News vs. Noise: AI Spending Meets Skepticism

A megacap earnings beat still got punished, and that’s the tell on AI spending

Alphabet’s Q2 report beat estimates on revenue and cloud growth — Google Cloud revenue rose 82% [1] — but the stock still sold off more than 7% after the company raised 2026 capital spending guidance to $195–205 billion and posted its first negative free cash flow, the cash a company has left over after covering its operating expenses and capital investments, which it can use to pay down debt, return money to shareholders, or reinvest in the business [24], in 22 years as a public company [2, 3]. A beat that gets sold is a signal in its own right: investors have stopped rewarding AI spending simply for existing.

That reaction lines up with what’s happening underneath, in the debt markets that fund this buildout. AI-infrastructure borrowers have issued roughly 14 times the bonds they did in 2024, leverage ratios for this group (a measure of how much of a company’s financing comes from debt rather than equity [26]) have roughly doubled in six months, and credit spreads on their debt (the extra yield investors demand to hold a company’s bonds over safer Treasury debt of the same maturity [25]) have widened as bond investors start pricing them like capital-intensive infrastructure companies rather than high-margin software businesses [4]. This is becoming a financing story as much as a growth story.

We’ll get a fast, direct test of whether this was a one-off or the start of a pattern: Microsoft and Meta report Wednesday, Apple and Amazon report Thursday, and chipmakers Qualcomm and SK Hynix also report this week [5]. Watch whether the market keeps drawing the same line between spending and returns, or whether Alphabet’s reaction was the outlier.

China’s grip on a metal nobody watches is becoming a real chip-supply problem

China’s tungsten export curbs — exports are down roughly 50% year-over-year — are squeezing the supply of tungsten hexafluoride (WF6), the gas used to form the tungsten “plugs” that wire together the microscopic layers inside every advanced logic and memory chip [6, 7, 9]. Tungsten is used here specifically because of its high melting point and its ability to fill impossibly small spaces evenly; there’s no simple substitute sitting on the shelf [10].

Japanese and Korean WF6 producers, who supply most of the chip industry outside China, are already warning customers of supply gaps and have raised prices as much as 70–90% [8]. That matters because chipmakers need 12 to 24 months to qualify any new material before they’ll trust it in production [9]. Even if an alternative source were identified today, this doesn’t resolve before 2028 at the earliest — it’s a genuine multi-year constraint, not a headline that fades with the next trade announcement.

We’re treating this as a structural, not a tradeable, risk for now: something that raises the cost and lengthens the timeline of the broader chip and AI-hardware buildout over the next several years, worth watching alongside — not instead of — the financing question above.

An open, near-frontier AI model just undercut the idea that model-makers keep the pricing power

Last week, Chinese AI lab Moonshot released Kimi K3, a 2.8-trillion-parameter open-weight model — meaning its code and weights are published for anyone to use or modify — that performs near the level of leading closed models at a fraction of the cost to run [11, 12]. That’s a direct hit to an assumption a lot of AI valuations have been built on: that the labs building the best models keep durable pricing power over the businesses built on top of them.

If a near-frontier model is available cheaply and openly, the model itself stops being the scarce, high-margin asset. Value shifts down the stack, toward the compute and inference infrastructure needed to actually run these models at scale, and toward the memory and specialized chips that infrastructure depends on [11].

This is the same repricing question the market is already working through elsewhere in AI: the durable winners may increasingly be the picks-and-shovels layer, not the labs. We’re watching whether other frontier labs respond by cutting prices, restricting access, or leaning harder into proprietary data and distribution advantages that a published model can’t copy.

An impressive early “beat rate” is hiding how narrow this earnings season still is

With only a few dozen S&P 500 companies reported so far, the season’s beat rate is running around 86–88%, and headlines are calling it one of the strongest starts in years [13, 14]. That number is small-sample and lopsided.

Strip out Alphabet alone and the season’s blended earnings growth rate (a figure that combines actual reported results with analyst estimates for the companies that haven’t reported yet [13]) drops from roughly 38% to 26% [15]: one company is doing an outsized share of the lifting, and that same company’s stock got sold off on its own results (see above). A beat rate this early tells you who has reported, not how the market broadly is doing.

We’re not treating a strong early beat rate as confirmation that the season, or the market, is broadly healthy until more of the index, and more sectors, have actually reported.

One earnings beat that popped and faded isn’t evidence the selloff is over

Intel’s Q2 results, out Thursday, beat badly: revenue of $16.1 billion against expectations of $14.4 billion, with earnings per share roughly double what analysts expected, driven by a 59% surge in its data-center and AI business [16, 17]. The stock jumped as much as 12–13% within minutes of the release, then gave back more than half that gain by the next morning [18].

That round-trip is the tell. A single company’s beat, even a strong one, doesn’t validate a market-wide call that a broader tech and momentum-stock selloff has bottomed, especially when the stock’s own reaction couldn’t hold overnight. Commentators reaching for “the worst is over” off one earnings report and a couple of green trading days are reading a lot into a thin sample.

We’re not repositioning around a headline earnings beat that the market itself couldn’t sustain for twelve hours. One data point is not a trend.

“Strong earnings can offset higher yields” is a mismatch of two different levers

The 30-year Treasury yield closed above 5% for its longest stretch since 2007 last week, after a July bond auction, the periodic sale through which the U.S. Treasury issues new debt to investors and sets its yield [27], already drew the weakest demand for 30-year debt in roughly two decades [19, 20, 21]. That’s the backdrop against which the idea that strong corporate earnings can simply offset higher long-term yields has been circulating [22].

That argument doesn’t hold up mechanically. A company’s value is the sum of its future profits discounted back to today’s dollars, and the discount rate used for that math is tied directly to long-term interest rates. When yields rise, the same future earnings are worth less today, regardless of whether those earnings are actually growing. A good quarter confirms demand is real; it does nothing to reset the rate the market is using to value tomorrow’s profits [22]. Those are two separate levers, and earnings don’t substitute for the other one.

We’re treating the level and duration of long-term yields, not any single earnings print, as the more important signal for how much investors should be willing to pay for future growth right now.

Beijing’s emergency buying propped up Chinese tech stocks for a day, and that’s not a demand story

Last week, Chinese authorities rolled out broad state support (including record inflows into tech-focused funds) specifically to halt a slide in Chinese technology shares [23]. It worked, at least in the near term: the selloff paused and shares bounced.

State-directed buying can arrest a selloff on command, but it doesn’t reflect a change in underlying demand, earnings, or the regulatory and financing pressures still facing Chinese technology companies. It’s a policy backstop, not a fundamentals story, and policy backstops don’t compound the way organic demand does — they hold a line until the next test.

We’re not reading a government-engineered bounce as confirmation that the underlying risk has eased. If anything, the fact that intervention was needed at all is a reminder of how fragile sentiment there still is.