The AI Capex Trade Did Not Break Today. It Bifurcated.

Microsoft +15.2%. Meta −9.0%. Same day, same sector, near-identical capex direction. The 24-point spread is not about how much they spend — it is about whose balance sheet the spending lands on, and whether a paying customer is attached to it.

@dailyanalysts · Thursday, July 30, 2026 · Prices as of 12:00–12:02 p.m. ET · Option B: Major Market News Analysis

1. The Finding, Stated Plainly

The market has stopped pricing AI capital expenditure and started pricing AI funding structure. That is the single most important thing that happened this week, and almost every write-up of today's tape is missing it because they are still framing the question as "is AI spending too high?"

Microsoft and Meta reported within ninety minutes of each other last night. Both are spending historic sums on AI infrastructure. Both saw free cash flow collapse. One is up 15.2% and one is down 9.0%. If the market were punishing capex, they would both be down. If the market were rewarding AI ambition, they would both be up. Instead it split them, and the axis of the split is financing.

Today's tape

One number that reframes the day: Microsoft is a $3.34 trillion company, roughly 6% of the S&P 500. A 15.2% move at a 6% weight contributes about 0.9 percentage points to the index. The S&P is up 1.06%. By my calculation, Microsoft alone is responsible for roughly 85% of today's entire index gain. This is not a market rally. It is one balance sheet being re-rated in public.

2. What Microsoft Actually Printed (Primary Source)

From the Microsoft FY26 Q4 earnings release dated July 29, 2026, and the accompanying financial statements:

Full-year FY26: revenue $331.8B (+18%), operating income $155.2B (+21%), net income $133.7B (+31%), capex $115.9B against $182.9B of operating cash flow. Azure crossed $100B for the first time. M365 Copilot passed 30 million paid seats.

Two things tempering my enthusiasm — which almost nobody flagged

1. $0.27 of the $4.81 was discrete items, disclosed by the company itself: a $3.2 billion gain on the Anthropic investment, plus lower-than-expected voluntary-retirement expense, offset by severance and Xbox impairments. A mark-to-model gain on a private AI position is not operating earnings, and it is not repeatable on command. Strip it and the beat is good, not spectacular.
2. The depreciation schedule changed. Per Stifel's note, Microsoft extended the useful life of offices and datacenters from 15 years to 25 years. With $313.1 billion of net PP&E on the balance sheet (up from $205.0 billion a year ago), that is a large, mechanical reduction in annual depreciation that will flatter FY27 gross margin. Management flagged it as an offset to mix-driven margin compression. It is an accounting tailwind, not an operating one, and it should be discounted when you underwrite FY27 EPS.

The sell side was near-unanimous. Thirteen firms raised or held: Goldman to $640, Wells Fargo to $650, Citi and Morgan Stanley to $600, JPMorgan and Wolfe to $550. Evercore's Kirk Materne: "Christmas came a bit early for MSFT shareholders." The most useful line came from Evercore, noting management's call for positive free cash flow in FY27 "removes a key overhang around a more draconian capex scenario." (analyst round-up)

3. What Meta Actually Printed (Primary Source)

From the Meta Q2 2026 earnings release, July 29, 2026 — reading the condensed consolidated statements, not the headline:

Guidance: Q3 revenue $61–64B against consensus of $63.15B. FY26 expenses raised to $165–169B. FY26 capex narrowed to $130–145B — the bottom was raised. Tax rate raised to 15–17%. And a new disclosure: youth-related trials scheduled this year "may ultimately result in a material loss."

Every major firm cut its price target and not one downgraded the rating. JPMorgan went to Neutral $640 from $725 while modelling 2027 capex of $243 billion, +70% year over year. Wells Fargo cut from $835 to $640 and wrote the most honest sentence on the street: "Meta incremental ROI case remains the most opaque." Evercore listed four reasons for the selloff, including that management is "prioritizing selling intelligence over compute… deferring a nearer-term, more visible revenue / FCF offset." (full round-up)

4. The Story Everyone Missed: Sopaipilla Investor, LLC

On Tuesday, Meta and BlackRock announced a $14 billion venture to build a 1-gigawatt campus in El Paso. It was reported as a partnership. It is not a partnership. It is a deconsolidation, and it is the most important financial engineering event in the AI complex this year.

