The AI Capex Bet, Six Months In: What the Mag 7 Actually Spent (and What the Market Learned)

In February 2026, we called the Mag 7's 2026 AI infrastructure spend a $700 billion bet — up roughly 60% from 2025, with the hyperscalers alone guiding toward $650 billion. Six months and a full Q2 earnings season later, that number is already low. Actual guidance from the companies that have now reported puts the real 2026 figure meaningfully higher, and — more importantly — the market has stopped treating "more AI capex" as one undifferentiated story. It's now pricing capex very differently depending on whether the company reporting it can show hard evidence the spending is converting into contracted revenue.

That's the real update. Not just bigger numbers — a genuinely different market reaction function than the one we described in February.

What Actually Happened, Company by Company

*Microsoft's headline capex number needs an asterisk we'll come back to below — a meaningful chunk of the "reduction" is an accounting change, not less real spending.

Every hyperscaler that's reported so far raised its number from what we had in February. Alphabet and Meta both came in above even their own April guidance. Tesla's capex is now explicitly framed by management as rising "for the next couple of years," not stabilizing. Amazon reported Q2 2026 the same day this update was written: quarterly capex hit $54.2 billion, up from $32.2 billion a year ago, pushing trailing-twelve-month capex to $169.0 billion — up 64% year-over-year and tracking well above the original $200B full-year guide's implied run-rate. Amazon didn't reiterate a specific new full-year number on the call, but the trajectory speaks for itself. Apple reported the same day and did not move: it still guides no capex number and remains the smallest spender in the group. What changed at Apple wasn't the spending line — it was the reason the spending line stopped protecting it, and that's the Apple section below.

The Big Change Since February: The Market Stopped Treating Capex as One Story

Our February piece flagged the bull/bear debate but treated the market's reaction to all this spending as roughly uniform — broad skepticism, priced in across the board. That's no longer true. This earnings season produced a genuinely clean natural experiment, because three hyperscalers reported heavy capex guidance within about a week of each other, and the market drew three different conclusions:

• Alphabet raised capex to $195–205B, posted its first-ever negative free cash flow quarter (−$5.9B), paused buybacks to zero, and raised $49.6B in fresh equity — stock fell 7% in two sessions, even though Google Cloud operating margin nearly doubled (20.7% → 35.6%) and Cloud CEO Thomas Kurian said existing customers are growing spend 50% year-over-year.
• Meta's free cash flow collapsed to $784 million from $8.5 billion (Meta's own reported basis, which subtracts finance-lease principal; our series uses operating cash flow minus capex and reads $1,746M for the same quarter), capex hit $31.1B for the quarter, and buybacks went to zero while Meta issued $24.9B of new debt — stock fell 9%, and Zuckerberg spent part of the call fielding live questions about why Meta doesn't build a cloud business the way its peers have.
• Microsoft's capex guidance came in below the Street's model, and management backed it with a $678 billion commercial backlog, up 84% year-over-year and up $51 billion in a single quarter — stock rose 8.9%, adding roughly $260 billion of market value in one session.
• Amazon's quarterly capex hit $54.2 billion (TTM: $169.0B, +64% YoY) — but AWS grew 37%, its fastest pace in 18 quarters, while AWS operating margin simultaneously expanded to 39.4% — stock rose more than 8% after hours, even though Q3 guidance landed a touch below consensus.

The pattern is consistent: the market isn't punishing capex size. It's punishing capex that shows up without an independently verifiable demand signal attached, and rewarding capex that shows up with one. Microsoft's backlog number and Amazon's AWS reacceleration-plus-margin-expansion combination are about as close to hard proof as this earnings season produced that AI infrastructure spend is converting into contracted, forward-visible revenue rather than a hopeful multi-year bet — and tellingly, both got rewarded with almost identical stock reactions (roughly +8-9%). Alphabet's argument (expanding Cloud margins) is a good one but requires more interpretation; Meta currently has no cloud business at all to point to as a monetization proxy, which is likely why its reaction was the harshest of the four.

Reportedly, the four largest hyperscalers combined are now sitting on close to $2 trillion in cloud backlog — a figure that, if anywhere close to accurate, is the strongest evidence available that this capex wave is chasing real, already-committed demand rather than speculative overbuild. The market's current instinct is to treat capex as a cost until proven otherwise; that backlog number is the best available argument it's actually an investment, and the divergent reactions above show investors are, at least selectively, starting to make that distinction rather than punishing every hyperscaler equally.

