Microsoft Proved AI Capex Can Pay. Meta Showed What Happens When It Can't.

The Nasdaq is one of the cleanest tells in the opening snapshot — but not for the reason the headline number suggests. A 3.3% surge in the Nasdaq-100 (QQQ closed at $683.55, up from $661.73)^[1] looks like a broad risk-on session. It is not. The Technology Select Sector ETF (XLK) gained 5.50% while Health Care (XLV) fell 1.64% and Communications (XLC) dropped 2.68%^[1]. When the spread between the best and worst major sector ETFs in a single session exceeds eight percentage points, the index number stops describing the market and starts describing an argument — specifically, an argument about whether $175 billion in AI infrastructure spending earns its cost of capital.

The argument has a protagonist and an antagonist, and they reported earnings twelve hours apart.

Microsoft: The Quarter That Reset the Tape

Microsoft surged 14.2% — reportedly the largest single-day market-value increase in stock market history^[2] — after a fiscal fourth-quarter print that beat on every line investors had flagged as a risk. Revenue landed at $90.0 billion, up 18% and above the $87.62 billion consensus. Adjusted EPS came in at $4.74 versus $4.24 expected, a 12% bottom-line beat^[3].

The number that reset the tape was Azure. Growth accelerated to 43% year over year, up from 40% in the prior quarter, against a Street looking for 40%^[3]. That pushed full-year Azure revenue past $100 billion for the first time^[2]. But the more consequential figure sits on the balance sheet: commercial remaining performance obligations — contracted future revenue, not pipeline — jumped 84% year over year to $678 billion, a $51 billion sequential build^[3]. Microsoft 365 Copilot crossed 30 million paid seats, up from just over 20 million in April^[3].

The distinction matters because it is the first hyperscaler print this cycle where the coverage ratio moved in the right direction. Capital expenditures including finance leases hit $41 billion in the quarter, up 69% year over year^[3]. But the backlog grew $51 billion sequentially against that $41 billion bill. Operating margin ticked up to 45% even as gross margin compressed on AI infrastructure costs^[3]. Capex is outpacing depreciation’s drag on the income statement. That is the arithmetic of a buildout that is amortizing on something close to the promised schedule.

One detail deserves scrutiny: the reduction in calendar-2026 capital expenditure guidance to approximately $175 billion from roughly $190 billion is an accounting change, not a spending cut. Microsoft lengthened the assumed useful life of office and data center properties to 25 years from 15, effective in fiscal 2027^[3]. The buildings and the silicon are still being bought on the same schedule; the depreciation is being spread across a decade longer. Management said so directly on the call. Anyone reading that $15 billion delta as capital discipline has misread it.

Meta: When a Record Quarter Prints a Cash Problem

Meta Platforms fell 9.2% — and the divergence with Microsoft is the story of the session^[4]. Revenue reached $60.80 billion, up 28% year over year and roughly $600 million above consensus — the largest revenue quarter in company history. Advertising revenue rose 27%, with ad impressions up 14% and average price per ad up 12%^[3].

Demand is not the issue. Conversion of that demand into cash is the issue, and it has become severe.

Diluted EPS came in at $6.18 against $7.22 expected, a miss of nearly 14%^[4]. Total costs and expenses surged 55% year over year, dragging operating margin to 31% from 43%^[3]. The cash flow statement is where the stock broke: operating cash flow rose nearly 25% to $31.86 billion, but capital expenditures including finance-lease principal payments reached $31.08 billion, consuming 97.5% of it. Free cash flow came in at $784 million against $8.55 billion a year earlier — a 91% collapse^[3]. Meta also raised the floor of its full-year capex guidance to $130 billion from $125 billion, the second increase this year after an original sub-$100 billion plan^[4].

What would have to be true for Meta’s spending to be the right call? The company needs a revenue line large enough to offset a capex bill that is roughly doubling from last year’s $72 billion. Microsoft has a cloud business with a counterparty who pays for compute at a contracted rate. Meta’s compute is consumed internally — model training, recommendation systems, personal agents — and monetizes indirectly through advertising yield on a lag nobody can size with confidence^[3]. Management flagged an ambition to build a business serving external customers, but that is a plan, not a backlog.

