JUICED EARNINGS: WHAT'S INSIDE BIG TECH'S RECORD PROFITS

Price-to-earnings ratios, revenue growth rates, operating margins, free cash flow, return on invested capital: these are among the key biometrics that let investors understand the relative health of a business.

Earnings are another, generally reliable, way to check up on a company's vital signs: how its products and services, such as cloud computing contracts, digital advertising, and semiconductor sales, are translating into actual revenue. Investors use this information to decide whether to buy, hold, or sell a stock, and to judge what price is fair to pay for a claim on future profits. High earnings signal that the underlying business is performing well; low earnings signal that it is not, or that management is spending heavily today in pursuit of growth tomorrow. Simple, right? But beyond this single, monolithic metric, however, the process of determining the figure can get considerably more complicated. In the age of artificial intelligence and increasingly creative financial engineering, it is becoming harder to know how ostensible versus useful these numbers actually are.

Big Tech's second-quarter 2026 results were, by the headline numbers, extraordinary: Alphabet reported net income of $112.1 billion; Amazon reported net income of $62.6 billion, up 245% year over year. However, read past the top line and a large share of that profit did not come from selling search ads, cloud capacity, or anything else to a paying customer; it came from marking up the value of stakes these companies hold in private artificial intelligence companies, chiefly Anthropic. That is one of three mechanisms currently doing quiet, substantial work inside “record” tech earnings, alongside circular financing arrangements among AI infrastructure players and increasingly aggressive depreciation assumptions on the chips underlying the buildout.

None of these mechanisms is improper on its own. All three are worth understanding before treating this quarter's earnings per share as a clean read on demand.

Alphabet's own Q2 2026 earnings release discloses a $99.0 billion pre-tax gain on equity securities, a figure that, net of tax, lifted net income by $77.1 billion and added $6.26 to diluted earnings per share. Put another way: the house made more money on its side bet than on the casino. Alphabet does not break the gain out by individual holding, but the quarter coincides with a period in which Anthropic (in which Alphabet has invested) raised capital at a sharply higher valuation.

Amazon's disclosure is more direct: its own earnings release states that “second quarter 2026 net income includes non-operating pre-tax other income of $53.4 billion, primarily from our investments in Anthropic,” against operating income of just $27.5 billion for the quarter. In both cases, the accounting is standard and permitted for minority equity stakes in private companies: no cash changed hands, and no rule was bent. But it means two very different kinds of “profit” now sit inside the same net income line: profit from selling something alongside profit from someone else's private funding round. The two behave very differently when markets turn, and coverage of “blowout” tech earnings rarely separates them for the reader.

The same distortion is not confined to these two companies’ income statements; it shows up in the S&P 500’s own aggregate growth rate. First-quarter 2026 index earnings grew a reported ~28% year over year; stripping out an estimated $69 billion in combined unrealized gains at Alphabet, Amazon, and Nvidia brings that down to roughly 16%, as cited by Acadian Asset Management. The pattern intensified in the second quarter, when the index’s blended earnings growth as of the writing of this article reached 50.4% year over year but fell to 32.0% excluding Alphabet and Amazon alone; per FactSet Earnings Insight, the two companies, combined, account for “about 71% of the increase in dollar-level earnings for the index since June 30.” In dollar terms, that is roughly 43% of the Q1 growth rate and 37% of the Q2 growth rate riding on equity markups rather than on operating results across the rest of the index.

A second, related mechanism runs through the AI infrastructure buildout itself. Nvidia agreed in 2025 to invest up to $100 billion in OpenAI as part of a joint plan to deploy 10 gigawatts of Nvidia systems, infrastructure that OpenAI is, in significant part, buying from Nvidia. That $100 billion figure did not hold. Nvidia’s own November 2025 disclosures said there was no assurance a definitive agreement would follow, and by February 2026 Jensen Huang said the commitment had never been binding to begin with. What closed instead, on March 31, 2026, was a $30 billion Nvidia stake in a $122 billion OpenAI funding round (announced at up to $110 billion in February 2026, before closing larger), a stake the companies describe as unconnected to the 10-gigawatt deployment plan. By mid-2026, reporting indicated Nvidia was weighing additional financing arrangements, including a possible guarantee on roughly $250 billion of debt tied to an OpenAI data center project, on top of financing OpenAI's chip purchases directly. Nvidia's own credit default swaps widened sharply on the news, a sign that some credit investors see concentration risk in a structure where the seller is also underwriting the buyer's ability to pay. That $250 billion figure also proved provisional: when Nvidia and OpenAI signed a final agreement in August 2026 on the project, an Ohio data center, capping Nvidia’s disclosed payment obligation at $105 billion. Given how often the reported figures on this buildout have moved, both should be read as reflecting the most recent public disclosures rather than a final accounting, and are subject to further revision as the parties disclose more detail.

