, a price the market applies only when it doubts durability rather than profitability.

The Problem Adobe Has Added

It reports Thursday having changed almost everyone who answers for it.

On September 3 the board named Anil Chakravarthy, who runs the enterprise experience business, chief executive from December 1. Shantanu Narayen becomes executive chair after 18 years. David Wadhwani, who ran creativity and productivity for nearly 5 years and was the other internal candidate, is leaving. The finance seat has been interim since June, when Dan Durn left, as it happens for Marvell. The shares fell 6.7% on Friday on nearly twice normal volume.

Read the choice, not the churn. Faced with a creative franchise under AI pressure and a faster-growing enterprise business, Adobe promoted the enterprise leader and lost the creative one. That’s coherent and arguably right. It’s also an admission about where the board thinks the growth is, a week before a print already about that question.

Potential Directions: Oracle

The print is not about the quarter. Consensus revenue is near $19.13 billion and Oracle will probably land close to it.

What decides the reaction. The backlog first: anything below $638 billion would be the first sequential decline of the AI era and would become the story. Second, the prepaid share, which sets how much of the build Oracle funds itself. Third, spending, where $70 billion net is the number to beat either way. Fourth, any change to the $90 billion and $8.05 guide.

The bull path. Backlog grows again, cloud infrastructure holds near 90%, spending is reaffirmed rather than raised, and management details financing costs. That makes Oracle a growth story with a heavy but understood balance sheet, and a stock under 20 times guided earnings has room.

The base path. Backlog roughly flat, revenue in line, spending reiterated, margins stepping down as flagged. Nothing resolves and the stock keeps trading on the financing question.

The bear path. Backlog falls, spending rises again, free cash flow deteriorates faster than expected, or a large contract is restructured. Any of those turns the discussion to the $97.6 billion of net borrowings, and the multiple stops being the constraint.

Potential Directions: Adobe

What decides the reaction. Recurring revenue, and whether the roughly $500 million trade-off implied on the second-quarter call is developing as expected. Then the AI-first line, which needs to compound fast enough to matter before the base decelerates. Then whether the deferred pricing increases stay deferred, and the full-year target of $26.50 billion to $26.60 billion and 10.2%.

The bull path. Recurring revenue beats, the trade-off tracks, AI-first revenue keeps roughly tripling, and the incoming chief executive signals pricing returns next year. At 10.3 times free cash flow, a company merely proving it isn’t shrinking would be worth considerably more.

The base path. In-line quarter, targets maintained, AI metrics good and still small. The stock stays cheap, because cheap is a judgment about the next decade, not this quarter.

The bear path. Recurring revenue misses, the trade-off costs more than $500 million, or the pricing deferral extends again. Then the argument becomes whether the seat count itself is at risk, the only version that justifies the current multiple, and a transition management is poorly placed to answer it.

Bottom Line

Oracle has the demand and must buy the capacity to serve it. Adobe has the cash and must prove anyone still wants what it sells.

I’d watch 1 number at each. At Oracle it’s the backlog, because everything else is downstream of whether $638 billion is still growing. At Adobe it’s recurring revenue against the roughly $500 million that trade-off was sized at, the one line where strategy and results meet.

2 prints, 1 evening, and opposite ways to be wrong.

This is research and commentary, not personal investment advice. The author may hold positions in names discussed.

The stock fell 2.7% the next day, while the S&P 500 rose 1.1% and Nvidia rose 1.8%.

A week earlier Marvell had beaten, raised its outlook for 2 fiscal years, and fallen 10.3%. Broadcom now sits 27.7% below its June high and Marvell 32.2% below its own.

The growth is not in dispute. What’s being priced is what each company gave up to get it.

Key Takeaways

• Broadcom’s semiconductor segment grew 127% and its software segment 29%. Software fell from 42.5% of revenue to 29.6% in a year, and gross margin fell 210 basis points sequentially to 75% on that mix.

• Operating margin went the other way, rising about 240 basis points to 67.9% as scale absorbed the dilution. Operating income rose 92% to a record $20.1 billion.

• On August 19 Marvell granted Google a warrant over 58.97 million shares, 6.7% of those outstanding. 97.7% of it is performance-based, and full performance vesting requires $120 billion of cumulative qualifying Google revenue, 10 times Marvell’s guided year.

• Marvell’s non-GAAP operating margin is 36.6%. Broadcom’s is 67.9%. On forward guided revenue both trade near 16.4 times.

Prices are the Friday September 4 close, the last completed session.

Both Beat. Both Fell.

Marvell went first, after the close on August 27. Revenue of $2.739 billion, up 37%, a company record. Data centre revenue of $2.171 billion, 79% of the total, up 46%. Non-GAAP earnings of $0.94 a share against $0.33 on a GAAP basis. It raised fiscal 2027 revenue to $12 billion from $11.5 billion and fiscal 2028 to $18 billion, and guided the October quarter to $3.15 billion, 52% growth against 37% just delivered.

The stock fell 10.3% the next session. Be careful attributing all of that to the print: the sector fund fell 3.5% and Nvidia 4.6% the same day. The sector move alone was roughly a third the size of Marvell’s decline.

Broadcom’s reaction is the cleaner signal. It fell 2.7% on September 3 while the market rose, close to 4 points of underperformance with no sector excuse available.

Then on Friday Marvell rose 7.1% to $223.55. Mostly a sector day too, with the semiconductor index up 3.5% and AMD up 4.7%, on 21.2 million shares against a 20-session average of 24.4 million. Lighter buying than the selling it followed.

What Broadcom Is Paying

Broadcom has 2 businesses. Semiconductor solutions did $20.839 billion against $9.166 billion a year ago, up 127%. Infrastructure software, largely VMware, did $8.752 billion against $6.786 billion, up 29%.

So software has gone from 42.5% of revenue to 29.6% in 4 quarters. Not because it shrank: it grew 29%, and is being diluted by a semiconductor business growing 4 times faster.

The 2 halves don’t earn the same way, and you see it in the gross line. Consolidated gross margin came in at 75%, down 210 basis points sequentially, which the company attributes to AI mix. AI semiconductor revenue carries a lower gross margin than software on average, so as accelerators take a larger share of the mix they pull the consolidated figure down.

Now the part that complicates the bear case: operating margin didn’t follow it down. It rose about 240 basis points year over year to 67.9%, and operating income rose 92% to a record $20.1 billion. Scale is absorbing the dilution faster than the mix creates it. The fourth quarter is guided to about 66%, on revenue up another 18%.

So the honest statement isn’t that margins are falling. It’s that Broadcom is accepting a structurally lower-margin mix in exchange for volume, and operating leverage is currently masking the cost. An investor who paid a software multiple is watching the software proportion fall by a third in a year. That’s a change in the thing being valued while you hold it, whether or not the operating margin shows it.

What Marvell Paid

Marvell’s cost is explicit, and it isn’t in the income statement.

On August 19 Marvell disclosed an expanded custom-silicon relationship with Google, covering inference accelerators, storage and network interface controllers, memory interface controllers and near-memory compute attached to Google’s tensor processing unit ecosystem.

The price was a warrant: Google can buy up to 58.97 million Marvell shares at $206.58, exercisable to August 2033. At that strike, exercising all 58.97 million shares for cash would imply an aggregate exercise price near $12.2 billion, though the warrant also permits cashless exercise. Against 876.9 million shares outstanding the cover is 6.7%. Marvell closed Friday at $223.55, so the strike is already 8% in the money.

Look at the vesting. A total of 1,360,867 shares vest on time over the first year, in 4 quarterly instalments of roughly 340,000. The other 57.61 million, which is 97.7% of the warrant, are performance-based: they vest in 240 tranches, 1 for every $500 million of custom-product revenue Marvell records from Google, to the end of fiscal 2033.

Full performance vesting therefore requires $120 billion of cumulative Google revenue. Marvell’s entire guided fiscal 2027 revenue is $12 billion. The warrant is calibrated to 10 times a full year of everything Marvell sells.

That tells you how large Marvell thinks this can be, and about bargaining power: it put real equity behind the relationship to secure Google’s long-term custom-silicon business. On the day it was disclosed Marvell rose 9.9% and Broadcom fell 4.6%, because Google had broadened its supplier base beyond Broadcom, which holds long-term agreements to develop Google’s future accelerators and supply components for its AI racks as late as 2031.

The Gap Between Them

These get discussed as rivals. The numbers side by side make that struggle.

Broadcom’s AI revenue in the fiscal third quarter alone was $16.7 billion, 1.4 times Marvell’s guided revenue for a year. Its guided fiscal fourth quarter, about $34.8 billion, is nearly 3 times Marvell’s guided year. Its fiscal 2028 line of sight is $230 billion against Marvell’s $18 billion of total revenue.

The profitability gap is wider. Marvell’s non-GAAP operating margin was 36.6%, and 16.8% on a GAAP basis. Broadcom’s was 67.9%.

What “Line of Sight” Carries

Broadcom’s $115 billion and $230 billion rest on what it called secured supply and conservative deployment assumptions. That’s real demand backed by secured supply, and I’m not waving it away.

But look at whose spending they describe. Broadcom named 6 custom accelerator customers, and Hock Tan said Anthropic is on track to become the largest in 2027 and stay there into 2028. Google and OpenAI are others.