The 23-Point Spread: What the Market Is Actually Pricing

The 23.4-percentage-point same-day divergence is the widest between two Magnificent Seven members this cycle^[3]. Both companies spent the same order of magnitude on AI infrastructure. One reported a $51 billion sequential backlog increase with expanding margins. The other reported free cash flow of $784 million and a twelve-point margin contraction. The market did not reward AI spending or punish it. It priced the demonstrated return on it, name by name.

This raises a question that cuts to the structure of the trade. If the cloud offset scales with the spend — as Microsoft’s backlog suggests — then Meta’s problem is Meta-specific and the differentiation thesis holds. If it does not scale, the market has to conclude that hyperscaler capex is being deployed against demand that has not materialized, and this becomes a sector de-rating rather than a dispersion trade. Alphabet hit the same wall last week when free cash flow turned negative for the first time on record alongside a raised 2026 capex outlook^[3].

The Fed, the 30-Year, and the Discount Rate Constraint

The Federal Reserve held the target range at 3.50%–3.75% for a fifth consecutive meeting on Wednesday, and the vote was 9-3 — with the Cleveland, Minneapolis, and Dallas presidents dissenting in favor of a quarter-point hike^[5]. Markets had priced roughly 35% odds of an increase going in, making this one of the least predictable meetings in years^[5].

Chair Warsh declined, again, to offer forward guidance. The committee needs to “observe market reaction to developments direct and unfiltered,” he said^[5]. By declining to guide, he has effectively deputized the bond market as a policy instrument. The bond market delivered its verdict inside ninety minutes: the 30-year Treasury yield surged 12 basis points to 5.21%, the highest since 2007 — a nineteen-year peak. The 10-year rose to 4.677%^[3].

The latest FRED macro snapshot shows the 10-year at 4.61%, the Fed Funds rate at 3.63%, CPI inflation at 3.46% year over year, and the yield curve (10-2Y) at a positive 0.35%^[6]. Real GDP grew 2.66% year over year and unemployment held at 4.2%^[6]. The HY credit spread sits at 2.84%, up only marginally — credit markets are not signaling stress^[6]. But consumer sentiment collapsed to 44.8, down 14% year over year and 10% month over month^[6], the kind of reading that historically accompanies an economy where inflation is eroding real purchasing power even as the headline growth number stays positive.

The historical analogs the FRED model surfaces are instructive. The most similar periods are mid-2006 — specifically June through August 2006, when the Fed Funds rate sat between 4.99% and 5.25% and the yield curve was flat to inverted^[6]. That was the late stages of a hiking cycle that preceded a housing-led slowdown. The parallel is not exact — today’s Fed has already cut to 3.63% and the curve is positively sloped — but the combination of elevated inflation, a central bank publicly divided on the next move, and long-end yields grinding higher is a pattern with precedent, and the outcome was not benign.

A 5.21% long end is the binding constraint on the equity thesis. Discount-rate math is unsentimental: the longer the duration of the cash flows, the more a twelve-basis-point move at the back end costs in present value. The names carrying this market — hyperscalers committing $130 billion to $175 billion annually against returns that arrive in 2028 and beyond — are the longest-duration equities in the index by construction.

PCE, Oil, and the Inflation Feedback Loop

Thursday morning brought June PCE data. Core PCE slowed to 3.3%, with the headline index rising 3.7% year over year^[5]. On the surface, that is disinflation. But June’s benign print was built substantially on a 9.2% decline in gasoline during a ceasefire window that has already closed^[3].

The oil market is the feedback loop the equity tape is not fully pricing. Brent gained 7.9% on Wednesday to close at $90.74 after the United States launched strikes against Iran in retaliation for missile attacks on American forces^[3]. Crude pulled back Thursday — Brent near $88.93, WTI near $83.09^[3] — but both choke points for Middle East oil exports, the Strait of Hormuz and the southern Red Sea, are contested simultaneously. U.S. commercial crude inventories posted their largest draw since mid-June, and the Strategic Petroleum Reserve declined for an eighteenth consecutive week, to the lowest level since 1983^[3].

