06-reference

stratechery apple earnings amazon aws barbell

2026-08-10·reference·source: Stratechery·by Ben Thompson
stratecheryappleamazonawstrainiumchip-supplytsmcmemory-cyclehyperscaler-capexai-infra

"Apple Earnings, More on Amazon's Earnings" — @BenThompson

Why this is in the vault

Two separate earnings threads worth tracking for the RDCO capex/memory-cycle work: Thompson identifies Apple's post-earnings drop as a logic-chip (TSMC advanced-node) supply constraint, explicitly distinct from the memory shortage everyone else is discussing, and extends last week's AWS follow-up with a "barbell" demand framework plus a prediction that Amazon will start selling Trainium externally, mirroring Google's TPU-sales move.

The core argument

Two linked sections in a short-form Update.

Apple Earnings. Apple shares fell 7% in after-hours trading on September-quarter guidance of 9-11% revenue growth (vs. ~12% Street estimate) and gross margins coming in "just short" of estimates, despite a strong quarter (iPhone +22%, Mac +29%). Tim Cook's earnings-call comments (via the Six Colors transcript) attributed the guidance miss to supply constraints hitting Mac primarily and, to a lesser extent, iPhone/iPad in the June quarter, worsening "significantly sequentially" into September — framed by Cook as "a demand forecast issue, to be candid," not a pure supply issue. Thompson's own read: Apple pulled forward N3 wafer starts with TSMC a year ago (he cites his own prior piece speculating Apple's advanced-node ordering would get squeezed by AI-driven demand from Qualcomm, AMD MI350, Microsoft Maia 200, and Amazon Trainium 3), and "the bill is coming due" now — the iPhone 17 Pro to 18 Pro transition (18 Pro moves to N2) means Apple likely sells out of 17 Pro with ongoing shortages across the M5 family. He explicitly separates this logic-chip shortage from the memory-price shortage dominating other coverage, and frames the stock reaction as the market re-pricing Apple away from "safe harbor" status once sustained higher growth got called into question.

More on Amazon's Earnings (follow-up to Wednesday's Update, sourced from the Q2 2026 AWS earnings-call transcript). Jassy and CFO Brian Olsavsky gave three reasons AI is helping AWS's non-AI "core" business: proximity of inference to existing data/ops (AWS's structural lead), AI driving genuinely non-AI compute demand (post-training RL and agent tool use run mostly on CPUs — Graviton cited at "30% to 40% better price performance"), and AI accelerating general cloud migration ("strong linkage between AI spend and core growth"). Thompson flags an inconsistency in Jassy's frontier-model stance: Jassy first argues AWS doesn't need its own frontier model ("not going to be one model to rule the world"), then predicts Amazon will have one of "at least a half dozen" comparably good models within a few years — Thompson finds the first answer "much more believable." Jassy's demand framework is a barbell: AI labs (heavy compute) at one end, enterprises capturing cost/productivity value at the other, with the "messy middle" of current enterprise production workloads as the eventual largest segment (AWS chips now over $25B annual revenue run-rate; AWS/AI framed as a potential $1 trillion revenue business). Thompson extends his own 2024 "Enterprise Philosophy and the First Wave of AI" argument: the middle may get absorbed by both ends expanding — labs enabling new challengers on one side, AI eliminating job categories at incumbents on the other — rather than the middle itself transforming. Closing prediction: Amazon will likely start selling Trainium externally/standalone (not just via AWS), same playbook as Google selling TPUs, to capture customers who'll never be AWS cloud customers (e.g., Meta) or who'll eventually build their own datacenters (the labs) — which Thompson calls "very pragmatic" and "inherently skeptical about the importance of owning the frontier."

No sponsor block in this issue. Self-promotion limited to an intro plug for Thursday's in-person Sharp Tech episode (Microsoft's middleware AI plan, Meta's market tussle, Situational Awareness) and the standard Stratechery Plus podcast-suite footer.

Mapping against Ray Data Co

The direct actionable piece is the logic-chip vs. memory-chip distinction: investing-edgar-watch and the standing 01-projects/investing/theses/2026-05-18-memory-cycle-v1.1.md thesis have been tracking memory-price pressure as the dominant capex/margin story this earnings season (see the 2026-08-05 Amazon note, where Amazon's own capex raise was attributed to memory costs). Thompson's Apple section is a reminder that logic-chip (advanced-node TSMC capacity) scarcity is a separate, compounding constraint that hit Apple specifically because it pulled forward wafer orders — worth flagging in the next capex-anchor pulse so memory-driven and logic-driven supply stories don't get collapsed into one undifferentiated "chip shortage" headline.

Second, the "barbell" demand framework and the externally-sold-Trainium prediction extend the agent-deployer thesis already logged against the 2026-04-30 Amazon note ([[2026-04-30-stratechery-amazon-earnings-trainium-commodity]]): if the "messy middle" of enterprise production workloads gets squeezed by both barbell ends rather than transformed in place, that's a live risk case against RDCO's own MAC bet (SMB-layer agent deployment), which is explicitly betting on that middle segment persisting and being reachable via commodity-cost infrastructure. Worth a gut-check pass on that assumption rather than pure reinforcement.

Related

Sponsorship

Not applicable — no sponsor block in this issue. Only self-promotion present: intro plug for the Thursday Sharp Tech episode and the standard closing Stratechery Plus podcast-suite cross-promo footer. No bias implication beyond ordinary in-house cross-promotion.