06-reference

stratechery google frontier case amazon earnings

2026-08-05·reference·source: Stratechery·by Ben Thompson
stratecherygooglealphabetamazonawsanthropichyperscaler-capexai-infrafrontier-modelsharness-thesis

"Google Earnings, The Frontier Case, Amazon Earnings" — @BenThompson

Why this is in the vault

Third earnings-season Update in a row (after the Meta and Microsoft pieces), and the one that actually closes the loop on capex-justification logic: Jassy's AWS remarks are the clearest public explanation yet of why hyperscalers argue their AI capex pencils out, and Thompson's Google section adds a live, sourced theory (Pichai's "bridge deal" comments) that Google is renting Anthropic compute demand at scale — both directly load-bearing for the RDCO capex-anchor tracking loop.

The core argument

Three linked sections in one Daily Update.

Google earnings. Alphabet raised 2026 capex guidance to $195-205B (from $190B) alongside Cloud revenue of $24.77B (vs. $22.46B expected), up 82% YoY (vs. 63% last quarter, 32% a year prior) at 36% margins; Cloud backlog grew to $514B from ~$460B a quarter earlier. Stock initially fell over 6% on the capex number, then recovered to roughly 10% above pre-earnings levels. Thompson's theory: Anthropic sought capacity with high commitment willingness, Google allied with Anthropic to hedge its own coding/agentic weakness, made a multi-year compute commitment at high prices, starved its own research teams of compute to serve it, and is now paying "bridge" compute costs until new capacity comes online — evidenced by Pichai's earnings-call description of an unnamed large customer deal as "highly ROI-positive... in the lifetime of the deal."

The Frontier Case. Pichai touted an upcoming "Gemini 4," but the actual recent ships were to the Flash line — the models that protect Search, not chase consumer-chat share. Thompson reads this as the correct framing: CBO Philipp Schindler said Gemini "supercharges" ad-matching and "AI Max" campaigns "unlock... billions of net new searches that weren't really monetizable before." Thompson holds this up as the frontier-investment justification he'd wanted from Meta (referencing his own prior "A Script for Mark Zuckerberg" piece) — Google gets extra credit because it also monetizes spare compute by selling it, which Meta can't.

Amazon earnings. AWS revenue hit $42.2B, up 37% YoY (vs. $40.6B expected) — the fifth straight quarter of accelerating growth, fastest since Q4 2021 — and the stock had its biggest surge since 2012. Full-year 2026 capex guidance rose to $220B from $200B, with CEO Andy Jassy attributing the increase specifically to higher memory costs. Jassy's prepared remarks (Thompson compares their industry importance to Box's 2014 S-1 popularizing SaaS cohort economics) laid out the capex logic plainly: data centers take ~2 years to build then generate revenue for 30+ years without repeat capital; servers/networking break even in ~3 years on 5-6 year useful lives, mostly under 5-year contracts. Jassy: "we will still not have enough capacity to meet all the demand we have in 2026," expecting the same into 2027. Thompson's read: the market absorbed an extra $20B in component costs without punishing the stock, and Google's own post-earnings recovery is a second data point for the same phenomenon — the capex-justification narrative is being bought broadly, not just at Amazon.

No sponsor block in this issue. Self-promo is limited to the standard Stratechery Plus podcast-suite footer and an opening aside about Thompson "vibe-porting" an app, discussed on that day's Dithering episode.

Mapping against Ray Data Co

Two concrete, actionable pieces for the capex-anchor work. First, Amazon's guidance bump is explicitly attributed to memory-cost inflation, not new build volume — this is a direct external data point for the 01-projects/investing/theses/2026-05-18-memory-cycle-v1.1.md thesis and its backtest work: a hyperscaler naming memory pricing as the marginal driver of a capex raise is exactly the kind of anchor evidence investing-edgar-watch should be flagging when it diffs quarterly 10-Q capex deltas, distinct from the pure accounting-policy shift (useful-life reclassification) logged in the 2026-08-04 Microsoft note two days prior — worth cross-referencing both in the next capex-anchor pulse so the two effects (real memory-price pressure vs. reporting-standard change) don't get collapsed into one "capex accelerating" headline.

Second, the Google/Anthropic "bridge deal" theory is a live instance of the harness-thesis vocabulary already tracked in this vault (see the 2026-08-04 and 2026-05-06 Stratechery notes): Google is functioning as compute-landlord to a frontier-model lab it's simultaneously hedging against, which is a different competitive posture than Microsoft's harness/middleware position or Meta's own-frontier bet. This doesn't change the standing memory-cycle thesis but adds a third data point (alongside Microsoft's ASC 842 shift and Meta's earnings-day slide) for how differently each hyperscaler's earnings narrative is being read by the market this cycle — genuinely new information, not restated capex totals, so it clears the bar for filing rather than being a pure "AI capex is still going up" restatement.

Related

Sponsorship

Not applicable — no sponsor block in this issue. Only self-promotion present: standard Stratechery Plus podcast cross-promo footer (Sharp Tech, Dithering, Sharp China, Greatest Of All Talk, Asianometry) and an opening personal note about a side project discussed on that day's Dithering episode. No bias implication beyond ordinary in-house cross-promotion.