Here is the actual structure, per the bond marketing documents summarised by Global Data Center Hub:

And it is now a repeatable, named template. Meta is literally naming these vehicles after pastries — Beignet for Louisiana, Sopaipilla for El Paso. When a company builds a naming convention for its off-balance-sheet financing subsidiaries, it is telling you there will be many more.

How big is the gap between reported debt and real obligation?

Moody's put numbers on it last week across six names — Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave (CNBC summary of the note):

Do the arithmetic that Moody's stops short of doing. Direct debt is $460B of a $1.66 trillion obligation stack. The off-balance-sheet portion is 2.6 times the on-balance-sheet portion. When a strategist tells you the hyperscalers have fortress balance sheets, they are describing 28% of the liabilities. Moody's own framing: "The transition from asset-light to asset-heavy models requires unprecedented levels of investment and capital raising," and these moves "threaten credit quality."

5. Where the Risk Went: Your Bond Fund

This is the second-order effect that connects the story to assets outside tech, and it is the part consensus has not absorbed.

Risk that leaves an equity holder does not evaporate. It gets sold to someone. In this case, it is being sold to the investment-grade credit market — and the credit market has started to notice. From Protos and Axios, both citing ICE, FactSet and DTCC data:

The supply side is equally stark. Through July 22, five companies — Alphabet, Amazon, Meta, Microsoft and Oracle — raised nearly $302 billion of debt and equity in 2026, against $108 billion for all of 2025 and under $50 billion annually before 2024. Goldman counts $489 billion of AI-related supply already this year, versus its full-year 2025 estimate of $322 billion. Alphabet alone priced an $84.75 billion equity raise last month.

The third-order effect: what actually kills this

The SPV channel works only if two conditions hold: the tenant's credit stays investment grade, and long rates do not rise faster than the lease escalators. Both are being tested this week.

The 30-year Treasury topped 5.211% on Wednesday — the highest since 2007 — after the Fed held at 3.50–3.75% on a 9–3 vote. Ed Yardeni's read, which I think is exactly right: "Warsh talked hawkishly. But he did not deliver a rate hike. So bond yields rose. Arguably, Warsh failed his first credibility test… the Fed has to raise short-term rates to lower long-term rates." (CNBC) Fed funds futures now price a 60% chance of a September hike.

This is the mechanism by which higher rates actually kill AI capex — and it is not the one people describe. It is not the hyperscalers' equity cost of capital. It is the SPV debt market that now sits between the hyperscaler and the shovel. A 1GW campus financed at A+ spreads over a 5.2% thirty-year is a materially worse project than the identical campus financed twelve months ago. Every subsequent Sopaipilla prices wider. At some spread, the tenant stops signing the lease — and capex guidance gets cut not because demand fell, but because the financing stopped clearing.

6. Today's Melt-Up in Chips Is a Liquidation Clearing, Not a Repair

Micron +15.5%. SanDisk +22.7%. Nebius +28.5%. CoreWeave +22.8%. AMD +12.4%. On no company-specific news.

