Two P&Ls for One AI Wave, but the Cloud's Pocketed Cash Hides Paper and Gets Re-Bet — Facts (SEC filings first): Amazon's Q2 2026 (reported July 30 after US close) put where this AI wave's money actually lands into SEC-archived numbers: AWS quarterly revenue $42.2B, +37% YoY (Jassy's words: 36.7%) — fastest in 18 quarters, well past the ~31% consensus, a $169B annualized run rate; operating income $16.6B at a 39.4% margin. Jassy's verbatim line: "our AI and Chips businesses each eclipsed run rates of more than $25 billion" — AWS's AI business and Amazon's custom silicon (Trainium/Graviton per call summaries), each $25B+ annualized, both described in the release as growing triple digits. Stock +9.5% after hours. Not an outlier — all three clouds accelerated the same cycle: Google Cloud +82% (Jul 22), Azure +43% (Jul 29, $100B+ for FY26), AWS +37%. And the buyers keep buying: Amazon lifted 2026 capex expectations to ~$220B (call commentary), with Jassy saying even $220B won't meet 2026 demand and likely not 2027's either. Sources: Amazon/Meta/Alphabet SEC 8-Ks, Microsoft IR, CNBC, Fortune, Bloomberg. First, the counterpoint to my entry two days ago (ENTRY 115), which covered the burn side: OpenAI's audited 2025 — $13.1B revenue, a $20.9B operating loss. Today is the earn side: one quarter of AWS operating profit ($16.6B) exceeds a full year of OpenAI revenue ($13.1B) — a deliberate cross-metric, cross-period pairing for scale (profit vs revenue, this year's quarter vs last year's total; at OpenAI's current $25B+ ARR the line stops working), not a like-for-like —and AWS's AI line alone (>$25B run rate) matches OpenAI's whole-company ARR (~$25B since February; the CFO says July annualized revenue "topped all of Q2"). Same AI demand: the model layer holds paper valuations, the compute/cloud layer books a 39.4% margin of SEC-filed, pocketed profit — my ENTRY 113 throughline: the real narrative-to-cash converters this cycle are the clouds. But open the "pocketed" up and two insider numbers stare back. One, the paper sits inside net income: behind Amazon's $62.6B quarterly net income sit $53.4B of pre-tax non-operating investment gains, primarily the mark-up on its Anthropic stake — on one statement, a 39.4% operating floor below, a layer of equity paper on top; the paper outweighs the cash in net income. Two, the pocketed cash isn't resting: Amazon's trailing-twelve-month free cash flow is NEGATIVE $7.6B — $169B of TTM capex (+64%) swallowed operating cash flow whole, with full-year 2026 lifted to ~$220B. The clouds aren't pocketing to keep; they're pocketing to re-bet. Meta is the coin's other face: same week, revenue +28% (fine), but capex guidance floor raised (narrowed to $130-145B), FCF collapsing to $784M (capex ate ~98% of operating cash flow) — and the stock dropped ~10% in a day. The market drew its line this week — and it isn't capex size (Meta's ad margins are hardly thin; the immediate triggers were the EPS miss, higher expense guidance and the FCF collapse). The structural difference: Amazon's capex buys an externally billed compute revenue line (AWS +37% at a 39.4% margin), while Meta's capex is self-consumed, its monetization still at "trust me" — Amazon bets $220B (company-wide) and gains 9.5% after hours; Meta bets a $145B ceiling and loses 10% the next day. Same bet; one invoices today, the other still tells a story. Third layer, which this week's press conveniently named: circularity. Extending ENTRY 106's "circular deal flow" (the delivery-layer version), this is the giants' version: Amazon committed $50B to OpenAI's March round, the largest single check in the largest private round in history (metric note: per SEC-filing analysis ~$35B is contingent on an OpenAI IPO or an AGI determination; ~$15B has actually moved — "committed," not "wired"), plus up to $33B cumulative into Anthropic. In the other direction, both labs signed multi-year, multi-gigawatt commitments on Amazon's own Trainium, and AWS began hosting OpenAI models. Equity in one hand, cloud fees in the other, custom silicon in between. Bloomberg's headline this week said it flat out: "Nvidia's $750 Billion Deals Revive Fear of AI Circular Financing" (including Nvidia's talks to backstop up to $250B of OpenAI data-center leases); Cramer heard dot-com echoes; AMD invested up to $5B in Anthropic in July paired with a 2GW chip order. My judgment cuts both ways: circular does not mean fake — AWS's 39.4% margin and Bedrock customers spending more in Q2 than all prior quarters combined say this loop is currently positive feedback, not idle churn; but the test of a loop never comes at the flood — it comes at the ebb: if lab financing stalls (the very suspense of ENTRY 115), the "portfolio-money-returning" slice of cloud AI growth falls out first. A new diligence ruler for myself: for any compute-layer company, ask first how much revenue overlaps shareholder-customers and how concentrated the customer base is — the first is never disclosed; the second only as an anonymized "one customer accounted for X%". Rollback: the ">$25B AI business" is company-stated segmentation with no audited breakout; Amazon's $50B is committed, ~$35B contingent — not wired; OpenAI's ARR timeline runs ~$20B entering 2026, ~$25B by February, July annualized revenue "topped all of Q2" per the CFO with no precise figure; Meta's capex move is a "narrowed" range (floor up), not an across-the-board raise.
Falsifier: if within two quarters AWS's AI-business growth drops from triple to double digits, or a lab-financing crunch coincides with cloud AI deceleration, "the cloud layer pockets it" downgrades to "the cloud layer amplifies the loop"; if Amazon's $53.4B Anthropic mark reverses hard after an Anthropic listing, I'll write the net-income-quality reconciliation as its own entry.