MSFT Beats, META Misses: Mag 7 AI Spend Diverges
Microsoft cuts 2026 capex to $175B while Amazon raises to $220B. Q2 earnings reveal a fracturing AI spending consensus across Big Tech.
Key Points
- Meta's $2.4B legal charge and $1.18B in severance costs collapsed Q2 EPS to $6.18 against a $7.19 consensus — a $1.01-per-share miss on a quarter where revenue actually beat.
- Microsoft cut its 2026 capex forecast to roughly $175B from ~$190B while simultaneously reporting Copilot seat additions more than doubling quarter-over-quarter, signaling improving AI unit economics.
- Amazon's raise of 2026 capex from $200B to $220B — driven by soaring memory chip prices and demand locked in through 2028 — puts it on a collision course with margin expectations heading into H2.
Meta beat on revenue and still cratered on earnings — that single-line summary captures the central tension running through the Magnificent 7's Q2 reporting season. A $2.4 billion legal charge and $1.18 billion in severance costs hit Meta's bottom line hard enough to push EPS to $6.18 against Street expectations near $7.19, a $1.01-per-share miss that had nothing to do with AI monetization and everything to do with legal and restructuring drag that management has not fully guided the market to expect. Meanwhile, Microsoft cut its capital expenditure forecast and watched its stock soar. The divergence is the story.
Microsoft's Capex Cut Is the Most Underreported Number This Earnings Season
The instinct when a technology company cuts its capital expenditure forecast is to read it as a demand signal — a sign that growth expectations are softening. Microsoft's revised 2026 capex guidance of roughly $175 billion, down from approximately $190 billion communicated in April, is not that. It is, if management's framing holds, a supply-side efficiency story: Copilot seat additions more than doubled quarter-over-quarter, indicating that the existing infrastructure footprint is generating more monetizable compute per dollar deployed than the April model assumed. That's a unit economics improvement, not a retreat.
The distinction matters enormously for how traders should think about Microsoft's AI trajectory. The original $190 billion figure reflected a build-first posture — spend aggressively on Azure capacity, data center expansion, and Copilot infrastructure in anticipation of demand that was expected to materialize over the following 12 to 18 months. The downward revision to $175 billion suggests that either demand is arriving faster than the physical buildout can absorb (a good problem), or that Microsoft found pricing and efficiency gains in its supply chain that reduced the cost of achieving the same compute capacity target. CEO Satya Nadella's framing around Copilot evolving from a chat assistant into autonomous AI agents points toward the former — agents generate persistent, recurring compute demand rather than the episodic query load that characterized early ChatGPT-style interfaces.
The Magnificent 7 earnings divergence makes Microsoft's outperformance more meaningful in context. Of the large-cap AI names reporting this cycle, MSFT posted the strongest post-earnings stock reaction in the cohort. That's a signal about relative positioning, not just absolute results. Investors have been wrestling with how to value AI capex commitments that span multiple years without clear line-of-sight to proportional revenue growth. Microsoft's combination of a capex cut and accelerating Copilot seat adoption is the first concrete evidence this cycle that an AI-native product is scaling at rates that change the capital intensity calculus.
Amazon Goes the Other Direction — and Means It
Amazon's decision to raise 2026 capex from $200 billion to $220 billion is the counterpoint that keeps the broader AI infrastructure thesis intact. CEO Andy Jassy cited two specific drivers: soaring memory chip prices inflating the cost basis of its AWS and AI infrastructure build, and demand visibility that Jassy described as "striking" and already extending into 2028 booking conversations. A $20 billion raise mid-year is not a rounding error — it reflects genuine constraint on the supply side of AI compute, where memory bandwidth and advanced packaging capacity are the binding resources, not wafer starts alone.
The memory chip pressure Jassy flagged is consistent with what TSMC's supply data shows. Wedbush's analysis of TSMC's capacity ramp indicates that 3nm wafer starts could hit 180,000 monthly in early Q4 — two to three months ahead of schedule — with demand pulled forward by Nvidia, AMD, and Broadcom simultaneously. When leading-edge wafer capacity is running ahead of schedule and Amazon is still raising capex by $20 billion, the implication is that the constraint has shifted from silicon to high-bandwidth memory and advanced packaging. That's a different supply chain problem, and it explains why memory prices are inflating even as logic capacity expands.
For Amazon investors, the $220 billion capex commitment compresses near-term free cash flow in a rate environment where the 10-year Treasury is yielding 4.63% and the Fed funds rate sits at 3.63%. The opportunity cost of capital is not zero. Amazon's AWS segment needs to demonstrate that the incremental compute deployed in 2026 generates revenue at a pace that justifies the raise — and Jassy's reference to 2028 demand being "locked in" suggests he's making a long-duration bet on pricing power that won't be validated for two years. That's a credibility wager, not a near-term earnings story.
Meta's Miss Is a Warning About Execution Risk Below the Revenue Line
Meta's Q2 result deserves a precise reading. Revenue of $60.8 billion was a genuine beat — advertising demand remains robust and the company's AI-enhanced targeting tools continue to drive click-through improvement for performance marketers. The EPS collapse to $6.18 versus $7.19 expected was entirely a below-the-line event: the $2.4 billion legal charge and $1.18 billion in severance costs are non-recurring in the accounting sense, but the legal liability trajectory is not resolved. Management has not provided forward guidance that fully accounts for the legal drag expected through H2 2026, which means the risk of a second consecutive EPS miss remains live through the Q3 print.
Reddit's quarter offers a useful counterpoint to Meta's execution struggles. RDDT posted a 31.58% EPS beat — $1.25 actual against $0.95 consensus — with revenue of $804.9 million, up from $499.6 million a year ago and well above the $730.3 million estimate. Reddit's growth is being driven by data licensing revenue tied to AI training agreements, a revenue line that carries high incremental margins and is structurally uncorrelated with the advertising cycle volatility that Meta must manage. The contrast illustrates how differently AI monetization is flowing through the income statement depending on business model: for Meta, AI improves ad targeting but doesn't create a new revenue vertical; for Reddit, AI creates a direct, high-margin revenue stream from the same content asset base.
The forward setup for the AI and tech complex now hinges almost entirely on one date: August 26, when Nvidia reports. Management guided for $91 billion in revenue for the upcoming quarter — a figure that implies continued hyperscaler GPU demand at a pace that validates every capex commitment made by Amazon, Microsoft, and Google this earnings season. If Nvidia delivers and raises, the AI infrastructure thesis is intact and the Meta EPS miss gets filed as idiosyncratic legal noise. If Nvidia guides cautiously — citing Blackwell yield issues, packaging constraints, or any softening in hyperscaler order cadence — every capex commitment in this earnings cycle gets repriced simultaneously. Traders should treat $91 billion as the floor to watch; anything below that number on guidance, regardless of the current-quarter beat, will reset the AI spending narrative heading into Q4.
The Weekly Investor
Daily market analysis for active traders. Free.
