On Wednesday evening, minutes after the closing bell in New York, two of the most valuable companies on earth reported quarterly results. Both were, at bottom, reporting on the same thing: spending on artificial intelligence that has grown so large it now visibly deforms the income statement. Both confirmed that the spending is still rising. Both said, in their own words, that demand justifies it.
By Thursday morning Microsoft was up almost nine percent in premarket trading and Meta was down more than nine.
That is not a mood swing, and it is not thin-tape noise. It is the first clean separation the market has made in this entire capital cycle. For two years, AI capex has been priced as one single trade — sometimes celebrated, sometimes punished, but always as a block. On Wednesday night the block came apart. The question is no longer how much a company spends. The question is whether the spending has a meter attached.
Two sets of numbers, one subject, opposite signs
Neither company walked into the print with the wind at its back. The Nasdaq-100 had fallen 1.8 percent on Wednesday to close more than eleven percent below its June high — formally, a correction. The semiconductor selloff ran into a fourth day: Sandisk lost 14 percent, while AMD, ARM, Micron and Seagate each dropped more than eight, triggered in part by SK Hynix reporting a sharply higher quarterly profit that was nonetheless smaller than Wall Street wanted. The S&P 500 fell 1.5 percent and the Dow Jones Industrial Average 2.2 percent, roughly 1,100 points.
Meta arrived carrying something heavier. The stock had fallen for nine consecutive sessions — the longest daily losing streak in the company’s history as a public company. From a July intraday high of $681.90 it had slid to $593.41, down nearly twelve percent and about $223 billion in market value, before a single figure was released. The market had, in effect, already rendered its verdict and was waiting only for confirmation.
It got the confirmation. And in the same minute, it got the control experiment from Microsoft.
Microsoft: the number that mattered was not revenue
Microsoft closed out a fiscal year that can fairly be called exceptional. Fourth-quarter revenue rose 18 percent to $90.0 billion, net income 31 percent to $35.8 billion, and diluted earnings per share 32 percent to $4.81. Operating income came in at $40.6 billion. Across the full fiscal year that added up to $331.8 billion of revenue and $133.7 billion of profit.
Azure made the headline. The cloud business grew 43 percent in the quarter — accelerating from 40 percent in the prior one — and crossed $100 billion of annual revenue for the first time, up 41 percent for the year. CFO Amy Hood guided to 45 percent growth in the current quarter in constant currency, comfortably above the roughly 41.4 percent consensus. Microsoft Cloud overall delivered $59.3 billion in the quarter and $214 billion for the year.
The number that actually mattered sat further back in the deck. Commercial remaining performance obligations — contracted revenue that has been booked but not yet recognized — stood at $678 billion, up 84 percent. Alongside it: more than 30 million paid Microsoft 365 Copilot seats and 50 million GitHub Copilot users.
That is the meter. Microsoft is building data centers that produce something it can invoice, and it disclosed an order book roughly twice the size of its annual revenue. Capital expenditures including finance leases hit $41 billion in the quarter, up 69 percent year over year — but they sat next to a figure that explained what they are for.
Meta: free cash flow down 91 percent
Meta’s advertising business, to be clear, was not the problem. Revenue rose 28 percent to $60.8 billion, beating the $60.17 billion consensus, and the company guided to $61 billion to $64 billion for the third quarter. Some 3.60 billion people use at least one of its apps daily.
Below that line it got ugly. Total costs and expenses rose 55 percent to $42.0 billion — twice the rate of revenue growth. Operating income fell eight percent to $18.8 billion and the operating margin dropped from 43 percent to 31 percent. Net income declined 14 percent to $15.8 billion and diluted earnings per share 13 percent to $6.18, against a consensus of $7.22.
But the figure that actually moved the stock was elsewhere. Free cash flow collapsed 91 percent — from $8.55 billion in the year-ago quarter to $784 million. Operating cash flow of $31.86 billion was almost entirely consumed by $31.08 billion of capital expenditure. And the spending is still climbing: full-year capex guidance was raised at the bottom end, from $125–145 billion to $130–145 billion, while the expense range moved from $162–169 billion to $165–169 billion.
Fairness requires noting that the $42 billion cost base included $2.4 billion of legal charges and $1.18 billion of severance and restructuring. On the call, CFO Susan Li pointed out that excluding those two items, operating income would have risen nine percent. That is a legitimate objection — and it leads directly to the least comfortable part of this story.
The catch: part of the gap is an accounting assumption
In the same breath, Microsoft lowered its capital expenditure outlook for calendar 2026 from roughly $190 billion to roughly $175 billion. It sounded like discipline — like a company applying the brakes that Meta refuses to touch.
It was not a brake. Amy Hood explained that from fiscal 2027 the company is extending the estimated useful life of its data centers and office buildings from 15 years to 25 years, and will classify more future data-center leases as operating rather than finance leases. Finance leases count inside the capital expenditure line; operating leases do not. Outside those two effects, Hood said, the calendar 2026 expectation is unchanged. The reported number therefore drops by $15 billion without a single dollar of reduced spending. For the current quarter Microsoft guided to more than $50 billion of capital expenditure — the company is raising its build plans and citing demand as the reason.
Which leaves an uncomfortable finding on the table. On Wednesday night, the market rewarded a company that spread its depreciation across more years and punished one that took legal and severance charges into the current quarter. A longer useful life reduces annual depreciation and flatters earnings — not once, but for years.
This is not an accusation. The assumption may well be correct: building shells, land and grid interconnects plausibly last longer than fifteen years. It may be equally wrong for the graphics processors inside them. The point is narrower and harder to dismiss: nobody voted on that assumption, and it is the most consequential unverifiable number in the entire earnings season. Anyone attributing the whole gap between these two stocks to business quality is underestimating how much of it came out of a footnote.
