Microsoft Rose 8 Percent and Meta Fell 8 on the Same Night. The AI Capex Trade Just Split in Two.

Two companies reported earnings after the close on July 29, within about an hour of each other. Both printed the largest quarterly capital expenditure in their history. Both spent that money on the same thing: GPUs, data centers, power. Microsoft finished the after-hours session up around 8 percent. Meta finished down about the same.

Four days ago we published a scoreboard for exactly this week, after Alphabet posted the first negative free cash flow quarter in its public history and got sold off for it. The thesis was that the market had started reclassifying hyperscalers from capital-light monopolies into capital-intensive infrastructure companies, and that backlog was the only defense. Tuesday night ran the experiment with a clean control group. Same night, same macro, same spend category, opposite verdicts.

The market is not grading AI capex anymore. It is grading whether the capex has a counterparty.

The Same Night, Side by Side

Read the revenue row again. Meta grew faster. Meta grew ten points faster than Microsoft and lost a tenth of its market cap after hours anyway. Growth was never the question this week. The question was what stands between the capex line and the shareholder, and the two companies gave opposite answers.

What Microsoft Sold the Street

Microsoft's answer was one number: commercial remaining performance obligations of $678 billion, up 84 percent year over year. That is contracted, signed revenue that has not been recognized yet, and it grew more than $50 billion in a single quarter. Management added the two qualifiers the street had been asking for since April: roughly 30 percent of it converts to revenue within twelve months, and the weighted average duration is 2.3 years. This is not a vibes number. It is a schedule.

Azure did its part too. Growth accelerated to 43 percent, up from 40 last quarter, and the business crossed $100 billion in annual revenue for the first time. Guidance for the September quarter is roughly 45 percent in constant currency. When your cloud is accelerating at that scale, a record $35.8 billion capex quarter reads as fulfillment cost for demand you have already signed, not as a bet. Free cash flow still came in at $19.6 billion. Squeezed, but decisively positive, and nobody had to open the financing section of the cash flow statement to check how the quarter was paid for.

One footnote deserves more attention than it got: $0.27 of the EPS beat came from discrete items, mainly a $3.2 billion mark-to-market gain on Microsoft's Anthropic investment. The cleanest AI quarter on record was itself partly an AI valuation gain. The circle in this market never fully closes. File it next to Alphabet funding Anthropic while Anthropic fills Google Cloud's backlog, and read our Nvidia guarantee piece for the fully load-bearing version of the same loop.

What Meta Could Not Sell

Meta's problem is structural and we said so on Saturday: it is the only company in this group spending hyperscaler money without a hyperscaler business model. There are no external compute customers, no RPO line, no backlog to point at. Every dollar of the roughly $31 billion Meta poured into infrastructure this quarter is a bet that better models raise ad pricing and engagement enough to pay for the buildout. That may even be true. But it is a thesis, not a contract, and Tuesday the market priced the difference.

The quarter itself was fine by any pre-2026 standard. Revenue up 28 percent to $60.8 billion beat estimates. But expenses grew 55 percent to $42 billion, EPS of $6.18 missed, and free cash flow collapsed 91 percent to $784 million. Meta also borrowed roughly $25 billion in the quarter to keep funding construction, which means the ad machine plus debt is now the funding stack for a buildout guided at $130 billion to $145 billion for the year. That guide is worth a second look: Meta raised the bottom of the range from $125 billion, not the top. Raising the floor is a commitment. It says the minimum spend went up regardless of what the return curve does.

A company one quarter from negative free cash flow, with no contracted revenue against the spend, funding the gap with debt. That sentence described Alphabet last Wednesday minus the backlog, and Alphabet fell 5 percent. Meta fell nearly twice as far because it is missing the one line that saved Microsoft.

Scoring Our Own Signposts

The July 26 piece ended with three things to watch this week. Here is how they resolved.

The partial on the first signpost matters most. Microsoft gave the street an ex-OpenAI growth rate for bookings but not for the backlog stock itself, and the concentration question does not go away because the aggregate number is enormous. If something like 40 to 45 percent of that $678 billion sits with a single counterparty that loses double-digit billions a year, the quality of the backlog defense depends on the credit of one customer. That is the thread connecting this quarter to the Nvidia guarantee story: everyone's cleanest number has OpenAI's balance sheet somewhere inside it.

Amazon Reports Tonight

The fourth print lands after the close today, and Amazon walks in holding both cards at once. It has the backlog: AWS remaining performance obligations stood at $364 billion last quarter, before counting the new $100 billion OpenAI commitment layered on top of the existing $38 billion contract. It also has the spend, with 2026 capex guided north of $200 billion. If the pattern from Tuesday holds, the market will look straight past the capex number to two things: how much of the OpenAI commitment shows up in RPO, and whether AWS growth accelerates the way Azure did. Backlog plus acceleration got Microsoft an 8 percent rally. Backlog with decelerating cloud growth is a different trade entirely.

Our Take

For a year the AI capex debate has been argued as one question: is the spending too much? Tuesday night made clear the market has moved on to a better question: who is on the other side of the spending? Capex with a signed counterparty is inventory for a demand schedule. Capex with a thesis is an expense with a story attached. Microsoft and Meta spent within $5 billion of each other and landed 16 points apart because they sit on opposite sides of that line.

The uncomfortable part is how much of the winning answer rests on counterparties whose own economics are unresolved. Microsoft's backlog leans on OpenAI. Its EPS beat leaned on an Anthropic mark. Amazon's tonight will lean on OpenAI again. The backlog defense works until the market starts asking the same counterparty question one level down, and the Nvidia CDS market suggests some people already are.

Three signposts from here. First, whether Amazon's print tonight confirms the split: backlog plus acceleration rewarded, everything else sold. Second, whether Meta uses the next quarter to manufacture a counterparty, through external Llama licensing revenue, a compute partnership, anything that puts a contracted number against the spend. Third, whether Microsoft discloses an ex-OpenAI RPO figure next quarter, because the longer the aggregate grows without that split, the more the concentration question compounds. We track provider economics and API pricing shifts as they land on our models tracker, and the capex story is increasingly the pricing story: the bill for all of this eventually shows up in someone's per-token rate.