Google funds capital spending from one of the largest cash-generating businesses in existence. Anthropic and OpenAI fund themselves by raising capital. Their demand is real; whether it converts on schedule depends on rounds not yet raised.

The market has built machinery for that gap, and Broadcom is inside it. In June, Apollo led a $35 billion financing with Blackstone and a group of banks to fund more than 1 gigawatt of Anthropic compute, under an arrangement meant to enable over 20 gigawatts for frontier AI labs through 2028. That cuts both ways: evidence the demand is bankable, and evidence that external capital is an integral part of the funding model. That isn’t a prediction of failure. It’s an observation that the visibility rests on someone else’s capital plan.

The Case Against Me

The strongest counter is valuation.

Broadcom’s enterprise value is about $1,738 billion against roughly $105.9 billion of fiscal 2026 revenue, the 3 reported quarters plus the guided fourth, or 16.4 times. Marvell’s is about $197 billion against $12 billion guided, which is also 16.4 times. The market pays the same multiple of next year’s revenue for a company earning 36.6% operating margins as for one earning 67.9% at 9 times the scale.

On revenue multiples alone Broadcom’s profitability makes the identical pricing hard to ignore, though Marvell’s faster growth complicates the comparison and both could be wrong in the same direction. The sell side leans the same way: Broadcom’s average target is $533 against $357.90, with 3 holds and no sells among 49 analysts; Marvell’s is $285 against $223.55.

The mix argument also has a ceiling. If Broadcom does $230 billion of AI revenue in 2028, the software percentage will be a footnote nobody remembers.

The Setup

Broadcom is in the weekly plan I publish and Marvell isn’t, so I’ll say which is which.

1. This weekend’s plan cut Broadcom to Breakdown and its status to avoid. The reclaim is now above $373.55, invalidation $342.33, first objective $380.62. Note where those come from: $373.55 is the 20-day exponential average and $342.33 is Thursday’s intraday low. At $357.90 the stock sits between them, below its 20, 50 and 200-day averages of $373.55, $382.84 and $368.11.

2. Check the geometry first. Buying $373.55 with invalidation at $342.33 risks $31.22 to make $7.07, about 0.23 to 1. The status is avoid and the geometry is most of the reason.

3. Marvell’s levels are mine. It gapped from $241.45 to open at $225.26 on August 28 and hasn’t closed it. Friday’s $223.55 sits just under the lower edge.

4. The zone is $212 to $224, bracketing both exponential averages and Friday’s close, so the stock is already inside it. Invalidation is a close below $199, under the September 1 low of $200.62. From the middle that risks about $19, or 1.2 average daily ranges, and pays 1.2 to 1 at the $241.45 gap fill and 1.8 to 1 at the upper band.

5. Buying strength above $225.26 instead risks $26.26 to make $27.02, about 1 to 1. The pullback pays better than the chase, though not by much. Marvell’s October 6 investor day is meant to quantify the Google opportunity.

Bottom Line

Both delivered. Broadcom’s was one of its strongest quarters on record; Marvell raised its outlook for 2 years in one release.

Both stocks are about 30% below their highs, and the market appears increasingly focused not on whether the AI revenue arrives but on what winning it costs. Broadcom is paying in mix: the high-margin half is under 30% of revenue, and it shows at the gross line while scale protects the operating line. Marvell paid in equity, handing a customer an option over 6.7% of itself, calibrated to 10 times its current year’s revenue.

Neither is a bad business. But growth bought at a price is not the same asset as growth that arrives free, and the market has started charging for the difference.

This is research and commentary, not personal investment advice. Levels are illustrative; size positions to your own risk tolerance and time horizon. The author may hold positions in names discussed.

On Thursday it opened at $377.24, up 23.4%. It touched $384.56, a 52-week high. It closed at $356.47.

That’s still up 16.6%, and it’s a very good day. It’s also sellers taking back 29% of the gap inside a single session, on 5 times its 20-session average volume. Both halves of that are the story: the mechanism worked, and the market immediately started arguing about the price.

Key Takeaways

• Product revenue grew 37%, after 30% and 34%. Net revenue retention was 126%, and remaining performance obligations reached $9.00 billion, up 30%.

• The full-year product revenue guide went to $6.07 billion from $5.84 billion in May. That $230 million raise is about 3 times the $74.4 million by which product revenue beat the company’s own guidance.

• The chief executive attributes about half the acceleration to AI products. CoCo, its coding agent, ended the quarter above 9,100 accounts, up more than 2,000.

• Non-GAAP operating margin was 15.3%. The GAAP figure was negative 17.0%. Stock compensation was $456.4 million, 29.5% of revenue.

Prices are the Thursday September 3 close, the last completed session. Friday is pre-open as this is written.

The Test Software Has Been Failing

The bear case on application software is simple and mostly right. If you sell licences by the seat, and AI does what the person in the seat used to do, then your customer needs fewer seats. Better product, smaller bill. Priced per user, capability becomes deflation.

Snowflake isn’t primarily priced per user. Its model is consumption-based, with customers paying for compute, storage, data transfer and AI usage, and AI work is enormously compute-hungry, and it runs on the data customers have already put there. Every model that gets pointed at a company’s own records is billable. In that arrangement AI isn’t a substitute for the product. It’s a meter running faster.

That’s the theory. It’s been the standing bull argument for consumption-priced data infrastructure for 2 years, and mostly just an argument, because the numbers weren’t separating from the rest of software. This quarter they did, which is why the reaction was violent.

The Acceleration

Growth in product revenue by quarter went 26%, 32%, 29%, 30%, 34%, 37%. Read it properly: it bottomed at 26%, wobbled back to 29%, and has now got faster 3 quarters running. The next quarter is guided to 37% to 38%, so management is telling you it expects the acceleration to hold.

Deceleration is close to a law in software. Companies get bigger, the base gets harder, growth grinds down. Snowflake did that for years, from triple-digit growth at listing down to 26%. Reversing it at a $6 billion run rate is unusual, and it isn’t a single-quarter artifact.

The supporting measures agree. Net revenue retention was 126%, up from 125% 3 quarters earlier, which means the multi-year slide in how much existing customers expand has, at least for now, stopped. Remaining performance obligations, the contracted work not yet recognised, reached $9.00 billion, up 30%. That’s 7 points slower than revenue, and it’s the first number a sceptic should reach for. The company’s explanation is that renewals cluster in the fourth quarter, and the detail supports it: about 54% of that $9.00 billion is expected to convert inside 12 months, and that portion is up roughly 42% on the same estimate a year ago. The near-term book is growing faster than revenue even though the total isn’t. Customers spending more than $1 million a year numbered 828, up 27%.

On where it came from, chief executive Sridhar Ramaswamy put roughly half on AI products, naming CoCo, CoWork, AI functions and the AI Gateway, and the other half on the core business, naming broader product use, applications, notebooks and faster customer migrations. CoCo finished above 9,100 accounts and CoWork at 5,800.

That split is worth holding onto, because it says the AI contribution is real but not the whole thing. My own reading is that the line between the halves is blurrier than the decomposition suggests: migrations mean companies moving data in so they have somewhere clean to point a model, which is an AI story at 1 remove.

The Raise Was Bigger Than the Beat

The number I’d keep if I could keep only one.

Compare like with like, using the company’s own numbers on both sides of it. Snowflake guided this quarter to $1.415 billion to $1.420 billion of product revenue and delivered $1,491.9 million, beating its own midpoint by $74.4 million. It then raised the full-year product revenue guide by $230 million, to $6.07 billion from the $5.84 billion given in May, and took the full-year margin guide to 14.5% from 13.5%.

Management raised the year by about 3 times what it beat the quarter by. Put it the other way round. Flow the entire $74.4 million beat straight into the full-year number, and roughly $155.6 million of the increase is still unaccounted for. That residual is management raising its assumptions for quarters that haven’t happened. Companies expecting a one-off don’t do that, because walking it back in December is brutal.

What the 15% Margin Leaves Out

Now the part that didn’t make the headlines.

The 15.3% operating margin everyone quoted is non-GAAP. The GAAP operating margin in the same quarter was negative 17.0%. Those are 32.3 points apart, and the gap is mostly 1 item: $456.4 million of charges related to stock-based compensation, including employer payroll taxes, which is 29.5% of revenue.

I’m not calling that a scandal. Excluding it is the standard convention in software and it’s disclosed in plain sight. But it means “accelerating and profitable” is a sentence that only holds in the adjusted column. On a GAAP basis Snowflake lost $191.7 million in the quarter, or $0.55 a share, which is at least a real improvement from $0.89 a year ago.

One more piece of arithmetic before you take the cash story on trust. Adjusted free cash flow was $92.3 million, a 6.0% margin, against a full-year guide of 23.0%. Collection is weighted to the back half and this is a normal shape for Snowflake, but the full-year figure depends on quarters that haven’t happened.

Why the Market Sold a Third of It

At $356.47 the equity is worth $123.6 billion. Net the balance sheet it just reported, $4.33 billion of cash and investments against $2.28 billion of convertible notes, and enterprise value is about $121.5 billion. Against the $6.07 billion of product revenue it just guided to, that’s almost exactly 20 times guided full-year product revenue.

