TL;DR
The Mag 7's split is no longer a thesis — it's a three-tier cash flow regime, and this cycle the market started trading the tiers separately. Alphabet and Amazon printed record headline profits that were mostly paper marks on SpaceX and Anthropic stakes, while their actual operations bled cash to fund the AI buildout. Meanwhile Apple, the group's capital-light pole, floated its first serious move toward infrastructure in 15 years, and an AI-safety slowdown call from the labs themselves cracked the chip trade — sending chips, platforms, and software in three different directions in a single week.
The cash flow tiers are now the market's organizing principle
The June quarter hardened the group into three distinct free-cash-flow regimes, and the extremes are pulling further apart. Apple generated +$31.91B in FCF on just $2.46B of capex while Amazon posted the group's largest single-quarter capex at $54.21B and a −$8.82B FCF capital-allocation-and-free-cash-flow-divergence.
The new middle tier is the interesting one. Meta's capex consumed nearly all of its operating cash flow:
"Free cash flow collapsed to $784 million from $8.55 billion a year earlier." — capital-allocation-and-free-cash-flow-divergence
Meta's operating margin compressed to 31% from 43% as costs surged 55%, and its 2026 capex guidance of $130–145B now exceeds the cash it generates — a "mega cloud operator" bet that puts it on a collision path with the established clouds while its own margins contract capital-allocation-and-free-cash-flow-divergence. With the 10-year Treasury briefly above 5% and a Fed hike priced at 90%, the external funding cost of the negative-FCF tier is rising exactly when their cash generation is weakest ai-slowdown-debate-rotation
.
What to watch: Meta's Q3 report in late October — whether the $130–145B capex guide holds and whether operating margin stabilizes anywhere near 31%.
Headline earnings are increasingly not operating earnings
A new fault line opened this quarter: the basket's most eye-popping "beats" were mark-to-market, not operations. S&P 500 Q2 EPS rose 52.6% — the biggest increase since 2021 — but much of that was paper gains on tech stakes held by Alphabet and Amazon ai-stake-paper-gains-earnings-quality.
"Google's parent company — Alphabet — posted a roughly $98 billion net gain related to its stake in SpaceX." — ai-stake-paper-gains-earnings-quality
That's why Alphabet printed $112.19B in net income against operating income of $40.77B, and Amazon booked roughly $53 billion tied in part to Anthropic. Strip the marks out and, per FactSet's John Butters, earnings were still up 33.8% — strong, but the marginal "beat" narrative that moved individual tickers was substantially paper ai-stake-paper-gains-earnings-quality. This makes cross-company EPS comparisons inside the basket unreliable without adjusting to operating income — and it means Alphabet's headline EPS is now hostage to SpaceX's private valuation.
What to watch: Any SpaceX ($SPCX) mark-down flowing through Alphabet's other income line next quarter.
Apple's capital-light model is intact — but no longer static
The capex spread is the widest on record, and the Apple end of it now comes with an asterisk. Apple remains the group's FCF machine, but it is reportedly weighing its first enterprise hardware move since Xserve: an AI server combining its own M8 Ultra chips with Nvidia's NVLink Fusion networking apple-server-pivot-ternus.
"Such a move would signal Apple's push into AI infrastructure — targeting the fast-growing inference market as enterprise workloads shift from training AI models to running them." — apple-server-pivot-ternus
The lineage matters: the project was backed by John Ternus when he was hardware chief, before succeeding Tim Cook as CEO in April 2026, and it would mark a thaw in Apple–Nvidia relations after nearly two decades — notable because Apple is currently a tenant in Google's data centers running Nvidia chips for its revamped Siri effort apple-server-pivot-ternus. The 2029 timeline is distant and the project could be canceled, but it suggests the "capital-light vs. hyperscaler" divergence has a possible convergence point: the consumer pole considering selling infrastructure just as Meta considers renting out its compute capex-divergence-hyperscalers-vs-apple
.
What to watch: Whether Apple confirms, cancels, or re-shapes the M8 Ultra server program — and whether NVLink survives in the design.
Tesla is the mirror image of Nvidia
Tesla is now the only Mag 7 member whose revenue is accelerating while operating income goes to zero: operating income fell from $1.62B to $0.94B to $0.40B over the past year, landing at a 1.4% operating margin on $28.24B of Q2 revenue, with FCF at −$1.09B tesla-existential-autonomy-capex-pivot. The bull case is entirely a call on execution, not the existing P&L — and the strategic bets keep layering on top of each other.
"Tesla recently abandoned its Solar Roof product... Elon Musk... is pursuing a massive solar-cell manufacturing initiative in Texas, Project Crystal Sun, aiming for 100 gigawatts of annual U.S. solar production." — tesla-existential-autonomy-capex-pivot
That solar-manufacturing program sits on top of the autonomy/Cybercab pivot and the $16.8B Tesla–SpaceX Terafab chip fab — all funded by a business running a 1.4% operating margin tesla-existential-autonomy-capex-pivot. The stock trades at a P/E of 334, 25.6% below its 52-week high — the weakest technical position in the basket tesla-existential-autonomy-capex-pivot
.
What to watch: Whether the Terafab construction start slips, and whether any new capital commitment arrives before the auto margin recovers.
The basket cracked in one week — in three directions
The clearest evidence yet that these seven names no longer trade as one came the week of September 14, when the shock came from inside the AI ecosystem itself ai-slowdown-debate-rotation.
"The leaders of Anthropic, OpenAI and xAI warned of risks from rapid development, the starkest threat yet to the billions of dollars being poured into the industry that have driven markets to record highs." — ai-slowdown-debate-rotation
The damage was concentrated, not uniform: Nvidia fell 3.4% and the PHLX chip index tumbled 5.9%, while software names that had sold off on AI-disruption fears — ServiceNow, Adobe, Workday — rallied between 4% and 7.4% ai-slowdown-debate-rotation. Nvidia's own CEO pushed back with "We Are Going to Go as Fast We Can," and Altimeter's Brad Gerstner put a number on the debate: the AI labs must collectively lift their annualized revenue run rate from roughly $100 billion to $180 billion to keep the Nvidia trade alive ai-slowdown-debate-rotation
. Chips down, platforms mixed, software up — one week, three directions inside "tech."
What to watch: Whether the Fed hiked on September 16 and how the resulting discount-rate environment reprices the negative-FCF hyperscalers versus the FCF machines.
What surprised us
- Meta's FCF went to effectively zero and nobody flinched. A ~91% year-over-year FCF collapse alongside a 12-point margin compression, and the story is framed as a bold "mega cloud operator" pivot rather than a crisis capital-allocation-and-free-cash-flow-divergence
.
- Google Cloud is the acceleration story hiding in plain sight. 82% revenue growth with operating margin expanding from 20.7% to 35.6% and a $514B backlog — the single clearest acceleration in the group, growing faster while making more per dollar than anyone else in cloud cloud-infrastructure-growth-and-margin-divergence
.
- Apple is the only Mag 7 name trading above its analyst target ($336.13 vs. $328.22) — meaning the market has already priced option value into a 2029 server project that hasn't been confirmed to exist apple-server-pivot-ternus
- The AI labs called for their own slowdown — and the market's first read was to sell the suppliers, not the labs. The customer warning the vendor's revenue, not the other way around ai-slowdown-debate-rotation