The explanation arrived this morning via David Faber. Leopold Aschenbrenner's Situational Awareness — a fund that reached $45 billion at the start of July — was forced by its prime brokers (Bank of America, Goldman Sachs, JPMorgan) to sell its entire public equity book to meet margin calls. Citadel bought the whole thing. Its largest Q1 positions were Nebius, SanDisk, Micron and CoreWeave — all four down more than 35% this month. It was long AI infrastructure with leverage and short software, and both legs went against it. (CNBC)

7. Scorecard: One Big Win, One Clean Loss

Other open positions, marked to market: NVDA long $195–210 — not triggered, trading $193.15, above the $185 invalidation, see new CDS condition below. ARCC long $18.40–19.20 — trading $18.58, in zone, but Q2 missed on both lines (EPS $0.47 vs $0.4826; revenue $768M vs $790.1M); thesis weakened, holding with the stop. APO long $118–126 — trading $118.57, at the bottom of the zone. Short FSK — trading $10.47, −2.6% today, working. OBDC accumulate $10.20–10.80 — trading $10.71, in zone, but I am still waiting for the Aug 5 print before adding. KO/XLP defensive — unchanged.

8. The Calls

① SHORT / AVOID META  HIGH CONVICTION

Three independent signals. (a) Operating income fell 8% on 28% revenue growth; the margin went 43% → 31%. (b) Zero buybacks in H1 2026 versus $22.9 billion in H1 2025, alongside $24.9 billion of new long-term debt in a single quarter — the capital-return story is over and the balance sheet is in retreat. (c) The marginal capex dollar is now financed outside the company at A+, which means the $130–145B guide understates the economic commitment; JPMorgan already models $243 billion for 2027.

My fair value work, shown. Street 2027 EPS is about $34 (BofA anchors it: "at $542, Meta is valued at 16x our '27E EPS"). I think that number is too high. D&A was $6.356B this quarter, up from $4.342B. Net PP&E is $225.7B and rising roughly $50B per half-year. Even on the extended-life conventions the industry is adopting, quarterly D&A heads toward $12–14B by late 2027 — call it $50–56B annualised against roughly $25B in 2026. That is $25–30B of incremental pre-tax expense, ~$21–25B after tax, ~$8–10 per share of EPS drag. $34 − $8 = ~$26. At 18x, that is $468. My target of $455 assumes the multiple compresses slightly further before the monetization proof point arrives.

Why I am comfortable against consensus: thirteen firms cut their price targets this morning and zero downgraded the rating, with a minimum target of $640. When the entire street marks down the number but nobody marks down the recommendation, the de-rating is not finished. That gap is the trade.

② LONG BLACKROCK (BLK)  HIGH CONVICTION

Own the funder, not the builder. Two signals. (a) BlackRock just proved that GIP and HPS — acquired for $12.5B and $12B respectively, $24.5B combined — can jointly originate, underwrite and syndicate a $12.3 billion A+ rated infrastructure bond with a megacap sole tenant. That is the highest-margin, lowest-risk seat in the entire AI capital stack: infrastructure-equity economics with investment-grade tenant credit, and none of the depreciation. (b) The pipeline is not a forecast, it is a construction schedule — Goldman counts $489 billion of AI-related supply already in 2026 and Moody's projects ~$1 trillion of hyperscaler capex in 2027.

BLK revenue grew 26.5% year over year and the stock sits about 10% below its 52-week high of $1,219.94 while the assets it finances trade at records. Meta is 33% below its high. The market is paying up for the compute buyer and discounting the compute landlord. That is backwards.

③ MSFT — RE-ENTRY WATCH, DO NOT CHASE  WATCH

The quarter was the best any hyperscaler has printed this cycle and $678B of RPO at +84% is a genuine moat. But I will not buy a 15% single-session gap. Fair value on 26x FY27 non-GAAP EPS of ~$21 is $546; my risk-adjusted 12-month target is $520 because I am discounting the useful-life accounting benefit and the $3.2B Anthropic mark. Upgrades to HIGH CONVICTION the moment price enters the zone.