For equities, the transmission channel is inflation, not energy sector earnings. A sustained move back above $90 in Brent rebuilds the energy contribution to headline inflation into the August and September prints — the exact data the Fed said would determine September, and the exact mechanism the three dissenters cited in their vote^[3].

Semiconductors: Equipment Signal vs. Memory Problem

The semiconductor complex bounced more than 7% on Thursday, led by Lam Research’s 18.3% surge on record quarterly revenue^[3]. Semiconductor capital equipment revenue is the cleanest available read on whether hyperscaler capex commitments are converting into actual purchase orders, and record revenue says they are^[3].

But the complex entered the session down roughly 25% since late June, after surging more than 87% in the second quarter — its best quarter on record^[3]. A 7% bounce does not repair a 25% drawdown. The Philadelphia Semiconductor Index’s 50-day rolling standard deviation of daily returns reached its highest level since May 2020 earlier this month^[3].

The concentration is part of the mechanism. Micron, AMD, Nvidia, Intel, and Broadcom together account for roughly 36% to 37% of the semiconductor ETF^[3]. When memory pricing is the marginal variable and memory is heavily weighted, the index becomes a leveraged bet on a single input cost. Qualcomm dropped 4.8% despite beating revenue estimates — a two-cent EPS shortfall cost the stock nearly 5%, which is what happens to semiconductor multiples when positioning is at the 98th percentile of the past five years and prime brokers are calling for collateral^[3].

After the Bell: Amazon and Apple Deliver

Two prints after the close added another chapter. Amazon reported net sales of $200.6 billion, up 20% year over year, with AWS revenue of $42.2 billion — up 37%, marking a fourth consecutive quarter of acceleration^[7]. Operating income was $27.5 billion, up 43%^[7]. Amazon also said it will increase capital spending on technology, mostly AI, by an additional 10%^[7]. The capex line was the one the market was watching most closely: if AWS growth scales with the spend, the differentiation thesis holds. AWS growing 37% on an already enormous base is consistent with Microsoft’s evidence that cloud demand is real and compounding.

Apple posted fiscal third-quarter revenue of $109.4 billion, up 16% year over year and a June-quarter record, with gross margin at 50.1%^[7]. The margin guide is the whole story — the AI-driven memory chip shortage has already forced price increases on Macs and iPads, and whether Apple is absorbing that cost or passing it through will shape the September-quarter outlook^[3].

What to Watch Next

• September FOMC meeting: With three dissenters already voting for a hike and no forward guidance from the chair, September is a live meeting. Incoming inflation prints between now and then — particularly the July and August CPI and PCE releases — are the entire input. Any upside surprise feeds straight into the long end of the Treasury curve and from there into equity valuations.
• Oil and the Middle East: Brent above $90 rebuilds the energy contribution to headline inflation. The Strait of Hormuz and the southern Red Sea are both contested. Watch crude inventory data and any escalation in the Iran conflict for the inflation read-through.
• The 30-year Treasury at 5.21%: This is the discount rate applied to the longest-duration equities in the index. A move through 5.25% would compress the multiples the market pays for hyperscaler cash flows that arrive in 2028 and beyond. A stabilization or reversal would be the precondition for a broader equity rally.
• Semiconductor positioning: Prime brokers including Goldman Sachs and JPMorgan have been issuing margin calls to hedge funds with concentrated AI exposure^[3]. Positioning in semiconductors remained at the 98th percentile even after four consecutive weeks of net selling in early July^[3]. Forced unwinds amplify moves in both directions.
• Hyperscaler capex progression: The three-way sort is now Microsoft (contracted offset, FCF-positive commitment), Alphabet (negative FCF, rising capex), and Meta (heaviest spend as a share of cash generation, no cloud revenue line). Whether Amazon’s AWS acceleration and capex increase reinforces or complicates this sorting is the question the after-hours prints just began to answer.

The base-rate reading is straightforward: earnings are strong, the economy is growing, and the Fed is on hold. The historical analog — mid-2006, late in a hiking cycle with long-end yields rising and a divided central bank — is a reminder that strong fundamentals and a hostile discount rate can coexist, and that the market’s verdict on which one matters more can change quickly. The 23-point Microsoft-Meta spread is the market saying it has already started sorting.