Then look at where the price sits relative to the people who cover it. Going into the print the average target was $331.66 and the median $327.50, both below Thursday’s close, and this isn’t a hostile group: 44 of 51 ratings are buy or better. The stock ran past its own bulls. The average has since jumped to about $410, so that pre-print figure is already stale. The sequencing is the point: the market repriced before the analysis did.

That’s the honest reading of the fade. It wasn’t disbelief in the quarter, it was the multiple absorbing it. At the $384.56 high the acceleration wasn’t a surprise any more, it was the base case, and a base case has no upside in it.

For scale, the stock is still about 11% below its November 2021 closing high of $401.89, set at a far smaller revenue base.

The Case Against Me

The bull answer is decent. A company growing 37% and getting faster, retaining 126%, with $9 billion of remaining performance obligations and margins expanding, is not obviously expensive at 20 times guided product revenue. Plenty of slower software trades near that.

The better objection to my own caution runs the other way. Consumption pricing cuts both directions, and Snowflake has already demonstrated this: the same meter that’s running fast now ran slow through 2023 and 2024, when customers optimised their spending and growth collapsed from the 90s into the 20s. Nothing about the model prevents that happening again. A consumption business reprices immediately when its customers decide to spend less, and AI budgets are the least battle-tested line in enterprise software.

So the acceleration is real and the mechanism is real. What isn’t established is that this is a durable rate rather than the fast part of an adoption wave.

The Setup

Snowflake isn’t in the weekly plan we publish, so these levels are new and we’ll be explicit about them.

Read more

The market values AST at $24.3 billion. What it has bought so far is 13 satellites in orbit and $46.3 million of first-half revenue, none of which came from the network everyone owns the stock for. That revenue is gateway hardware and government milestones. The service itself hasn’t earned anything yet.

So this is an execution chain: manufacturing into launches, launches into orbital capacity, capacity into service, service into recurring revenue. Wednesday shortened none of those links. It moved the price.

Key Takeaways

• The $92 target didn’t raise the ceiling. The sell side already carried a mean near $79.60, a median of $82.50 and a high of $108, and most covering analysts still rate it hold or worse.

• First-half revenue was $37.8 million of product and $8.4 million of services, from gateway deliveries and government work. The network itself hasn’t earned anything.

• The 45-satellite milestone has already moved. In May it was framed for 2026; it’s now framed for early 2027 on launch availability.

• Even the equipment guide is a stretch. The second half needs $51.9 million to $76.9 million a quarter against a $54.3 million record.

Prices are the Wednesday September 2 close, the last completed session.

What Wednesday Actually Repriced

Berenberg’s Michael Filatov initiated at Buy with a $92 target, describing AST as the only company to have demonstrated true cellular broadband from space to unmodified smartphones, and citing more than 60 mobile operator agreements covering roughly 3 billion subscribers.

Every word of that was equally true on Tuesday, when the stock closed at $55.80.

Look at where $92 actually sits. It came in below the existing high of $108 and above the median, so it added nothing to the top of the range, and the shares moved 12% on more than double the prior week’s average volume.

The insider purchase is real and I’d rather see it than not. Director Adriana Cisneros bought 10,822 shares on August 31 at $57.00 to $58.87. That’s $619,000 against a $24.3 billion market value, roughly 1 part in 39,000.

So the move was a re-rating on sentiment. That’s allowed. It matters because of what sits underneath it.

What the $24 Billion Is Actually Buying

Split the first half properly. Of the $46.3 million, $37.8 million was product revenue and $8.4 million services: gateway hardware delivered to operators, and milestones under United States government contracts. It’s real revenue and it’s growing fast. It just isn’t the network.

Put numbers on what the network still needs. AST says roughly 25 satellites give about half a day of coverage, and 45 to 60 give continuous coverage across key markets including the United States, Europe and Japan. It has 13, and has activated an initial 3,000 digital cells across the continental United States.

Now watch the chain move in real time. In May, management framed roughly 45 satellites in orbit during 2026. The current framing puts 45 satellites and continuous service in early 2027, on launch availability. That breaks nothing, and I’m not calling it a crisis. It’s a worked example of why this is the thing to watch: a major milestone has already slid by about 2 quarters.

There’s a second link AST doesn’t control. It describes 2026 beta service as scaled non-commercial usage with partner operators, and hasn’t committed to when that turns commercial. Its president has said the carriers, AT&T and Verizon among them, decide when to open the service.

This is the part the multiple debate keeps missing. At 139 times guided revenue AST looks absurd, and the reply that 2026 revenue is irrelevant to terminal value is fair. Both miss the point. The question isn’t whether 139 times is expensive. It’s how many unfinished milestones are already inside $24.3 billion.

Even the Equipment Guide Is a Stretch

AST reported $14.7 million of revenue in March and $31.5 million in June, so $46.3 million in the first half, and reaffirmed full-year revenue of $150 million to $200 million. That leaves $103.7 million to $153.7 million for 2 quarters.

Put all 3 cases against the $54.3 million record, set in the final quarter of 2025. The bottom needs $51.9 million a quarter, 4.5% under the record. The midpoint needs $64.4 million, 18.5% above it. The top needs $76.9 million, 42% above. So the low end asks for 2 near-record quarters; the midpoint asks for the record beaten twice.

There’s a mechanism behind it. This revenue is lumpy and contract-shaped rather than recurring, drawn from roughly $1.3 billion of contracted commercial and government backlog, so it steps rather than compounds, and the $54.3 million quarter proves the step exists. But June came in at $31.5 million, below what analysts expected, and the run rate entering the second half sits far under what the guide needs. Backlog is a schedule, not a quarter.

The Financing Dependency

AST consumed roughly $1.37 billion of free cash across the 4 quarters through March, about $340 million a quarter, and adjusted operating expenses rose again in June, to $119.1 million from $91.2 million. In the March quarter it raised about $1.13 billion net across debt and equity.

At June 30 it held $2.29 billion of cash and equivalents, or $2.72 billion including restricted cash, against $2.96 billion of long-term debt. It’s net borrowed on either definition. In July it sold another $1.15 billion of convertible notes at 1.625%, lifting its own pro forma cash above $3.7 billion.

Be careful with that, because the obvious overstatement is that AST is about to run out of money. It isn’t. More than $3.7 billion against roughly $340 million a quarter is years of runway, and management says liquidity covers anticipated requirements for at least the next 12 months.

The accurate claim is narrower and still forceful. This network has been built on external capital rather than its own cash flow, and the economics stay sensitive to the cost of that capital until service revenue carries the load. That’s why Fed Chair Kevin Warsh signaling that inflation was back at the top of the agenda hit this name harder than most. A rate scare is a valuation input for most stocks. Here it reaches the funding of the asset.

Redwire Is the Control, Not the Comp

Redwire broke down the same week for none of the same reasons, and it’s here as the financing-risk control, not the valuation comparison. Nobody should argue it deserves AST’s multiple.

It did $214.0 million in the first half against guidance of $450 million to $500 million, so its second half needs 10% to 34% growth, and it puts visibility into that midpoint at 90% on record backlog of $542.1 million. It holds $557.0 million of cash plus restricted against $48.1 million of debt, close to 20% of its market value in net cash, and burned $35.3 million last quarter. It rose 24% across August, then gave ground into month end, closing at $10.33.

That’s the contrast. AST has vastly more upside if the network works. Redwire has vastly more room to be late. When what you own is an execution chain, the ability to absorb delay is worth something, and only one of these has it.

The Case Against Me

The counter that most threatens my argument is backlog. AST carries roughly $1.3 billion of it and booked over $125 million of new United States government awards in the June quarter. Converting $103.7 million inside 6 months is a scheduling question, not a demand question, which is why the guide is a stretch rather than a fantasy. But backlog converts on a timetable the company hasn’t demonstrated, and the first half converted $46.3 million.

The other counter is that I’m marking a venture asset against a revenue line that was never the point. If the constellation gets to 45 satellites and service starts, 2026 revenue becomes a footnote. More than 60 operator agreements across 3 billion subscribers is a distribution position nobody else has assembled, and AST really has shown broadband to an ordinary handset from orbit.

I accept most of that. This isn’t a short thesis and I’m not arguing the technology can’t work. My objection is to the trigger. A stock that can only be valued on execution ought to move on execution.

The Setup, and What Would Change It

Read more

Measured from Thursday’s close, before the selling started, the damage is nearly the same: 26.1% for PG&E, 26.7% for Edison. Both ended with a 14-day momentum reading of 22, deep into oversold. They were traded as one position. They aren’t one risk. Edison has a live claim whose gross estimate runs to 65% of its market value. PG&E has a rule change and no fire.

Key Takeaways

• The SB 492 compromise leaves out the $6 billion per-incident cap on Wildfire Fund withdrawals, any mechanism to refill the fund, repeal of its 2028 sunset, and Newsom’s proposal to stop insurers suing utilities. It has not been voted on yet.

• Edison’s Southern California Edison faces the Eaton Fire, with gross potential liability estimated on the sell side at about $13.5 billion before recoveries, against a $20.8 billion market value.

• PG&E has no equivalent live event. BMO left its earnings estimates untouched at $1.65, $1.82 and $1.98 for 2026 to 2028.

• BMO raised its explicit liability discount by $4 a share. PG&E’s shares fell $4.68, more than the whole increase, to 0.86 times book.