Free Cash Flow Is No Longer a Forecast — It's Actually Cratering

Our February bear case flagged FCF pressure as a risk to watch. Six months later it's not a risk anymore, it's a fact, at multiple companies simultaneously:

• Alphabet: first negative free cash flow quarter on record (−$5.9B), on capex that nearly doubled YoY.
• Meta: FCF fell to $784 million from $8.5 billion — a 91% collapse, narrowly avoiding negative territory. Both figures are Meta's reported basis; on our series the fall is 80.6%, to $1,746M.
• Tesla: FCF turned negative (−$1.09B) for the first time in more than two years, as capex jumped 142% YoY — the quarter is covered in our Tesla Q2 analysis.
• Amazon: trailing-twelve-month FCF swung to an outflow of $7.6 billion, from a positive $25.9 billion a year ago — a swing of more than $33 billion, driven by property-and-equipment purchases running at a $169.0 billion trailing pace, up 64% year-over-year. Detail in our Amazon Q2 analysis.

Four of the seven now, not three. Amazon's case is the most reassuring version of this pattern, though: trailing operating cash flow still grew 33% to $161.4 billion over the same period — the business is generating meaningfully more cash even as it spends faster than that cash arrives, a materially different story than a company where operating cash flow is flat while capex balloons. Capital return responses varied: Meta and Alphabet both paused buybacks to zero (versus $10B+ each a year earlier) and leaned on new debt or equity (Meta issued $24.9B in new long-term debt; Alphabet raised $49.6B in fresh equity, including its first-ever mandatory convertible preferred, plus $20.3B in new notes); Amazon's own capital-return posture didn't see the same dramatic pullback in this release. This is a genuine regime change from February, when the framing was still "capex is eating into FCF." Now, for a majority of the group, capex has fully eaten it, and financing structures are shifting to compensate — with Amazon so far the exception showing the underlying cash engine keeping pace.

The Tesla and Amazon figures above are as each company reported the quarter, and our own coverage of both prints carries the detail. Our stored series do not yet reach them — our Tesla free-cash-flow series ends at 2026 Q1 and Amazon's at 2025 Q3 — so a reader cross-referencing /stocks/tsla or /stocks/amzn will not find these quarters there. That is a coverage gap on our side, not a dispute about the figures.

One name outside the seven belongs in this section and was missing from it when we published. Oracle's free cash flow for its fiscal year ended May 31, 2026 was −$23,686M, against −$394M the year before, with two quarters burning more than 60% of revenue — while revenue in its closing quarter, ended 2026-05-31, grew +20.6%. It is the borrower's side of this same buildout, and we covered it separately in Oracle burned $23.7 billion building the AI boom.

The Contrarian Case, Now Tested: Apple Opted Out and the Wave Reached It Anyway

Our February piece listed Apple's capex at roughly $13 billion — by far the smallest of the group, and we noted it almost as an afterthought. Six months later, that restraint has become Apple's headline story. Apple overtook Nvidia on July 17 to become the world's most valuable company, up 22% year-to-date, specifically while Alphabet and Meta were falling on their own AI capex numbers.

Then Apple reported, on Tim Cook's final earnings call as CEO. The test came back split, and the half we got wrong is the more interesting one.

On the cash question, opting out worked exactly as advertised. Through calendar Q1 2026, Apple's trailing-twelve-month free cash flow was $129.2 billion, up from $98.5 billion a year earlier — a 31% increase, over the same stretch in which Alphabet posted its first negative FCF quarter on record and Meta's fell 80.6%, from $9,023M to $1,746M.

Those three are deliberately on one definition. Apple's TTM figure is derived from our own series, so the Meta comparison here uses our series too — operating cash flow minus capex — rather than the $784M reported figure quoted earlier in this piece. Comparing Apple's house-basis number against Meta's company-basis number would be mixing two definitions in a single sentence, which is exactly the error this comparison exists to avoid. Tesla and Amazon also saw free cash flow go negative over this period, as their bullets above record, but our series does not yet reach those quarters, so they are left out of the like-for-like comparison rather than being restated on a basis we cannot compute.

The companies that built have visibly weaker cash engines than the one that did not. That part of the thesis holds without qualification.

On the strategic question, the answer inverts the one we expected. The company that stayed out of the capex race is the one warning that everyone else's capex has broken the supply chain it depends on. Cook described the global memory chip shortage as a "hundred-year flood," with companies scrambling to secure manufacturing capacity — not hedge-speak, and not a phrase a CEO reaches for about a routine procurement problem. Management warned that supply constraints across advanced-node chips and memory costs will "significantly intensify" for iPhone, iPad and Mac in the coming quarters. Apple guided Q4 revenue growth to 9–11% against the ~12% expected, and the stock fell more than 6% after hours.

