Microsoft, Meta, Apple and Amazon Report Earnings July 29 and 30 as Investors Weigh AI Spending

Alphabet Beat Estimates. The Stock Fell Anyway.

Four of the biggest companies in the country report earnings this week. Microsoft and Meta go first, after the market closes Wednesday, July 29^[1][3]. Apple and Amazon follow Thursday, July 30^[2][3].

They're reporting into a market that just got rattled. Tech and chip stocks fell sharply during the week of July 20-24^[9][10][11]. The trigger wasn't bad news, exactly — it was Alphabet.

On July 22, Alphabet reported revenue of $119.8 billion, up 24% from a year earlier. Google Cloud revenue jumped 82%, to $24.8 billion^[6][7]. That's not a business nobody wants. Alphabet also raised its 2026 spending plan on data centers and chips to $195 billion-$205 billion, up from $180 billion-$190 billion^[6].

The stock fell anyway, and the wider selloff got worse the next day^[8][10]. Market coverage put Alphabet's July 23 drop at around 7.1%^[10]. That's the tension sitting under this whole week: the demand for AI computing looks real, and investors are getting nervous about it at the same time.

What the Spending Actually Hides

To see why a good quarter can still spook a stock, you need to understand capex — short for capital expenditure. It's money spent on things built to last for years: data centers, land, power hookups, chips.

Capex doesn't hit profit the way an ordinary expense does. The cash leaves the bank right away, but the cost gets spread out — "depreciated" — over however many years the equipment is expected to work. So a company can report record profit while its actual cash pile is shrinking fast, because the spending and the accounting move at different speeds.

That gap is the whole story this week. Amazon, Alphabet, Meta and Microsoft together plan to spend roughly $695 billion-$725 billion on this buildout in 2026. In 2025, the same four spent about $410 billion — an increase of about 77% in a single year^[3][4]. Free cash flow, the cash left after capex that actually funds stock buybacks and dividends, is getting squeezed across the group^[5].

There's a second number buried inside all this: how many years a chip is assumed to last. If a company books a chip's cost over six years, the annual hit to profit looks small. If the chip is really obsolete in three, the true yearly cost is double what's on the books. Nobody yet knows for certain which is closer to true, because AI-scale data centers haven't been running long enough to say^[5][11].

Building Too Little Looks Worse Than Building Too Much

Talk to the companies themselves, and the logic sounds less like faith and more like a bet on timing. A data center takes two to three years to build. If a company underbuilds and demand shows up anyway, the customer it turns away goes to a competitor — and often doesn't come back^[3][4]. Overbuilding just means a write-down later. Given that math, spending too much looks like the safer mistake.

Alphabet's cloud growth is the company's own evidence for this: 82% growth, they'd argue, is what a supply shortage looks like, not a demand question^[6]. Microsoft and Amazon sell cloud computing directly, so extra capacity is inventory sitting ready for a customer^[3][4]. Meta doesn't sell computing power; it needs AI to hold attention and sharpen ad targeting, which is where its money actually comes from^[15][16]. It just raised its own 2026 capex plan to $125 billion-$145 billion^[15][16].

Skeptical investors read the same numbers and see something closer to a bubble. Their strongest point is the depreciation question above: if these chips wear out faster than the companies assume, today's profits are flattered and a reckoning is coming^[5][11]. Their second point is simple arithmetic — AI-linked revenue isn't growing anywhere near as fast as the roughly 77% jump in spending^[3][4]. And they'd note that the market is already voting: Alphabet beat expectations on every headline number and the stock dropped anyway^[8][11]. Coverage further to the right, like The Epoch Times, has leaned into this as market discipline finally arriving — investors demanding "receipts" after years of open-ended spending^[11].

Apple's Bet on Doing Less

Apple reports the next day, July 30, and it's telling a different story entirely^[2]. Where its rivals are pouring tens of billions into data centers, Apple runs most AI processing on the device itself, backed by a comparatively small private cloud^[19]. Its pitch is that it doesn't need hyperscaler-scale infrastructure, because it sells the hardware people already hold in their hands.