Prices are the Monday August 31 close, the last completed session: $13.27 for PG&E, $53.98 for Edison. Tuesday is pre-open as this is written.

What the Compromise Contains

The text is a wildfire measure written for survivors, not utilities. It creates a “Fast Pay” route to victims, restricts selling wildfire claims to third parties including private-equity buyers, and bans bonuses for senior executives when their company causes a fire destroying 500 structures or more. It preserves the right to sue.

It’s also not law. Utilities were still pushing for late changes on Monday, and until the vote is taken the terms can move. What the text leaves out is the package the sector had been pricing. No $6 billion per-incident cap on what a single fire can draw from the Wildfire Fund. No mechanism to refill it once drained. The 2028 sunset stands. And the proposal to bar insurers from pursuing utilities through subrogation didn’t survive.

So the fund stays finite, unreplenished and on a clock, with every avenue to sue preserved. That’s a genuine deterioration for anyone who owns a California utility, and both stocks deserved to fall.

The market did discriminate, just not between these 2. Sempra, which owns San Diego Gas and Electric, fell about 2% on the same news, cushioned by Texas and Mexico infrastructure and a far smaller California footprint. The selling was calibrated to California exposure. Within California it wasn’t calibrated to anything.

Why Edison’s Fall Is Defensible

Edison owns the problem the bill declined to solve.

Los Angeles County fire investigators concluded that electrical arcing from an out-of-service Southern California Edison transmission tower caused the January 2025 Eaton Fire, which killed 19 people and destroyed thousands of homes. Sell-side estimates of the eventual bill run to about $13.5 billion.

Size that properly, twice over. First, the ratio moves as the stock does: $13.5 billion was 48% of Edison’s market value before the selling and 65% after. The 2 sessions removed $7.58 billion of equity, 56% of the whole estimated figure.

Second, and more important, that estimate is gross potential liability, not shareholder loss. Edison carries the first $1 billion of wildfire claims through a customer-funded self-insurance programme before it reaches the Wildfire Fund at all. It has committed about $1.6 billion to Eaton settlements, and it is already booking recoveries against those losses: at the first quarter, against $1.3 billion of recorded losses, it had recognised $917 million of expected self-insurance recovery alone. Add the fund, transmission-side recovery and whatever the regulator later allows in rates, and what lands on equity is a fraction of $13.5 billion. The market isn’t pricing $13.5 billion. It’s pricing the uncertainty around how much gets through.

The company isn’t conceding. It argues the tower had been idle since 1971 and so doesn’t trigger California’s strict-liability rule, and on August 12 a Los Angeles judge tentatively agreed, declining an insurers’ motion to hold Edison automatically liable. A real win, but tentative, on one motion, and it leaves the harder question open: whether the regulator finds Edison acted prudently.

That finding is the key determinant of how much of Eaton reaches shareholders, though not the only one: civil liability, negligence, settlement levels and what the regulator later allows in rates all matter. Under the 2019 law that created the Wildfire Fund, a utility found prudent generally doesn’t reimburse it. Found imprudent, it does, but reimbursement is still capped at 20% of its transmission and distribution equity rate base over a rolling 3 years. Losing the prudency finding doesn’t by itself remove that cap. It falls away only in narrower circumstances, chiefly the absence of a valid safety certificate or a finding of conscious or willful disregard.

So Edison carries a large contested claim into a regulatory judgment nobody can handicap, drawing on a fund the legislature has just declined to cap, refill or extend. Marking that down 27% isn’t panic. It’s what happens when the insurance around a live claim thins and the claim stays unresolved.

Why PG&E’s Isn’t

PG&E’s position is the opposite. It has no Eaton-equivalent live catastrophe. Its major legacy fires are substantially provisioned and far further through the process, with the bankruptcy’s Fire Victim Trust paying claimants 70 cents on the dollar, though not finished: Dixie still has litigation running and an open proceeding to recover costs. It holds its safety certification, the qualification that keeps the liability cap available, and has buried more than 1,300 miles of line in high-risk territory since 2021. None of that changed on Monday.

Here’s the arithmetic that decides it. BMO downgraded the stock and raised its explicit wildfire liability discount from $6 a share to $10, assuming uncapped liabilities beyond 2030. That’s a $4 increase in the modelled cost of the thing everyone is frightened of. Over the 2 sessions, PG&E’s shares fell $4.68.

The decline exceeded the entire increase in that discount. And that same analyst left its earnings estimates entirely unchanged at $1.65 for 2026, $1.82 for 2027 and $1.98 for 2028. The operating business it models is exactly the business it modelled a week ago.

At $13.27 that’s 8.0 times this year’s estimate, 7.3 times next year’s and 6.7 times 2028, at a 52-week low.

Book value tells the same story. PG&E trades at about 0.86 times accounting book equity, against roughly 1.19 for Edison. Book equity isn’t the regulated rate base and shouldn’t be read as one, but a utility trading below it is unusual, and the market is applying that discount to the company without a fire rather than the one with it.

PG&E’s own view is worth noting. It said publicly that SB 492 does not adequately address the financing risks in the liability framework. That’s an argument the bill is bad, not that the business broke.

The Case Against Me

There’s a real argument on the other side.

PG&E is the largest operator in California’s high fire-threat territory, and its share of any utility contribution to the Wildfire Fund is roughly 48%. A fund that can be drained with no way to refill it is a bigger prospective problem for PG&E than for anyone, since it’s most likely to need it and most liable to rebuild it. Edison’s claim has first call on money that exists today.

Read that way it isn’t mispricing at all: Edison’s problem is large but partly backstopped by a fund that can still pay, while PG&E’s risk is prospective and may arrive only after that protection has been weakened. Partly, not wholly: the cap covers reimbursements to the fund, not every conceivable Eaton loss, so Edison’s exposure has no clean ceiling either. This is a company that went through Chapter 11 in 2019 over exactly this risk.

One more piece of honesty. The 3 firms that cut PG&E left targets of $16, $21 and $24, all above the price, but Mizuho’s cut target on Edison is $70, also well above it. Targets above the market tell you how fast prices moved, not which name is mispriced.

I think that argument explains why PG&E fell. It doesn’t explain why it fell more than the bear case, to a 52-week low, at under 7 times 2028 earnings, with unchanged estimates. Prospective risk on a business whose numbers didn’t move is a de-rating. This was a liquidation.

Our trade plan

Read more

Then it did something it had never done: put a preliminary number on the following fiscal year, about 70% revenue growth, when analysts had 44%. The stock rose 8.7% on Thursday and gave back 57% of it on Friday. Demand volume has stopped being the question. What replaced it is harder: what Nvidia earns on that demand, and how much of its own balance sheet has to stand behind the buyers.

Key Takeaways

• Revenue $96.2 billion, up 106%, with data centre at $89.0 billion, up 117%. Gross margin was 75.0%, up 260 basis points on the year.

• The October guide cleared the bar by 3.6%, and a first-ever year-ahead figure of 70% sits 26 points above consensus. Management calls it supply-constrained.

• Margin falls to 74% next quarter, troughs at 71% to 72% in January, then settles at 72% to 73%. Even so, gross profit still grows about 64%.

• The balance sheet is now part of the product. $99 billion of equity investments, $25 billion more committed, and up to $105 billion of credit support behind a single customer’s data centre.

• At $217.55 the stock is roughly 23 to 24 times this fiscal year. The 14 times you’ll see quoted for next year is a bet on the denominator, not a fact.

Prices are the Friday August 28 close of $217.55, the last completed session. Nvidia reported on Wednesday August 26 after the market shut.

What Got Settled

Before the print, the argument here was that the hurdle had moved off the reported quarter and onto the October guide and the gross margin. Both were answered on the night, in opposite directions.

The guide was emphatic. $108 billion is 3.6% through the bar and a sequential step of $11.8 billion, larger than Nvidia’s entire quarterly revenue as recently as early 2023. Jensen Huang’s line for it: “AI has reached its inflection point. It’s doing useful work. Now, compute is revenue.”

Then the part with no precedent. Nvidia has never put a number on a year in advance. It offered one as a preliminary expectation rather than formal guidance: about 70% growth in fiscal 2028, against consensus near 44%. On the base Nvidia is tracking toward this year, 26 points is worth more than $100 billion of revenue, more than it sold all last quarter.

The framing is the tell. Management describes that number as supply-constrained. Nvidia isn’t forecasting how much it can sell, it’s estimating how much it can build. When a company moves from demand guidance to supply guidance, the volume question has stopped being the interesting one.

One caution on the composition. Data centre revenue was $89.0 billion, split between $48.7 billion of hyperscale and $40.3 billion of AI clouds, industrial and enterprise, the latter growing faster at 138%. That looks like the buyer base broadening, and partly it is. But Nvidia says some of that faster-growing bucket is hyperscalers buying through AI clouds. Customer concentration is falling faster than end-demand concentration, and only the first of those is visible in the segment lines.

Demand Quality

Gross margin was 75.0%, GAAP and non-GAAP alike, up from 72.4% a year earlier. That’s the durability test passed, and it’s why Thursday happened.

Then the path. Margin is guided to 74.0% in October, to bottom in the January quarter at 71% to 72%, and to settle at 72% to 73% through fiscal 2028. The finance chief’s word for memory was scarcity, driven in large part by the AI buildout itself, which is to say by Nvidia’s own customers bidding for the same components. And the settling happens after price increases take effect: Nvidia is raising prices and still lands 2 to 3 points below where it is today.