④ AVOID / SHORT RALLIES IN ORACLE  SPECULATIVE

Oracle is +6.1% today, two sessions after printing a 52-week low of $114.50, and 64% below its $345.72 high. This is beta, not repair. Here is the mechanism nobody has priced: S&P has Oracle at BBB− — one notch from junk — with 5Y CDS at 215bp versus 145bp at the start of the year. If Oracle loses investment grade, it exits the IG corporate indices, and every passive IG fund must sell it mechanically, regardless of price or valuation. On $100B+ of outstanding debt, that is a forced-seller event inside the bond market that then feeds back into the equity. Marked SPECULATIVE rather than HIGH because a single dominant signal (credit) is driving it, and as today demonstrated, the equity can squeeze violently in the meantime.

⑤ CRWV / NBIS — DO NOT BUY THIS BOUNCE  AVOID

No entry, no target — this is an avoid with a re-engagement trigger. CoreWeave's 5Y CDS at 855bp implies roughly a coin flip on default within five years. You cannot own an equity whose credit market prices 50/50 survival and call it an investment; at best it is a rental. Today's +22.8% and +28.5% are, in my read, the mechanical consequence of the Situational Awareness book clearing to Citadel. Re-engagement condition: CRWV 5Y CDS closing below 600bp. Until then, nothing.

⑥ NVDA — HOLD THE WATCH, ADD A CREDIT INVALIDATION  WATCH

NVDA at $193.15 is still below my $195–210 entry zone and above the $185 invalidation, so nothing has triggered. I am adding one new condition that did not exist a month ago: an NVDA 5Y CDS close above 120bp kills the idea regardless of price. The CDS doubled to a record ~82bp in four weeks on Bloomberg's report of more than $750 billion of prospective commitments — including a partnership with SK Group valued above $500 billion and talks over guaranteeing as much as $250 billion of OpenAI compute leases. When a company with essentially no net debt sees its default insurance double in a month, the market is not pricing leverage. It is pricing contingent leverage — guarantees. That is a genuinely new risk factor for Nvidia and it belongs in the invalidation.

9. Three Scenarios, With Triggers

BULL — "Structured finance holds" · 25%

Triggers, all three required: the next hyperscaler SPV bond prices at or inside Sopaipilla's spread; the 30-year closes back below 5.00%; Oracle 5Y CDS closes below 175bp. In that world the funding channel is validated, 2027's ~$1 trillion of capex proceeds on schedule, and the builders re-rate. QQQ above $700, MSFT $520–550, META reclaims $585.

BASE — "A two-tier market" · 55%

Triggers: 30-year holds a 5.00–5.40% range; September hike odds stay between 40% and 70%; Oracle CDS stays 175–250bp. Cash-funded contracted compute (MSFT, AMZN, AVGO) keeps working; debt-funded uncontracted compute (META, ORCL, CRWV, NBIS) trades at a persistent and widening discount. The index goes nowhere while dispersion inside it runs at the widest in years. This is where I am positioned. Note the corroborating breadth data: 71% of S&P 500 names are above their 200-day moving average, the highest in two years, yet the index has not made a new high since June 2. The average stock is fine. The index is hostage to a handful of balance sheets.

BEAR — "The SPV bid disappears" · 20%

Triggers, any one sufficient: a hyperscaler SPV bond deal is pulled or downsized; Oracle loses BBB−; CRWV CDS above 1,000bp; or the 30-year above 5.50%. Because the six names are 8.6% of IG duration-times-spread, an Oracle fallen-angel event does not stay inside Oracle — it forces index-fund selling and widens the entire high-grade complex. That is the contagion path from an AI equity story into a retail bond fund, and it is the single risk in this piece I would most want hedged.

10. Cross-Asset Regime Levels

The macro print, and why it matters here

Q2 GDP came in at 1.5% versus 1.8% expected, but the miss was federal spending (−0.3%) and inventories (−0.7%). Underlying demand was strong: personal spending +2.1%, and final sales to private domestic purchasers +3.9%. Core PCE was 3.3% annual, headline 3.7%. The savings rate fell to 2.7%, a four-year low — consumers are funding consumption out of savings, which is not a durable input to the 3.9% number. (BEA data via CNBC) That combination — decent real demand, sticky 3.3% core, a hawkish-but-passive Fed — is exactly the mix that keeps the long end elevated. Which keeps the SPV channel expensive. Which is the whole thesis.