None of this makes the resulting revenue fake; Nvidia's chip sales are real, GAAP-compliant revenue. It does mean a portion of the demand behind that revenue is financed, directly or indirectly, by the seller itself, a different quality of demand than an independent customer buying because it needs the product.

The same dependency shows up on the hyperscalers’ own revenue lines. An independent analysis estimates that roughly 70-75% of AI-related revenue at Amazon, Microsoft, and Google, and more than 60% of the AI backlog hyperscalers point to in justifying their capital spending, traces back to just two customers: OpenAI and Anthropic. Google’s dealmaking shows a version of the same pattern; in April 2026, Google agreed to invest up to $40 billion in Anthropic, and, according to reporting by The Information, Anthropic separately committed to spend $200 billion on Google’s cloud infrastructure and chips. Google’s own second-quarter 2026 earnings call put Cloud’s backlog at $514 billion, up more than $50 billion for the quarter, though the company attributed it to “a broad mix of customers” and did not mention Anthropic by name; the analysis cited above nonetheless treats relationships like this one as a significant driver of that growth.

Nvidia’s own numbers show how concentrated that exposure has become on the supplier side. Hyperscalers, who buy chips in part to serve OpenAI and Anthropic (the latter a company in which Nvidia has separately invested $10 billion), reported $49.0 billion of Nvidia’s $89.0 billion in data center revenue last quarter, or just over half, per Nvidia’s own August 2026 earnings call. Nvidia has also backed several neocloud, or GPU cloud, providers directly, agreeing in at least one case to buy back unused capacity if the provider cannot find customers for it. An independent analysis puts total neocloud exposure at 30% or more of Nvidia’s revenue, which would put over 70% of Nvidia’s revenue as coming from customers it has itself invested in. The risk that follows is straightforward; if one of those customers cannot pay, Nvidia loses the sale and the investment at the same time.

The third mechanism is more technical, but no smaller in scale: how long a hyperscaler assumes a GPU or server will remain useful for purposes of spreading its cost across the income statement. Meta has extended its assumed server life from four years in 2022 to five and a half years currently, a change reported to be worth roughly $2.9 billion in reduced depreciation expense in 2025 alone. In November 2025, investor Michael Burry publicly argued, in a post that CNBC reported on but said it could not independently verify, that Oracle and Meta were overstating cumulative 2026-2028 earnings by roughly 27% and 21%, respectively, by extending useful-life assumptions on hardware he believes should be depreciated over two to three years given Nvidia's roughly annual chip-release cycle. He put the industry-wide understatement of depreciation at about $176 billion over that period. The hyperscalers' counterargument, echoed by research firms including Yardeni Research, is that old chips don't go to waste; once they're too slow for training new AI models, they get shifted to easier jobs and keep earning for years, the same pattern that made earlier rounds of server-life extension look reasonable in hindsight.

Multiples are a bet on repeatability. The premium currently being paid for AI-linked earnings assumes this quarter's numbers are a clean, repeatable signal of demand, an assumption the companies themselves are doubling down on; Alphabet and Meta have each raised 2026 capital spending guidance more than once this year, Alphabet to $195-205 billion, up from $180-190 billion, and Meta to a $130-145 billion range. That is a lot of capital riding on a demand signal whose composition is growing more dependent on private-market valuation marks, vendor-financed demand, and depreciation assumptions that credible analysts dispute by tens of billions of dollars.

Price-to-earnings, the multiple most commonly quoted, is exactly the ratio these gains flow into and one reason the markups have not shown up as an obvious valuation warning sign. The S&P 500’s forward twelve-month P/E stood at 20.0 as of August 7, 2026, only modestly above its five- and ten-year averages of 19.9 and 19.0. On that measure alone, the index looks only mildly rich relative to its own recent history.

Price-to-free-cash-flow, a multiple that is harder to move with a one-time unrealized mark, tells a more pointed story. On Crewcial’s internal analysis of data provided by CapIQ of S&P 500 (ex-Financials) constituent free cash flow and price, that multiple has climbed from the mid-20s three years ago (25.6x in the third quarter of 2023, 27.7x in the fourth) to 38.0x in the second quarter of 2026, its highest level in the series and a reading that predates results reported since this note was drafted. Viewed through free cash flow rather than earnings, the index is priced meaningfully richer than it was three years ago, even where the P/E has stayed close to its own historical range.

Ultimately, we would treat this quarter's headline earnings-per-share figures as a starting point for further diligence. For an institutional portfolio, the practical takeaway is to ask AI-exposed managers a simpler question: how much of their thesis depends on GAAP earnings continuing to grow at this pace, versus operating cash flow and contracted backlog that would hold up on their own. It is worth asking what price that thesis carries, too. Measured in P/E, the index looks fine; measured in free cash flow, it does not. The earnings are real; whether they are repeatable is a separate question, one the market has not yet had to answer.