Here’s the balance the bearish reading misses. On the revenue Nvidia is pointing to, a 72.5% margin still produces roughly $490 billion of annual gross profit against something near $300 billion this year. Gross profit grows about 64% even as the margin falls. The reset costs real money, roughly $17 billion a year at that scale, but it doesn’t stop the compounding.

So the danger isn’t 72.5%. It’s 72.5% becoming 69%, or 70% growth becoming 45%, or both at once. That’s the sensitivity that matters, and it’s why Friday’s reaction was about revision risk rather than about this quarter.

Friday itself deserves a caveat. The margin outlook gave investors a reason to reassess Thursday, but the macro backdrop gave them a reason to act: Fed comments at Jackson Hole pushed yields higher and hit long-duration technology broadly. Nvidia fell 4.6%; the Nasdaq fell too, if far less. Attribute the direction to the margin, not the whole magnitude.

The Balance Sheet Is Now Part of the Product

This is the part that deserved more attention than it got, and it’s bigger than the phrase “circular financing” allows.

The filings show $99 billion of equity investments and a further $25 billion committed. In August, Nvidia agreed to provide up to $105 billion of credit support behind an OpenAI data centre in Ohio, backing an initial 4.25 gigawatts with an option on 3.75 more. Separately it signed memorandums with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500 billion of third-party capital for customers. Those are preliminary arrangements, and Nvidia’s own filing warns they may never become definitive agreements.

Management’s own framing is the most revealing part: the AI labs for which it expects to lean on its balance sheet could be about a quarter of next year’s business.

Read together, Nvidia is running 3 businesses at once. It sells accelerators. It owns equity in the companies that generate the demand for them. And it increasingly acts as a market maker for AI infrastructure, using guarantees, commitments and relationships with outside capital to turn theoretical compute demand into financeable projects.

When it works, this is a widening moat, not a warning. Nvidia sells the systems, makes the project financeable, appreciates alongside the labs it owns, and deepens its software’s installed base.

The risk isn’t that any of it is improper. It’s correlation. Nvidia’s revenue, its customers’ access to funding, and the value of its own investments have become the same bet. If AI returns hold, capital gets cheaper, orders grow, and the stakes appreciate together. If they don’t, financing tightens, utilisation falls, guarantees start to matter, and the equity marks fall, all at once and for the same reason.

One thing that isn’t a risk right now. Data Center Hopper shipments to China were under 1% of data centre revenue, they’re dilutive to gross margin, they carry a 25% tariff Nvidia says it can’t pass through, and the company took a roughly $400 million charge on excess H200 inventory. There’s no China data centre compute revenue in the outlook. That’s excluded upside, not a free option.

What You’re Paying

At $217.55 Nvidia is worth about $5.25 trillion, roughly 23 to 24 times consensus earnings for the fiscal year ending January 2027.

The 14 times you’ll see quoted for fiscal 2028 deserves more scepticism than it usually gets, because it’s a statement about the denominator dressed as a statement about the price. At Friday’s close, $13 of earnings is 16.7 times, $14.50 is 15.0, $15.50 is 14.0 and $16 is 13.6. Saying the stock trades at 14 times next year is saying you believe next year’s earnings are about $15.50, which requires both the 70% revenue estimate and the 72% to 73% margin to hold in the same year.

That’s the honest version. The stock may look cheap because the market doubts the denominator, not because investors have forgotten arithmetic. The average analyst target near $305 sits about 40% above Friday’s close, which tells you where the sell side has landed on the same question.

The Setup

For all the drama, the price says almost nothing happened. Nvidia closed between its 20-day exponential average at $215.67 and its simple 20-day at $218.05, with a 14-day momentum gauge at 52 and a trend-strength reading near 17, which says no trend is established either way. It sits 8% below the May high of $236.54, after the largest quarter in its history. Both the daily and weekly timeframes reclaimed a full trend on the print.

These are the levels from this week’s plan, unchanged.

Read more

Key Takeaways

• Adjusted earnings of $5.90 included $2.53 from investment gains, which the company discloses. The operating figure was $3.37 against a $3.27 estimate.

• The same distortion runs through the guidance. Full-year adjusted earnings guidance rose $2.60 at the midpoint, and $2.53 of that is the quarter’s investment gain.

• What actually improved: contracted revenue due within a year grew 14%, and the finance chief says new order value is growing at its strongest rate in 4 years.

• Recognized revenue hasn’t turned yet. Growth was 11%, but $456 million came from Informatica; without it the rate was about 6%.

• The cheap-looking multiple is doubly flattered: by investment marks, and by a $25 billion buyback funded with debt.

Prices are the Friday August 28 close, the last completed session. Salesforce reported on Wednesday August 26 after the market shut.

The Number That Wasn’t

Give Salesforce credit for the disclosure. The release states plainly that gains on strategic investments of $2.613 billion raised adjusted earnings per share by $2.53 and the statutory figure by $2.43, and Reuters and others picked it up. But an 80% beat is a loud number and a $2.53 footnote is a quiet one.

The arithmetic is worth doing slowly. Adjusted earnings came in at $5.90 against an estimate of $3.27, a surprise of $2.63. Of that surprise, $2.53 was the investment mark. That’s 96%. The business itself delivered $3.37, which beat by 3% and grew 16% on the year rather than the 103% the headline showed.

The same distortion runs through the outlook, and it’s the better statistic. In May, Salesforce guided full-year adjusted earnings to $14.06 to $14.12. It now guides $16.67 to $16.71. The midpoint rose $2.60. The quarter’s investment gain was $2.53. So 97% of the apparent uplift in the full-year earnings guide is that same one-off mark, not an improved forecast for the software business.

None of this is an accusation. Salesforce’s adjusted definition includes these gains as a matter of course, and did last quarter too, when they added $0.51. But a number that moves on the private valuation of a company you don’t control tells you nothing about whether Salesforce sold more software.

The Number That Was

It sold more software. That’s the part worth owning.

Contracted revenue due within the next year reached $33.5 billion, up 14% both as reported and in constant currency, a point ahead of guidance and, on the currency-adjusted basis, an acceleration from 13% in each of the previous 2 quarters. Total contracted revenue was $66.3 billion, up 11%. Adjusted operating margin held at 34.1%.

Robin Washington, the finance chief, put the important sentence on the record. New annual order value, the worth of business newly signed, is growing at the strongest rate in 4 years, “keeping us on track for second-half organic revenue reacceleration.”

That’s the sequence a real software recovery follows: orders turn, contracted balances follow, recognized revenue arrives later, cash later still. If you want a reason the stock deserved to re-rate, it’s this and not the earnings line.

Now the discipline. Recognized revenue hasn’t followed yet. Growth was 11%, and $456 million of it came from Informatica, which Salesforce didn’t own a year ago. Strip that out and the underlying rate was about 6%. Read the finance chief’s sentence again: reacceleration is described as on track for the second half, not as something that happened. The guidance raise says the same. Of the $300 million added in constant currency, $200 million is the pending Contentful and Fin acquisitions and $100 million is organic, against a $100 million currency headwind. The organic raise and the currency hit cancel.

The AI numbers need a similar reading. Agentforce recurring revenue passed $1.5 billion, up more than 240%, and with Data 360 the pair is near $3.9 billion. Agentic workflows ran 3.2 billion actions in the quarter, up 97% on the previous one. But from this quarter Salesforce broadened what counts as Agentforce revenue to include Slackbot and Headless 360, so the 240% isn’t a clean like-for-like. And at $1.5 billion, Agentforce is about 3% of the revenue Salesforce expects this year. Real, growing fast, not yet moving the total.

So investors are paying for a turn in the leading indicators, not in reported revenue. That’s the risk and the opportunity.

The Anthropic Mark

The same day it reported, Salesforce announced Claudeforce: Claude inside Agentforce and Slack, and Salesforce data inside Claude through a plug-in with 37 prebuilt sales skills, in open beta next month. Claude becomes the first outside model to run fully inside Salesforce’s trust boundary, making one supplier the default. Salesforce will spend roughly $300 million on Anthropic tokens this year.

The shareholding did rather better than that. The filing attributes $2.7 billion of unrealized gain in the quarter to Anthropic, about 9 times the entire year’s planned token spending, and the stake has gone from roughly 22% of Salesforce’s venture book in January to about 45% of an $11.3 billion portfolio in July. Salesforce doesn’t disclose what it paid for it.

Compare it properly and it’s still striking. The $2.7 billion is pre-tax, while the $3.37 of operating earnings is after tax. Apply the roughly 20% adjusted tax rate implied by the company’s own per-share disclosure and the Anthropic contribution is nearer $2.2 billion, against about $2.8 billion the software business earned. A minority holding in someone else’s company was worth about 78% of everything Salesforce sold.

Worth understanding how the mark happens. Salesforce carries private holdings under the measurement alternative: carrying value moves when an observable transaction in the same security occurs. This isn’t a company deciding its stake feels more valuable, it’s a funding round repricing it. The gain is unrealized, isn’t repeatable, and reverses if private AI valuations do.

What You’re Paying

Salesforce is worth about $211 billion. Against full-year guidance of $16.67 to $16.71, that’s 15.3 times earnings, which reads cheap for a business compounding bookings at 14%.