The quarter underneath was strong: revenue $109.42B (+16% YoY, a beat) and diluted EPS $2.02 (+29%). That EPS needs its asterisk, and our full analysis of the print carries it — roughly $0.11 of the $2.02, and about 2 points of the 50.1% gross margin, came from a one-time tariff refund. Ex-refund, EPS was nearer $1.91: still a beat on the $1.89 consensus, just a smaller one than the headline implies.

The chart above was already telling this story; the prose stopped one day short of it. The money-flow graphic carries AAPL ~$13B as its smallest source node and Memory ~$50B as a supplier node. Apple sits at one end of that diagram buying from a category the other seven are bidding up at the other end. Restraint protected Apple's free cash flow. It did not — and structurally could not — insulate Apple from a capex wave that consumed the memory supply Apple buys. You can decline to build data centers. You cannot decline to compete for the parts that go inside them.

The supply side of that squeeze is already on this site and reads differently next to Apple's guidance: Micron sold out its entire 2026 HBM capacity back in February, and Sandisk's NAND pricing swings track the same shortage from the seller's side. What is a record year for the memory makers is a cost problem for the largest buyer of consumer electronics components in the world.

One escape route also narrowed the same day. U.S. senators published a letter urging Apple to abandon a reported plan to qualify memory from Chinese suppliers CXMT or YMTC — the workaround Apple had reportedly been exploring to ease precisely this shortage — and asked the company to confirm its position by August 21.

So the six-month answer to "was sitting out the AI capex race smart?" is: it kept Apple's cash flow intact and made it the most valuable company in the world, and it bought no protection at all from the thing the spending set off. Those are both true, and only the first one was in this article yesterday.

Nvidia: Still the Cleanest Beneficiary, and the Numbers Got Bigger

Nvidia's own capex remains modest — it's still the picks-and-shovels supplier, not a hyperscaler builder — but the demand signal keeps strengthening. Where our February estimate had Nvidia capturing $320–330 billion in data-center revenue for the year, the company's own more recent guidance points to roughly $91 billion in data-center revenue for a single quarter — an annualized run-rate meaningfully above our original full-year estimate. Wedbush's Dan Ives was on CNBC in late July reiterating a bullish stance, arguing "we're in the third inning of the AI revolution" and that demand is outpacing supply by a ratio of roughly 12 to 1. Whatever one thinks of that framing, the underlying fact — every hyperscaler capex number in this update went up, not down — is directly consistent with it.

The Eighth Name in the Chart: SpaceX

Our updated chart adds one company outside the traditional Mag 7: SpaceX, which went public June 12 at a $1.77 trillion valuation — the largest IPO in history — and is now trading roughly 50% below its post-IPO peak. It belongs in this conversation because its own IPO filings disclosed an AI segment (xAI, Grok, the X platform, and orbital/AI data-center ambitions) that posted a $2.47 billion operating loss in Q1 2026 alone, on $818 million of revenue against $3.29 billion of costs and expenses. That's the same "spend now, prove it later" pattern running through every name in this piece.

One important caveat, visible directly in the chart: the SpaceX figure shown is that Q1 AI-segment operating cost, not annual capex guidance the way the other seven companies' bars are. SpaceX hasn't reported as a public company long enough to have a comparable full-year capex number yet — its first-ever earnings report lands August 4, 2026. We've labeled it explicitly for that reason; treat it as directionally illustrative of scale, not as a like-for-like comparison with the other bars.

Not Just GPUs: Where the Capex Chart Actually Goes

Our updated capex breakdown adds a category we left out in February: CPUs. Data-center CPUs (Intel and AMD's server chips, plus custom ARM designs like AWS's Graviton and Google's Axion) are a real, separate line from the GPU/accelerator spend that dominates the headlines — our own coverage of Intel's Q2 print found data-center CPU demand "outpacing supply," a genuine and distinct capex destination worth tracking on its own rather than folding into a generic "servers" bucket. We've also added a "not itemized" category to both this chart and the original: the seven companies' combined capex doesn't fully reconcile against the publicly discussed spending categories, and that gap deserves to be shown honestly rather than implied away.