11. How I Would Position This

Timeline

• Next 1–2 weeks: the single highest-information event is the pricing of the next AI SPV bond deal. Watch the spread versus Sopaipilla. That one print tells you more than any earnings release.
• Next 4–6 weeks: September FOMC, currently 60% priced for a hike. Counterintuitively, a hike would help the long end and therefore help the SPV channel — Yardeni's point, and I agree with it.
• Next 3–6 months: Meta must either produce non-advertising AI revenue or cut capex. JPMorgan's $243B 2027 estimate is the marker. If Meta guides 2027 capex above $200B in the January print without a monetization proof point, $455 prints.
• Next 6–12 months: Oracle's credit rating is the systemic variable in this entire complex. Watch S&P, not the stock.

12. The Summary

The main thesis: AI capex is no longer priced as a single trade. It has split along the axis of who funds it. Microsoft spends $116 billion a year out of $183 billion of operating cash flow against a $678 billion contracted backlog, and the market paid 15.2% for that in one session. Meta spends comparably, funds it with $24.9 billion of new debt in a quarter plus a $12.3 billion off-balance-sheet vehicle in which it is the sole tenant, cancelled its buyback, watched operating income fall 8%, and lost 9.0%. Same sector, same day, same direction of spending, opposite outcome. Financing is now the variable.

What consensus is missing: the risk leaving these equities is landing in investment-grade credit. $1.2 trillion of debt-equivalent lease commitments against $460 billion of reported debt. AI CDS volumes up ~600% year over year. Six tech names at 8.6% of IG duration-times-spread risk. Oracle one notch from junk. Most people who are now exposed to the AI buildout do not own an AI stock — they own a bond fund.

Actionable steps:

1. Book the Microsoft win at $449.88. Do not chase. Re-entry $415–430, target $520.
2. Short Meta $528–560, target $455, stop on two closes above $585. HIGH CONVICTION
3. Buy BlackRock $1,050–1,100, target $1,275, stop below $1,010. HIGH CONVICTION
4. Avoid or short Oracle rallies $124–135 toward $98. SPECULATIVE
5. Do not buy the CoreWeave/Nebius bounce. Re-engage only below 600bp CDS.
6. Rotate index exposure to equal-weight while breadth is 71% and the cap-weighted index is stuck.
7. Watch three numbers, in order: the 30-year at 5.40%, Oracle CDS at 250bp, and the spread on the next hyperscaler SPV bond.

Sources

• Microsoft Corp., FY26 Q4 Earnings Release and Financial Statements, July 29, 2026
• Meta Platforms, Inc., Second Quarter 2026 Results, July 29, 2026
• Global Data Center Hub — Sopaipilla Investor, LLC $12.3B senior secured notes structure, July 26, 2026
• Moody's Ratings on hyperscaler credit quality, via CNBC, July 24, 2026
• Axios — hyperscaler debt, CDS and capital raising, July 27, 2026
• Protos — AI credit default swap levels and DTCC volumes, July 29, 2026
• Q2 GDP and June PCE, BEA via CNBC, July 30, 2026
• CNBC / David Faber — Situational Awareness forced liquidation, July 30, 2026
• Sell-side reaction to Microsoft FY26 Q4, CNBC, July 30, 2026
• Sell-side reaction to Meta Q2 2026, CNBC, July 30, 2026
• Meta Q2 earnings coverage, CNBC, July 29, 2026
• Yardeni on Warsh and the long end, CNBC, July 30, 2026
• S&P 500 breadth above the 200-day, CNBC, July 30, 2026

Analysis and opinions are my own and are marked as such throughout. All price data captured 12:00–12:02 p.m. ET on July 30, 2026. Not investment advice. Positions and levels are stated so they can be scored — I publish my losses alongside my wins.