Apple passed Nvidia on Friday. By the closing bell it hadn't, and the correction is the story.

The headline that travelled furthest on Friday was that Apple had reclaimed its title as the world's most valuable company. It was true for part of an afternoon. Apple traded at roughly $4.88 trillion while Nvidia, down as much as 3.9 percent shortly after the open, briefly slipped to about $4.82 trillion — and for a window of a couple of hours the ranking that everyone tracks and nobody can act on had changed hands. Then it changed back. Nvidia recovered through the session to close down about 2.2 percent, a little above $4.9 trillion, ahead of Apple by something on the order of six billion dollars. On a $4.9 trillion base that gap is roughly twelve basis points. It is, in any honest accounting, a tie.

I am not raising this to score a point against the coverage, though a fair amount of it described a mid-session snapshot as a closing fact and never went back to correct it. I am raising it because the ranking was always the least informative thing in the story, and the scramble to report it consumed attention that belonged to the number underneath. Two companies swapping the top slot for ninety minutes is trivia. What produced the swap is not. Nvidia lost that ground because the entire semiconductor complex was being sold, hard, in the worst week the sector has had in more than a year — and Apple gained it by standing still.

What actually happened to the money

Take the week rather than the day. The Philadelphia Semiconductor Index fell about 10 percent across the five sessions to 17 July, its steepest weekly decline since April 2025. By Friday's close it had entered a bear market, down 20.2 percent from the record close it set on 22 June — four weeks, top to bear, which is fast even by the standards of a sector that has always traded like a cyclical wearing a growth-stock costume. The Nasdaq 100 lost 4.1 percent on the week and the S&P 500 gave up 1.6 percent, which tells you how concentrated the damage was: the index barely moved relative to the thing inside it that was breaking. Across global semiconductor names, something in the region of $3.3 trillion of market value has come off since that June peak — a reported figure, and one I would treat as an estimate rather than a measurement, but directionally consistent with everything else.

Now the context that almost every account of the selloff omitted, and that anyone actually positioned in this sector would put first. The same index is still up more than 60 percent year to date, against roughly 9 percent for the S&P 500. It had risen about 105 percent from its March low to that June high. A 20 percent drawdown from a record set four weeks ago, in an index that has doubled inside five months, is not a crash. It is the first serious repricing of a position that had become extremely crowded, and the distinction matters enormously for what you do next — because a crash invites you to buy the dip, and a repricing invites you to ask what the thing is worth.

Friday's specific catalysts were mundane and cumulative rather than singular. TSMC fell about 7.3 percent and Netflix dropped somewhere between 7 and 9 percent, both on earnings that the market decided it did not like. The selling started overseas and arrived pre-opened: the KOSPI fell 6.4 percent, the Nikkei 4 percent, the CSI 300 3.6 percent. The US majors closed down between 0.8 and 1.4 percent, with the Nasdaq's 1.4 percent the worst of them. And running underneath all of it was a story from the previous fortnight that the market has been slowly digesting: a Chinese lab shipping a frontier-grade open-weight model, which invites the one question the entire trade is built on not asking — whether the capability everyone is spending hundreds of billions to buy can be had for meaningfully less.

The three events that actually moved the thesis

Strip out the noise and three discrete events in the past three weeks did real damage to the AI trade's central premise, which is that compute is permanently scarce and therefore permanently priced at a premium.

1. 1 July — Bloomberg reported that Meta is building a business to sell its surplus AI compute to outside customers. Meta rose 8.8 percent. Micron fell 10.6 percent, AMD 6.9 percent, Nvidia 1.25 percent. Read the divergence: the market rewarded the seller and punished the suppliers, because a hyperscaler with capacity to spare is evidence against scarcity.
2. 16 July — Google's flagship Gemini 3.5 Pro was reported to be months late after missing internal coding targets. Alphabet closed down about 4.4 percent, erasing roughly $200 billion. The read-through was not about Google. It was that spending does not reliably convert into capability on schedule.
3. Through July — SK Hynix was reported in trade press to be slowing its HBM4 line conversion in favour of higher-margin DDR5. I have not found this confirmed by a wire service and it should be held loosely. But if a memory maker is reallocating capacity away from the AI-specific product, that is a supplier expressing a view about demand.

Each of those is survivable alone. Together they describe a market that spent two years pricing compute as a permanently scarce good and is now, tentatively, testing whether it is merely an expensive one. That is a repricing of a premise, not of a quarter, which is why it has been indiscriminate about which chip names it hits.

The market spent two years pricing compute as permanently scarce. It is now testing whether it is merely expensive. That is a repricing of a premise, not of a quarter.

Why the money went to Apple

Apple did not rally into the top slot. It closed Friday down about 0.1 percent — flat, in a week when the things around it fell double digits. The rotation into Apple is less a bet on Apple than a preference for the one mega-cap in this cohort that is not spending like the others. As one UK wealth manager put it in the coverage, Apple is simply less exposed to capex intensity. That is the whole trade, stated plainly.

The fundamentals underneath are genuinely strong and worth stating without spin. Apple's March quarter delivered revenue of $111.2 billion, up 17 percent, with diluted earnings per share of $2.01, up 22 percent, on a March-quarter iPhone record and more than $28 billion of operating cash flow. The December quarter before it was $143.8 billion, up 16 percent, the best iPhone quarter the company has ever had. This is not a defensive stock being bought for the absence of a story. It is a business compounding in the high teens while generating cash rather than consuming it — which, in the current cohort, has become an unusual property.