2 things flatter it. The first is the one above: $0.51 of first-quarter gains and $2.53 of second-quarter gains mean $3.04 of that guide is marks rather than operations. Strip them and you’re paying nearer 18.8 times, about 22% dearer than the screen says. Cut the other way too. If you remove the investment earnings you should acknowledge the investments themselves, and the $11.3 billion venture book still sits inside the market value. Net it off and the operating multiple is nearer 17.7 times.

The second is more interesting. In May, Salesforce launched a $25 billion accelerated buyback funded by $24.8 billion of new debt. That lifts earnings per share by shrinking the share count, and it’s why cash flow guidance is so weak: the company attributes roughly 5 points of headwind on operating and free cash flow growth to the debt issuance itself. Revenue is guided to grow 11% to 12% and cash flow 4% to 5%. That gap is financial engineering, not AI capacity. So the falling multiple partly reflects a smaller denominator rather than a better business, and with the balance sheet levered, enterprise value now tells you more than the price-to-earnings ratio does.

One accounting note. Adjusted operating margin was 34.1% against a statutory 20.5%, about 14 points apart on the full-year guide, mostly stock compensation and acquired-intangible amortization. Neither is a cash cost this quarter. Neither is economically free either: stock compensation transfers value through dilution unless buybacks offset it, and amortization reflects acquisition capital already spent.

The Setup

The strangest fact in the story: after its best day in 6 years, Salesforce is still 4.9% below its 52-week high of $269.11, set last December. That’s how deep the de-rating had gone, and the June low near $146 embedded a durability discount that accelerating orders make harder to justify.

It’s also violently extended. Friday’s close sits 23.8% above the 20-day exponential average at $206.79 and above the upper volatility band at $239.94, with the 14-day momentum gauge near 81, deep into overbought. Average true range is $10.93, so the 2-day move was worth more than 4 normal sessions.

Read more

Those 2 figures aren’t the same kind of thing, and Palo Alto publishes the arithmetic that shows why. Strip out CyberArk and Chronosphere from its most recent quarter and revenue grew about 14%. Acquired growth is real revenue. It just isn’t evidence about how well a company sells. Fiscal 2027 is when Palo Alto has promised to start showing more clearly which is which.

Key Takeaways

• CrowdStrike added $332.8 million of net new recurring revenue, a record, up 51% on the year, and raised its full-year outlook for that measure by 630 basis points.

• Palo Alto’s last reported quarter showed 31% revenue growth. Excluding the 2 acquisitions it discloses, the rate was about 14%.

• Its own organic figures confirm the method: next-generation recurring revenue grew 28% against a 60% headline, and obligations 22% against 36%.

• The 2 growth engines are the same size. Each added 6.0% to its base in its latest quarter. CrowdStrike’s is accelerating at 51%; Palo Alto’s organic engine grew 18%.

• CrowdStrike trades at 117 times implied fiscal 2027 free cash flow against 68 times Palo Alto’s guided fiscal 2026 adjusted figure, which is about 2.4 times larger.

Prices are the Wednesday August 26 close, the last completed session: $189.18 for CrowdStrike, $339.31 for Palo Alto. Thursday pre-market at about 8:10 was $207.15 and $356.50.

What CrowdStrike Proved

Revenue rose 26% to $1.47 billion and recurring revenue reached $5.84 billion, up 25%. Both fine, neither surprising. The number that moved the stock is the one in between: net new recurring revenue of $332.8 million, the most the company has ever added in a quarter, and 51% more than the same quarter a year ago.

That figure is the hardest one in software to manufacture. It’s the quarter’s incremental recurring revenue, net of expansion, contraction and churn, and it can’t be flattered by a long tail of old contracts. CrowdStrike has taken it from $194 million to $333 million across 6 quarters, a company that has finished digesting the 2024 outage and gone back to winning. Management raised the full-year outlook for that measure by 630 basis points, to 34% growth at the midpoint, and lifted revenue guidance to $5.99 billion to $6.01 billion.

The rest supports it. Non-GAAP operating income grew 46% to $371.6 million, a 25% margin, and operating cash flow rose 59%. Falcon Flex, which lets customers draw down across modules, now carries $2.29 billion of recurring revenue, roughly double a year ago, and 51% of subscription customers run 6 or more modules. CrowdStrike buys companies too. SGNL, announced in January and closed in February, carried about $637 million of recorded consideration. CyberArk alone carried $21.1 billion, roughly 33 times as much.

What Palo Alto Will Report

Palo Alto closed Chronosphere on January 29 and CyberArk on February 11. The quarter it reports Tuesday, ended July 31, is the first full one with both inside the numbers, which is why the guided rate steps up to 32% when the full year is guided to 24%.

Palo Alto discloses what the acquisitions contributed to each headline metric, so the organic rate is a subtraction anyone can do. Last quarter revenue was $3.002 billion against $2.289 billion a year earlier, up 31%, and $388 million of it came from the 2 deals. Take that out and the rate is 14.2%. On the other 2 lines Palo Alto publishes the organic rate itself: next-generation recurring revenue grew 28% rather than the 60% headline, and contracted future obligations grew 22% rather than 36%.

That makes the revenue line testable. The finance chief said it out loud: “Excluding the impact of CyberArk and Chronosphere, NGS ARR was $6.5 billion, up 28% year over year,” and gave 22% for obligations. Run my subtraction against those 2 and it returns 27.9% and 22.7%, close on figures the company rounds before publishing, which licenses using it on revenue, where no organic rate is given and the inputs are exact.

So the number likely to lead every write-up on Tuesday is 32%. The context that matters is that the quarter before it grew 31% on the headline and about 14% once the disclosed acquisition revenue comes out. Palo Alto doesn’t guide an organic revenue rate, so Tuesday is where you find out whether that engine accelerated. The 2 deals behind the gap were announced at a combined $28 billion and recorded at closing at about $24 billion.

The Engines Are the Same Size

Here’s the fair version, and it cuts against the easy read.

Palo Alto’s organic next-generation business added $370 million of recurring revenue last quarter. CrowdStrike added $332.8 million. Palo Alto’s is bigger. Run each as a rate against the base it was added to and they’re identical: 6.0% and 6.0%. Those are different quarters, though, and matched on the calendar Palo Alto is ahead. Across February to April, CrowdStrike added 4.9% of its base, its weakest quarter of the year seasonally, against Palo Alto’s 6.0%.

The difference is the direction. CrowdStrike’s incremental recurring revenue is 51% larger than a year ago. Palo Alto’s organic addition is 18% larger. One engine is accelerating and the other is growing steadily, and over a few years that gap compounds into something enormous. That, and not the headline rate, is what CrowdStrike’s premium is actually buying.

One caveat. Palo Alto’s next-generation figure covers its growth business only, with an older firewall estate attached that grows slowly. CrowdStrike’s covers everything. So the like-for-like is engine against engine, not company against company. And no organic split is perfectly clean once a company sells what it bought alongside what it built.

What You Pay for the Difference

At Thursday’s pre-market price CrowdStrike is worth about $211 billion. Against its own fiscal 2027 guidance that’s 35 times revenue, 165 times adjusted earnings, and 117 times the free cash flow its at-least-30% margin framework implies.

Palo Alto is worth about $291 billion, and against fiscal 2026 guidance that’s just over 25 times revenue, 94 times adjusted earnings and 68 times its guided adjusted free cash flow: $4.28 billion against CrowdStrike’s implied $1.80 billion. Neither is a reported annual figure yet. And the comparison flatters CrowdStrike, because Palo Alto’s fiscal 2026 ended in July while CrowdStrike’s fiscal 2027 runs to next January. Palo Alto targets a 40% cash margin by fiscal 2028, so matched forward the gap is wider.

One adjustment runs the other way. CyberArk was paid for substantially in stock: across the quarter it closed in, Palo Alto’s share count went from 703 million to 813 million, up 15.6%. Per share, the growth is worse than the revenue line suggests. That’s the cost of buying rather than building, and the part the 32% headline hides most completely.

Tuesday

The report itself is close to settled. Consensus and guidance agree on about $3.35 billion of revenue and $0.96 to $0.98 of adjusted earnings, and the $8.90 billion to $8.95 billion of recurring revenue was set when the quarter closed on July 31.

The event is the fiscal 2027 guide, and there’s a specific bar. Palo Alto has told investors it targets $20 billion of next-generation recurring revenue by fiscal 2030. From a fiscal 2026 exit near $8.93 billion, hitting that requires compounding at 22.4% a year. That makes the guide a burden-of-proof number, not a pass or fail. Below 22.4% and the later years have to carry more; above it and fiscal 2027 starts ahead of the straight line. Neither settles the target, since growth needn’t arrive evenly.

The second thing is quieter and matters more. Palo Alto has said it intends to provide segment revenue for Network Security, Cortex and Identity beginning in fiscal 2027. It hasn’t said they arrive Tuesday, so watch first whether they do. Right now the consolidated figure makes it impossible to see whether the firewall business is healthy or melting under an acquired growth story. That’s the real asymmetry here: not that Palo Alto is cheap, but that the reason it’s cheap is about to become checkable.