The discount rate got more expensive the same evening
Hours before the results, the Federal Reserve had left its policy rate unchanged at 3.50 to 3.75 percent for a fifth consecutive meeting. The interesting part was not the decision but the tally: nine to three. Beth Hammack of Cleveland, Neel Kashkari of Minneapolis and Lorie Logan of Dallas all dissented in favor of an immediate quarter-point hike. Not since September 2016 had three voters dissented in the same direction.
Chair Kevin Warsh, who has dispensed with his institution’s forward guidance, remarked that he had asked for a good family fight and had gotten one. On substance he was blunter: there is no soft inflation target and no soft implicit target, he said, there is only a target, and it is two percent. September is now firmly a live meeting.
The bond market drew its conclusion immediately. The 30-year Treasury yield jumped 10.5 basis points to 5.201 percent and touched 5.244 percent intraday — the highest level since July 2007. The 10-year yield rose five basis points to 4.657 percent.
Long duration meets long rates
This is where the two strands meet, and it is why Wednesday amounted to more than one good report and one bad one.
A 25-year useful life is, at its core, a statement about duration. It says: this asset pays back over a quarter of a century. At precisely the moment technology companies are lengthening the payback period on their investments, the capital market is demanding more yield at the long end than at any point in 19 years. The cash flows move further out and the rate used to discount them moves up. For valuations, those two things happening together is the worst available combination.
Which is why the distinction the market is now drawing is economically sensible rather than merely nervous. When capital is expensive and committed for a long time, evidence that somebody at the end of the wire is paying an invoice matters more. A $678 billion contracted backlog is that evidence. An advertising business growing 28 percent while costs grow 55 percent is not — at least not yet.
What this means for a US portfolio
The most useful exercise this morning is to sort the AI complex by whether the meter exists, not by how big the spending is. Oracle is the obvious test case: it has reported an enormous contracted backlog of its own, but a much larger share of it is concentrated in a handful of counterparties whose own funding is unresolved. A backlog is only as good as the balance sheet standing behind it — the same question raised by the circular financing arrangements now common across the sector, in which the chip supplier effectively underwrites the customer’s obligations.
The same lens applies down the stack. Nvidia and Broadcom sell into the build; Micron, AMD, ARM and Seagate just took four days of indiscriminate punishment following one memory-maker’s earnings miss, which is exactly the kind of selloff worth separating by actual exposure. Amazon reports tonight and faces the identical question with AWS. Apple, which spends almost nothing relative to its size, is quietly becoming the control group for this entire debate.
The rate strand deserves its own line in the portfolio. A 30-year yield above 5.2 percent is a direct headwind for homebuilders, REITs and long-duration growth, and a tailwind for the net interest margin at regional banks — the KRE complex has been trading as the mirror image of the Nasdaq-100 for weeks now. On the tax side, one point is worth remembering as positions get re-examined this week: the difference between short-term and long-term capital gains rates is a real, quantified cost of acting on a single night’s news. Selling a winner eleven months in converts a preferential rate into an ordinary-income rate. Unlike a useful-life estimate, that number is not subject to management judgment.
The case against this reading
Four objections deserve serious weight before anyone turns one evening into a regime.
First: Meta’s advertising business is accelerating. Twenty-eight percent growth at that revenue base is remarkable, and third-quarter guidance up to $64 billion does not suggest erosion. What is being punished is capital allocation, not operations. Those two things can separate again — they have separated before at this very company.
Second: nine down days before the print means a substantial share of the move was positioning rather than fundamentals. Anyone selling after the ninth red candle is selling to a buyer who already receives a twelve percent discount.
Third: part of Azure’s acceleration is tied to OpenAI-related bookings — Microsoft explicitly disclosed commercial bookings growth excluding that effect, which tells you the question is live. A $678 billion backlog is contracted revenue, not earned revenue, and the creditworthiness of some counterparties in this industry is itself unresolved. Praising the meter is only half the job; the other half is asking who reads it and who ultimately pays.
Fourth: the useful-life assumption is open, not wrong. These assets may genuinely last 25 years, in which case the old schedule was too conservative and the new number is the more honest one.
What gets decided today
Thursday delivers both the numerator and the denominator. At 8:30 a.m. Eastern the June core PCE deflator arrives together with the first estimate of second-quarter GDP. The headline index is expected to fall 0.1 percent month over month, leaving the annual rate at 3.7 percent, with the core rate up 0.2 percent for an annual 3.3 percent. It is the Fed’s preferred inflation gauge — and it lands one day after the Fed’s decision.
After the close, Apple and Amazon report. For Amazon the test repeats exactly: AWS is the third major cloud that has to show its investment produces something billable. For Apple the question inverts, because the company invests almost nothing relative to its scale — a model that is suddenly finding admirers again.
Premarket, futures recovered: S&P 500 up 0.4 percent, Nasdaq-100 up 0.7 percent, Dow up 0.2 percent. The 30-year yield held near 5.24 percent at its multi-decade high and gold added almost one percent to $4,135. Overnight, US forces struck roughly a dozen Iranian targets after pausing two weeks of attacks last weekend; Brent rose 1.5 percent to $92.10 and WTI 0.9 percent to $85.23. In Europe, the DAX slipped 0.2 percent while the CAC 40 gained 0.6.
The practical takeaway from this night is smaller and more useful than the price swings suggest. For every company in this capital cycle, it is worth laying three numbers side by side: the contracted backlog, the free cash flow, and the assumed useful life. The first tells you whether anyone has ordered. The second tells you what is left after the building. The third tells you how long management believes the buildings will last — and it is the only one of the three that can be changed by a decision.
Try TradingView Free for 30 Days
Plus get a $15 discount on your first subscription through this link.