There is an irony worth marking. Apple touched the top of the world's equity markets during Tim Cook's final weeks in the job. The company announced in April that Cook will hand the chief executive role to John Ternus on 1 September, moving to executive chairman. Whatever the ranking said on Friday afternoon, the handover happens with Apple at or near the largest market capitalisation any company has ever held.

The number that gets tested this week

Here is the figure I would put above all the others, and it has nothing to do with who is briefly worth the most. Across Amazon, Microsoft, Alphabet and Meta, capital expenditure for 2026 is tracking toward roughly $725 billion on compiled estimates, up about 77 percent on last year's $410 billion. Amazon is around $200 billion, Microsoft around $190 billion, Alphabet between $175 and $185 billion, and Meta raised its own guidance in April to a range of $125 to $145 billion. Roughly three-quarters of that total is AI-related.

Now the ratio that actually describes the exposure. Capital spending across this group has reached approximately 100 percent of operating cash flow, against a ten-year average nearer 40 percent. Read that as a sentence about position rather than accounting: these companies are now spending, on physical infrastructure, essentially everything the business generates. The analysis from Epoch AI — one source, so weight it accordingly — puts capex growing around 70 percent a year against operating cash flow growing about 23 percent, with the two curves crossing for the group around the third quarter of this year, and staggered by company: Oracle already through it, Amazon crossing now, Alphabet around the first quarter of 2027, Meta around the third, Microsoft not until late 2028.

What that means concretely is that the marginal dollar of this build-out has stopped coming out of profits and started coming from somewhere else — the balance sheet, the debt markets, the bond desks that have spent the past year absorbing an extraordinary volume of AI-linked issuance. A company funding capital expenditure from cash flow can slow down whenever it likes. A company funding it from debt has committed to a schedule, and schedules are what turn a soft patch into a forced decision.

This is precisely what the market is now demanding be justified, and the demand arrives with a calendar attached. Bloomberg's framing on Sunday was that pressure is building on the biggest AI spenders to justify their expenditures to “beleaguered traders with increasingly itchy fingers hovering over their sell buttons.” Alphabet is scheduled to report on Wednesday, Microsoft and Meta a week later on the 29th, Apple and Amazon on the 30th, with Nvidia not until late August. Four of the five biggest spenders in the world report inside nine days, into a tape that has just taken 20 percent out of the sector selling them the equipment.

What breaks if the thesis is wrong

So state the downside, since nobody else seems keen to. The bull case requires that this capex converts into revenue on a timeline the debt schedules can tolerate. The bear case does not require a collapse in AI demand — that is the mistake in most of the pessimistic commentary. It only requires that conversion take longer than the financing assumed. At roughly 100 percent of operating cash flow, there is no slack left to absorb a delay. A year of slippage does not dent earnings; it forces a choice between cutting the build-out, which validates every bear, and funding it further from credit markets that will by then be repricing the risk.

And the second-order exposure is broader than the sector. Torsten Sløk at Apollo has made the point that an AI slowdown would risk tipping the wider economy into recession — which follows arithmetically once a single category of corporate investment gets large enough relative to everything else. This has stopped being a sector position. It is a macro one, held by index funds on behalf of people who have never knowingly expressed a view on inference costs.

Which brings me back to Apple, and to why I would be careful about the rotation even as I understand it. Apple is being bought as the AI mega-cap that is not exposed to AI capital spending. That is half right. It is not exposed on the spending side. It is thoroughly exposed on the cost side, and Cook has said so himself in unusually plain language: in a June interview he described component price increases as “unavoidable,” called the memory situation “unsustainable,” and characterised it as a “hundred-year flood” he had not seen the like of in more than forty years in the industry. Independent teardown estimates have the memory bill of materials on a flagship iPhone rising from roughly $50 to something nearer $200 across a single generation. The same AI build-out that is inflating everyone else's capital budget is inflating Apple's cost of goods sold, and it arrives at the consumer as a price rise into a demand curve nobody has tested at that level.

Nor is Apple independent of the compute it is not buying. Its rebuilt Siri is reported — and I stress reported, since neither company has confirmed it — to run on a custom Google model at a cost in the region of a billion dollars a year. Apple's apparent immunity to the capex cycle is, in part, an accounting artifact of renting the capability instead of building it. That is a defensible strategy and possibly a brilliant one. It is not an absence of exposure. It is a different exposure, held off balance sheet, to a supplier that is itself one of the four companies whose spending the market has just started to question.

So the ranking flipped for ninety minutes on Friday and flipped back, and by Monday morning the correction had not travelled a fraction as far as the error. Fine — the ranking never mattered. What matters is that a sector priced for permanent scarcity has just been repriced by 20 percent in four weeks, that the four largest spenders in the world are about to explain themselves inside nine days, and that they will be doing it while committing roughly a hundred percent of their operating cash flow to an asset base whose payback schedule is still, in every meaningful sense, a forecast. The trade is not that Apple is worth more than Nvidia. The trade is that the market has stopped taking the build-out on faith. Wednesday is the first time it gets to ask out loud.