Which points somewhere. CrowdStrike earned its move: organic reacceleration is the hardest thing here to fake, and the quarterly addition is 72% larger than 6 quarters ago. But at 117 times cash the premium is fully paid, and chasing the gap is paying for information the market already has. Palo Alto is the higher-variance side, priced at a discount that Tuesday either justifies or removes.

The Setup

Both stocks trade as one thing: they bottomed on the same 3 days in June, peaked within a session of each other in August, and Palo Alto rose 5% on Thursday’s pre-market on somebody else’s results.

Palo Alto closed at $339.31, about 15% below its August 13 high of $398.88. The pre-market at $356.50 sits almost exactly on the 20-day exponential average at $355.73 and the Ichimoku baseline at $353.71, the level it broke down through. Average true range is $15.22, roughly 4.5%, and an earnings move is usually larger.

Read more

From this year Intuit stops adding back stock compensation to its adjusted numbers, worth $5.81 a share, while the estimate it supposedly missed was built on the old one. Put the guide back on that basis and it’s about $28.81, roughly 5% above consensus rather than 15% below. The revenue guide really is light. But the number that should worry you is neither of those. It’s 39 million, and the problem behind it isn’t confined to TurboTax.

Key Takeaways

• The $4 shortfall in earnings guidance is a definition change: Intuit now includes share-based compensation, and on the old basis the guide is about $28.81 against estimates near $27.30.

• Revenue guidance of $23.28 to $23.51 billion, growth of 9% to 10%, sits below the $23.74 billion consensus and below the 14% Intuit delivered in fiscal 2026.

• TurboTax units fell 2% to 39 million. The chief executive says price is now the number 1 reason customers leave.

• The slowdown is concentrated: Intuit names Desktop, TurboTax and Credit Karma, and Global Business Solutions still guides to 13% to 14%. TurboTax Live, the lever that carried tax revenue while units fell, slows from 37% to mid-teens.

• The same pattern runs through the business side. Online paying customers grew 3% to 8.9 million while revenue per online customer rose 15%.

Levels and readings are the Tuesday August 25 close of $357.46, the last completed session. Intuit reported after it; the pre-market print at 7:12 Wednesday was $317.90.

Start with the number that isn’t what it looks like.

The Miss That Isn’t

Beginning in the first quarter of fiscal 2027, Intuit no longer excludes share-based compensation from its non-GAAP measures. Its reasoning is that stock comp is a recurring cost of paying people, so leaving it out misrepresents the operating result. Intuit quantifies it twice: $2.02 billion of operating expense, and $5.81 per diluted share.

So the guide and the estimates it was measured against, which run from about $27.30 to $27.34 depending on the provider, are not counting the same thing. Add the stock comp back and guidance is $28.69 to $28.93, roughly 5% above the number it was reported to have missed. Intuit’s own framing agrees: it calls the guide 23% to 24% earnings growth, which only works against a fiscal 2026 base restated the same way.

This deserves credit rather than suspicion. Few large software companies volunteer to stop pretending equity compensation is free, and doing it in the same release as a slowdown makes the headline read like a collapse in profitability that didn’t happen.

The screens will disagree for a while, too. On the new definition the stock traded near 15.5 times forward earnings at Tuesday’s close and 13.8 times in Wednesday’s pre-market; on the old one, about 12.4 and 11.0. Same company, same day, 3 points of multiple depending on which convention your data provider has caught up with.

The Miss That Is

Revenue guidance of $23.279 billion to $23.512 billion, growth of 9% to 10%, sits below the $23.74 billion FactSet consensus. That shortfall is roughly 1% to 2%, which isn’t dramatic in itself. The deceleration is what stings: fiscal 2026 grew 14%.

Intuit named the causes itself, and they’re worth quoting because the popular summary has them wrong. The deceleration is “primarily driven by Desktop Ecosystem, TurboTax, and Credit Karma.” Mailchimp isn’t on that list. It’s guided to flat or down 1%, which is poor, but small and already poor.

They aren’t the same kind of problem. Desktop grew 6% last year and is guided to decline in the low single digits as customers migrate to the online products, which is a transfer inside the company rather than a loss. Credit Karma grew 20% and guides to 11% to 13%, which management calls prudence after a strong year rather than deterioration. Neither is an emergency, and Global Business Solutions, the largest piece at $12.9 billion, grew 16% and still guides to 13% to 14%.

That leaves TurboTax, and it is the one that matters.

39 Million

US TurboTax units fell 2% in fiscal 2026, to 39 million. Online units fell 2% to 34.9 million and desktop fell 7% to 4.1 million. Revenue still grew 7%, because TurboTax Live revenue rose 37% and now makes up 53% of the franchise. Intuit sold tax software to fewer people and made more money doing it.

The mix underneath is worse than the headline. TurboTax Live customers grew 38% while total units fell 2%, which strongly suggests the count of people filing on their own without paid help fell by considerably more than 2%. That base is the top of the funnel, and it’s shrinking faster than the segment total lets on.

Sasan Goodarzi said the quiet part out loud on the call: “Consumers have more lower-cost alternatives and we lost quality DIY customers to lower-cost providers this year. Price is now the number 1 reason customers leave TurboTax.”

That single admission explains the whole guide. Intuit is deliberately accepting lower revenue per customer to stop the unit decline, which is why TurboTax is guided to 2% to 3% and the Consumer segment to 4% to 6%. It’s sacrificing price to rebuild a funnel it let weaken, and the fiscal 2027 guide is the invoice.

And the offset is fading. TurboTax Live, which carried the segment while units fell, is guided to mid-teens growth this year, because fewer customers are left to upgrade after 2 strong years of conversion. So Intuit is cutting price into the base at the exact moment its best way of making more per customer slows down.

And it isn’t only tax. On the business side, total online paying customers finished the year at 8.9 million, up 3%, about 2 points slower than the year before, while revenue per online customer rose 15%. Across both sides of Intuit, monetisation has been doing considerably more of the work while customer acquisition slowed. That’s the equation management is trying to reverse, and rebuilding it will weigh on fiscal 2027 growth.

So the question isn’t whether the guide is soft. It’s whether the customer count turns, and the honest scorecard is wider than units alone: units, share of total e-filers, revenue per DIY customer, and how many of those filers end up on Credit Karma too. Reaching 40 million with revenue per customer collapsing and no crossover wouldn’t be a win. If units keep sliding while price falls anyway, Intuit will have discovered that the pricing power was the business.

What They’re Buying

Management’s defence is specific. Customers who use both TurboTax and Credit Karma generate roughly twice the revenue of single-product customers, and Credit Karma members filing through TurboTax grew more than 50%. On that arithmetic a cheaper return isn’t a discount, it’s an acquisition cost for a larger relationship. The open question is causality. Dual-product customers do generate twice the revenue, but Intuit has to show that a lower entry price creates more of those relationships rather than discounting people who’d have filed anyway. They may simply be the higher-value customers to begin with. Intuit is also widening distribution, including inside the AI assistants people already use, which is where its Anthropic and OpenAI agreements show up.

There’s evidence it works somewhere. Intuit’s 3 big bets, assisted tax, money and mid-market, grew 34% together and now make up 30% of revenue. Operating margin expanded again and is guided to expand again next year on savings from the roughly 3,000 jobs cut in May. Beyond fiscal 2027 the stated goal is still double-digit growth. This isn’t a company in trouble. It’s one deliberately giving up near-term growth to rebuild customer acquisition, and asking the market to wait.

One detail tells you when the answer arrives. The first quarter is guided to about 11%, faster than the 9% to 10% full year. With the TurboTax reset landing in tax season, the cleanest test comes in the January to April quarter.

The Setup

Intuit closed Tuesday at $357.46, already down 3.4% before the report landed, and 49% below its 52-week high of $705.08 set last September. The pre-market print of $317.90 takes that to 55%. Tuesday’s bar was constructive: above the 20-day and 50-day exponential averages at $340.54 and $326.48, with ADX near 30 indicating a trend rather than a drift. The gap erases that.

What it doesn’t erase is the base underneath. Late July and early August produced a run of lows between $299.63 and $316.05, and the 78.6% retracement at $302.14 and the lower band at $303.90 sit in that same area. The stock is gapping back into ground it spent 3 weeks building on. Average true range is $14.05, about 4%, wider than the zone itself, so size for that rather than trying to be precise about the entry.

1. The zone: $300 to $312. Invalidation is a close below $295, under the whole of that base. Beneath it there’s little until the June low at $252.84.

2. From the middle of the zone you’d risk about $11 to reach the 50-day at $326.48, close to 2 to 1, and roughly 3 to 1 to $340.84.

3. Reclaim: a close back above $327, over the 50-day and the Ichimoku baseline at $325.96. Objectives $340.84 at the 61.8% retracement, where the 20-day average also sits, then $368.02 and $395.21. Mind the geometry. Against the same $295 stop that’s $32 of risk, so only the last objective pays about 2 to 1; the first pays 0.4 to 1 and the second 1.3 to 1. Entering on the reclaim costs most of the reward.

4. Not at the open. The opening print isn’t confirmation of anything, and a gap this size can produce noisy price discovery before a level holds.

Bottom Line

Both of these are true, and most of the coverage will only carry the first. Intuit guided fiscal 2027 revenue below consensus and decelerated from 14% to 9% or 10%, which is real. And it did not miss on earnings by $4, because the estimate and the guide were denominated differently, and like for like the guide is above the number. Anyone building a view off the headline comparison is working from an artefact.