That's also a financial choice. Every dollar Apple doesn't spend on data centers is a dollar it can hand back to shareholders through buybacks, which is exactly what its investor base rewards^[19]. Analysts expect Apple to report revenue of roughly $108.8 billion-$110 billion, with earnings around $1.89 a share and gross margins near 47.5%-48.5%^[19].

If the AI-spending worry deepens this week, Apple's restraint could start looking prudent instead of behind. The risk cuts the other way, too: if AI capability keeps compounding the way the hyperscalers are betting it will, a company spending far less on infrastructure could fall behind on the underlying technology.

A Cheaper Model Out of Shanghai

The selloff had a second trigger, and it came from China. On July 16-17, Moonshot AI released Kimi K3 at the World Artificial Intelligence Conference in Shanghai — a 2.8-trillion-parameter model released "open-weight," meaning the trained model itself is published for anyone to download and run on their own hardware, rather than paid for per query through a closed service^[12][13]. Moonshot says it approaches top U.S. models at far lower prices^[12].

The release hit chip stocks worldwide. Samsung and SK hynix fell alongside U.S. chipmakers, and one market tally attributed more than $3.3 trillion in lost global semiconductor value to the days following the launch — though that kind of single-event attribution is an estimate, and other things were moving markets the same week^[13][14].

There's a structural argument underneath the release, not just a marketing one. U.S. export controls limited Chinese developers' access to top-end chips, and Chinese labs say that scarcity pushed them toward training their models more efficiently instead of just buying more hardware^[12][13]. Moonshot is also reportedly preparing a Hong Kong listing targeting a valuation above $30 billion, with some reports citing $50 billion, on annual revenue of about $300 million^[20]. Whether efficient training is a durable edge or a one-time workaround is still an open question. Either way, it reframes the U.S. buildout for some observers as a possible overspend rather than just a valuation debate.

Who Actually Pays for the Buildout

There's a third argument running alongside the first two, and it isn't about investors at all — it's about electric bills. U.S. electricity prices rose 5.9% in May 2026 from a year earlier, while overall inflation ran at 4.2%^[17]. PJM, the largest U.S. grid operator, tied roughly $6.3 billion of coming consumer cost increases over three years largely to data-center demand^[21].

The mechanism here matters. A regulated utility doesn't profit like a normal business selling a product — it earns a rate of return set by regulators on the capital it invests in the grid. That system exists because it's how a monopoly utility attracts the investors who fund power lines in the first place, and setting that return too low can raise borrowing costs that eventually land on customers anyway. But it also means new transmission lines built for a data center get folded into everyone's rates unless regulators carve out a separate deal — so the utility has every reason to welcome the new load, and the real fight is over who gets billed for it, not whether to build it at all^[17][18].

Utilities and some economists push back with a genuine counterpoint: big, steady industrial customers have historically spread fixed grid costs across more electricity sold, which holds rates down for everyone else. That logic only breaks if capacity gets built for demand that never shows up^[18]. Data centers' share of U.S. electricity use is projected to rise from about 4% in 2023 to more than 8% by 2030, and the effect lands hardest in the specific counties where the buildings actually sit^[17].

Coverage of this week has split along familiar lines depending on which of these threads a newsroom picks up. The Epoch Times centered the shareholder-discipline story and left ratepayers out of it almost entirely^[11]. AP's wire coverage, read here on PBS, led with the household electricity numbers^[17]. Seoul Economic Daily covered the week almost entirely through the lens of what Kimi K3 meant for Samsung and SK hynix^[13]. Bloomberg framed the burden of proof as falling on Big Tech's management to "justify" the spending^[9]. None of that changes the underlying numbers — it just decides which of them gets top billing.

What none of this week's earnings calls can actually settle is the one number the whole debate hinges on: how long a 2026 AI chip really keeps earning its keep before it's obsolete. Bulls and bears are both extrapolating from about three years of history^[5][11]. The real answer is still a few years out.