One thing we didn't do: assign each company's spending to specific categories. It's tempting to draw a line from, say, Alphabet straight to "custom silicon" — Google (TPU) and Amazon (Trainium/Inferentia) do run the most mature custom-chip programs among the seven, with Meta (MTIA) and Microsoft (Maia) newer and smaller, and Apple and Tesla off in their own idiosyncratic lanes (on-device silicon, and a Dojo program that's been scaled back). But none of these companies discloses dollar-level splits by category, and we're not going to present a fabricated breakdown as if it were real data. The company-to-category story here is directional and qualitative, not something a chart should imply is precisely known.

Microsoft's Capex "Reduction," With the Asterisk

Because it matters for reading every hyperscaler's numbers going forward, it's worth restating clearly: Microsoft's headline capex guidance coming down from ~$190B to ~$175B is not simply spending discipline. Reuters reported Microsoft changed how it accounts for long-term data-center leases, spreading them over 25 years instead of 15 — which lowers the amount of spending that shows up in any single year's capex line without necessarily reducing what the company has actually committed to spend in total. It's the same building, recognized more slowly. Anyone modeling hyperscaler capex trends going forward should treat Microsoft's reported number with that caveat in mind, and watch for whether other hyperscalers make similar accounting adjustments as this spending cycle matures and the depreciation math gets more scrutiny.

Updated Bull vs. Bear Case

The bull case strengthened. Cloud backlog across the four largest hyperscalers is reportedly approaching $2 trillion. Microsoft's Cloud backlog specifically grew 84% YoY to $678B in a single quarter. Google Cloud margin nearly doubled in two quarters. Nvidia's demand signal (per Dan Ives) is still outpacing supply by a wide multiple. None of this reads like the demand side of the equation is weakening — if anything, the backlog numbers are the single best piece of new evidence since February that the spend is chasing real, contracted demand.

The bear case also strengthened, in a different way. FCF pressure moved from a forecast to a fact at three separate companies. Buybacks are being paused specifically to fund capex, and debt is stepping in — a capital-structure shift that increases financial risk if the AI-driven revenue takes longer than expected to show up. And the market has demonstrated, in real time, that it will punish capex increases harshly even at a company (Alphabet) that can show real Cloud margin expansion — meaning the bar for "acceptable capex" keeps rising, not falling, as the numbers get bigger.

What to Watch Now

1. Whether Apple's memory warning shows up in everyone else's cost lines. Apple has now reported, and its guidance problem was the price of components, not the price of data centers. If a "hundred-year flood" in memory is real, it reaches every company in the table above that buys the same parts — and it reaches them through cost of goods sold, where it is much harder to spot than a capex headline.
2. Whether other hyperscalers adopt Microsoft's longer lease-depreciation treatment. If so, "capex is coming down" needs to be read skeptically industry-wide, not just at Microsoft.
3. Whether the reported ~$2 trillion in combined cloud backlog holds up to scrutiny, and on what actual timeline it's expected to convert to revenue.
4. Apple's August 21 response to the senators' letter, and John Ternus's first quarter as CEO from September 1 — the two near-term markers of whether Apple's supply squeeze eases or hardens.
5. Buyback and dividend trends across the group — Alphabet and Meta both went to zero this quarter; watch whether that becomes the new normal across the cohort or reverses once FCF stabilizes.

The Bottom Line

Six months ago we called this a $700 billion bet the market was pricing with broad, undifferentiated skepticism. The bet got bigger — every hyperscaler that's reported has raised its spend, not lowered it — but the market's skepticism got more selective, not less. Microsoft and Amazon both proved that capex paired with hard demand evidence (a ballooning backlog, or accelerating growth plus expanding margin in the same segment) gets rewarded, each earning an 8%+ pop; Alphabet and Meta proved that capex without an equally hard, independently verifiable payoff story still gets punished, even from companies posting genuinely strong underlying quarters. And Apple, which spent the least, showed that third strategy's limit on the day it reported: not spending kept its cash flow intact and still left it exposed, because the components it buys are the ones everyone else's capex is bidding up. The debate hasn't resolved. It's just gotten a lot more specific about which companies get the benefit of the doubt, and why — and two of the four hyperscalers who've now reported have earned it.

This is a six-month update to our February 2026 Mag 7 AI capex analysis, incorporating confirmed Q2 2026 earnings data from Alphabet, Meta, Microsoft, Tesla, Amazon and Apple. Apple's quarter is covered in full in our Q3 FY2026 analysis; Apple free cash flow figures are from r40's own series for AAPL, trailing twelve months through calendar Q1 2026.