The thing worth tracking isn’t the multiple, which now depends on whose definition you use, and it isn’t the revenue guide in isolation, because part of that slowdown is deliberate. It’s customers. The apparent EPS problem is mostly definitional. The operating problem is customer acquisition, and it exists on both sides of the house. TurboTax has 39 million units and had 39.9 million a year ago; the business side added 3% more paying customers while charging each 15% more. Tax season is where that gets settled.

This is research and commentary, not personal investment advice. Levels are illustrative; size positions to your own risk tolerance and time horizon. The author may hold positions in names discussed.

Key Takeaways

• The October guide is the event, and the line is about $104.2 billion, which would be Nvidia’s first $100 billion quarter. Anything merely in line invites the same reaction as the last 4 prints.

• Gross margin near 75% is the durability test. DRAM contract prices rose about 90% in the first quarter of this year and 58% to 63% in the second, and high-bandwidth memory sits inside everything Nvidia ships.

• Server prices are already moving. Customers were told last week that systems containing Nvidia chips will cost more than 15% more from early next year, on memory costs.

• Options price a 5.4% move, about $280 billion, below the 6.5% implied before May and well under the 7.4% average of the last 12 quarters.

• Nvidia is about 8% of the S&P 500. Anyone holding a broad index fund has a position in this report.

Levels and readings below are the Monday August 24 close of $208.48, the last completed session.

Start with what the pattern actually shows.

Beating Stopped Working a Year Ago

4 consecutive beats, 4 consecutive declines. August 2025: beat 4.0%, fell 0.79%. November: beat 3.2%, fell 3.15%. February, the worst: beat 5.2%, fell 5.46%. May: beat 5.6%, fell 1.77%. The size of the beat has no visible relationship to the reaction.

Read that as a statement about where the hurdle sits rather than a prediction about Wednesday. Backward-looking estimates stopped being the binding constraint some time ago, which is why the most repeated line you’ll see this week, that Nvidia is likely to beat, carries close to no information.

And 4 observations are a pattern, not a probability. The setup is also different this time. Nvidia goes into this print after 7 straight lower closes rather than near its highs, and the options market is asking 5.4% against 6.5% before May and a 7.4% average across the last 12 reports. Less optimism is embedded than in February. The hurdle has moved, and it may also have come down.

The quarter itself.

What They’ll Print

Nvidia guided to $91.0 billion give or take 2%, gross margin of 74.9% reported and 75.0% adjusted, and operating expenses near $8.5 billion. Consensus has crept to about $92.2 billion and $2.09 a share, so the Street is asking for roughly 1% more than management already promised. The quarter laps $46.7 billion, which makes even the guide 95% growth.

Last quarter did $81.6 billion, a figure the filings and the aggregator feeds confirm independently, with data centre at $75.2 billion, up 92% on the year. Operating margin reached 65.6%, up from about 61% 18 months ago. Use operating margin rather than the headline net figure here: reported net income is inflated by $15.9 billion of gains on equity securities in that quarter alone, which says nothing about selling accelerators. Ignore the 49% printed in the April 2025 quarter as a baseline: it carried a $4.5 billion H20 charge and was closer to 59% without it.

Now the 4 things that decide the reaction.

2 Numbers and 2 Narratives

Number 1, the October guide. This is the event, and the line is about $104.2 billion, which would take Nvidia through $100 billion in a quarter for the first time. That number gives you a decision framework rather than a vibe. A guide meaningfully above it forces the Street to revise the forward earnings path higher. A guide that lands on it says Nvidia has not moved that path at all, which is what produced the last 4 reactions. A guide below it says growth is decelerating faster than investors assumed, and nearly 100% growth in the reported quarter becomes irrelevant.

Number 2, gross margin. Management targets the mid-70s this fiscal year. Be precise about where the pressure actually sits, because the popular version of this argument is wrong. Conventional DRAM contract prices rose about 90% in the first quarter of this year and another 58% to 63% in the second. High-bandwidth memory, the kind bolted to the accelerator itself, has not moved like that: TrendForce has HBM3e prices roughly flat to slightly higher in 2026, with the multiple-fold increase expected in 2027 as HBM4 becomes the mainstream generation. So today’s squeeze is at the rack level, in system memory and the rest of the bill of materials, rather than on the chip. The HBM cost step is ahead of Nvidia, not behind it.

That distinction is visible in the pricing. Bloomberg reported on August 22 that contract manufacturers building systems for Microsoft, Google and Oracle have told customers prices will rise more than 15% on systems shipping early next year, on memory costs, covering both Grace Blackwell and Vera Rubin configurations. Somebody is absorbing that, and the release should tell you who. The real question isn’t whether memory inflation squeezes Nvidia. It’s whether Nvidia can hold roughly 75% while ramping a new architecture, shipping ever more complex rack-scale systems, and operating in a supply chain whose component economics are changing underneath it. Hold the margin and it still sets the terms. Slide into the low 70s while revenue is booming and suppliers and customers are starting to take a share. That question outlives Wednesday.

Narrative 1, the Blackwell hand-off. Nvidia said in July that Vera Rubin had reached full production, and names cloud partners running it; Reuters’ preview this week still describes first shipments as an autumn event. The reconciliation matters less than the risk underneath it, which Reuters names directly: this is a hand-off from Blackwell, and any supply hiccup or slower upgrade cycle shows up in guidance fast. Transitions cost margin before they earn it. What matters is whether Rubin extends the cycle or interrupts it.

Narrative 2, China. This has moved and most previews haven’t caught up. Washington changed the licensing rule in January, and for months almost nothing shipped. That’s no longer true: ByteDance and Tencent each took delivery of roughly 10,000 H200s in recent weeks, the first real thaw. But volumes are running near 13% of the licensed ceiling, and the binding constraint is now Beijing rather than Washington, with companies steered to keep most units outside the mainland. Nvidia assumed no China data-centre compute revenue in its outlook, so this is unmodelled upside rather than a free option. It can still move expectations, mix and tone even when it isn’t in the guide.

Why this reaches past the ticker.

The Index Is Long This Whether It Wants to Be

Nvidia is worth $5.05 trillion and is the largest company in the S&P 500 at roughly 8% of the index, ahead of Apple near 7%. The top 10 names are more than a third of it. Anyone who owns a broad index fund has a position in this report and mostly doesn’t think of it that way.

That’s why a 5.4% implied move matters beyond the stock. It represents about $280 billion of market value, more than the entire market capitalisation of about 90% of S&P 500 members, and the correlation into everything else with AI exposure is high.

The macro backdrop is doing its own work. The 30-year Treasury yield touched 5.33% on August 18, its highest since 2007, on inflation that won’t settle, deficits near $2 trillion and foreign holders trimming. Long rates that high compress the present value of earnings arriving years out, and the cheap-looking version of Nvidia’s multiple is exactly that: about 22 times consensus for the fiscal year ending January 2027, falling to roughly 16 times the year after. The 97% growth figure is a single quarter’s year-over-year rate, so don’t set it against a full-year multiple. They measure different things.

What the price says.

The Setup

Nvidia sits below its 20-day and 50-day averages at $214.79 and $210.46, above the 100-day at $205.44 and well above the 200-day at $195.72, after 7 straight lower closes and a 7.5% slide from August 13. Sellers hold the directional readings, though ADX near 17 says no trend has established itself, so this reads as de-risking into an event rather than a breakdown. The faster stochastic-RSI gauge is at 1, about as oversold as that reading goes. Average true range is $6.40, roughly 3% of price, so the implied move is under 2 sessions of normal movement delivered at once.

A 5.4% move from Monday’s close is $197.22 or $219.74. The downside sits close to the 200-day average at $195.72 and the 50% retracement at $200.41. That’s a confluence worth noting, not a floor: the implied move describes expected magnitude, not where selling stops.

1. Accumulation: $195.50 to $201, holding the 200-day average and the 50% retracement. Invalidation is a close below $189, beneath the 60-day low at $189.80. From the middle of that zone you risk about $9 to reach $225, roughly 2.9 to 1.

2. Reclaim: a close back above $215, over the 20-day average. Objectives $225, then $232.28 at the 60-day high, then $236.54 at the 52-week high. The geometry is poor, about 0.8 to 1 to the last objective against the same stop, which is the arithmetic version of not chasing a gap up.

3. Not into the print. Implied volatility is elevated because an event premium is embedded in it, which is normal rather than a signal, and it typically collapses afterwards. Buying protection the day before is paying for the event at full price.

Bottom Line

The useful question isn’t whether Nvidia beats. It’s whether an October guide near $104 billion and a gross margin near 75% can both be true at once, because that pairing is what a company still setting the terms of its own supply chain looks like. The 4-quarter losing streak is evidence the market already thinks in those terms; it isn’t a reason on its own to expect a 5th. And something has changed this time: in February the stock went in near its highs, and it goes in now 7.5% off them, with options pricing a smaller move than before May and well under the 12-quarter average. Less optimism is embedded than the pattern implies. If you want to own it, the area where the implied downside and the 200-day average overlap is a more considered place to do that than $12 higher on Wednesday afternoon.

This is research and commentary, not personal investment advice. Levels are illustrative; size positions to your own risk tolerance and time horizon. The author may hold positions in names discussed.