Amazon just posted its best AWS quarter and its first negative-cash quarter — in the same release
Amazon posted its best AWS quarter in years and its first-ever negative-cash quarter — in the same release. The stock is no longer a cash-return story; it’s a duration bet.
The one-paragraph read
AWS reaccelerated to +37% (its fastest in 18 quarters) while expanding operating margin to 39.4% — the opposite of what the "AI capex crushes cloud margins" bear case predicted. But trailing free cash flow turned negative as capex hit ~$169B, the headline $5.75 EPS is ~3× inflated by a one-time $53.4B mark on the Anthropic stake, and management raised 2026 capex to ~$220B. So the stock is no longer a cash-return story — it's a duration bet that ~$220B/yr of AI capex converts to AWS revenue at a good return before depreciation and rates catch it. At ~26× forward it's a normal Mag-7 multiple — you're not overpaying, but you're underwriting future cash flow that doesn't exist today.
The scorecard
One quarter, two opposite signals — the operating business against the cash statement.
Q2 2026 — the facts everything hangs on
| Metric | Q2 2026 | YoY | Note |
|---|---|---|---|
| Total net sales | $200.6B | +20% | Beat ~$196.8B est. |
| AWS revenue | $42.2B | +37% | Fastest in 18 quarters; beat ~31% consensus |
| AWS operating income | $16.6B | +64% | 39.4% margin — expansion, not compression |
| Advertising | $19.8B | +26% | Compounding underneath AWS |
| North America (retail) | $116.2B | +21% OI | $9.1B op income · ~7.9% margin |
| International | $42.2B | — | $1.7B op income |
| Consolidated operating income | $27.5B | +43% | The durable earnings number |
| Net income / EPS | $62.6B / $5.75 | vs $1.68 | ⚠ includes $53.4B one-time Anthropic mark; clean ~$1.82 |
| TTM free cash flow | −$7.6 to −$11.6B | from +$18.2B | ⚠ capex exceeds operating cash flow |
| TTM capex | ~$169B | +~64% | vs operating cash flow ~$161.4B |
| FY2026 capex guide | ~$220B | ▲ ~$200B | Raised — memory costs + AI |
| AWS backlog (RPO) | ~$496B | record | The "demand-backed" defense of the capex |
| Q3 2026 guide | $197–202B | +9–12% | Decel — OI midpoint ≤ Q2; below ~$204B Street |
The central tension
The bull and bear don’t disagree on the facts. They disagree on one thing — and it’s the whole investment.
- AWS reaccelerates AND expands margin off a ~$169B run-rate — the opposite of law-of-large-numbers deceleration.
- Capex is demand-backed: record ~$496B backlog, capacity sold out through 2027 (two customers tried to buy all of 2026’s Graviton).
- Custom silicon → margin, not just spend — Trainium framed as saving "tens of billions" of capex per year.
- Ads (+26%) and North-America retail (+21% OI) compound underneath.
- ~26× forward is mid-pack for the Mag-7 — a normal multiple, not a stretched one.
- Free cash flow is negative; capex (~$169B) now exceeds operating cash flow — funded partly by ~$25B of new debt.
- The $220B capex keeps rising on memory-cost inflation, with ROIC unproven.
- Earnings are flattered by a reversible $53.4B non-cash Anthropic mark.
- Circular financing: Amazon funds Anthropic (~$25B), Anthropic commits >$100B back to AWS.
- A depreciation wave from the build is set to hit AWS margins in 2026–2027 — not yet realized.
- Google Cloud grows faster (63–82% vs AWS 37%) and trains its own frontier model on TPUs.
The one question
Is negative free cash flow a timing artifact — Jassy's 6–24 month lag, where AWS lays out cash before it can bill, backed by the $496B backlog — or a structural sink, where memory-cost inflation and a depreciation wave mean the capex never earns its return? Everything else is downstream of that single call.
Valuation — a normal multiple on a temporarily abnormal cash profile
AMZN sits mid-pack in the Mag-7 — richer than Meta (~23×), cheaper than Nvidia. The after-hours pop made the forward multiple richer, not cheaper, and you're valuing on a P/FCF that currently doesn't compute.
| Multiple (on $235.50) | Value | Context |
|---|---|---|
| Trailing P/E | 28.2× | — |
| Forward P/E | 26.3× | range 25.5–29× across trackers |
| EV / EBITDA | 15.6× | — |
| Price / Sales | 3.3× | — |
| PEG | 1.20 | — |
| Price / FCF | n/a | free cash flow is negative |
| Mag-7 avg fwd P/E | ~27× | vs S&P ~21× — ~10% premium, lowest in a decade |
Post-print targets split hard, which is itself the signal: BMO raised to $360, while UBS cut to $305 and Mizuho to $320 (mean ~$315–319; 79 buys / 4 holds / 0 sells). The multiple is only "cheap" if AWS ROIC shows up before depreciation and rates catch it.
AI-compute position — Nvidia's biggest landlord and its silicon challenger, at once
The most nuanced part of the picture, and where the skeptic pass caught the most spin. On the "winning" side: Trainium3 is throughput-competitive — a 144-chip UltraServer ≈362 PFLOPs vs Blackwell NVL72 ≈360 — and Bedrock breadth de-risks single-model dependence (>100k customers run Claude; GPT-5.5 landed on Bedrock in April 2026, ending Azure exclusivity). AWS is supply-constrained, not demand-short: power is "AWS's single biggest constraint," capacity short through 2027.
On the counterweight side: the marquee OpenAI $38B AWS deal is Nvidia GPUs — not Trainium — so AWS is the landlord there, not the silicon winner. Trainium's frontier-training demand is still overwhelmingly Anthropic. And Google, not AWS, has the clean story — TPU 8th-gen trains Gemini, and GCP is growing 2×+ faster off a smaller base. Trainium4's roadmap even adds Nvidia NVLink Fusion — an implicit admission the near-term plan is coexistence with CUDA, not displacement.
Caveat (downgraded in review): The "Trainium at 30–50% of Nvidia cost" figure is AWS marketing that narrows vs H200 spot and disappears once you count the Neuron-SDK migration tax; the "Uber 50% savings" stat was misattributed — Uber's quote was about Graviton4 ride-matching, not training.
Risk — each isolated, do not blur
- The FCF / capex duration bet (the central risk). FCF is negative and capex (
$169B) exceeds operating cash flow ($161.4B) — the "94% of OCF" figure circulating is wrong; it's ~105%, which is why FCF flipped negative. Jassy frames the hole as a 6–24 month timing lag, backed by the $496B backlog. - Depreciation time-bomb (forward, not yet realized). Depreciation from the exploding data-center base is set to flow into AWS operating income in 2026–2027, compressing margins even as revenue grows. It has not materialized — Q2 margin actually expanded to 39.4%.
- Earnings quality / Anthropic circularity. A $16.8B non-cash Anthropic mark was ~half of Q1 net income — it reverses if Anthropic's ~$1T valuation compresses. The magnitude of the Amazon→Anthropic→AWS revenue loop is genuinely unknown (estimates span $1.9B–$10B); the existence is real.
- Guided deceleration (already visible). Q3 revenue guide +9–12% (half of Q2's +20%), operating-income midpoint flat-to-down sequentially, below Street.
- Rate / consumer sensitivity. Negative-FCF, long-duration, debt-funded growth is exactly the profile that de-rates if rates stay high; thin-margin retail carries tariff and recession exposure.
- Regulatory / FTC antitrust overhang. A standing tail risk — not a near-term catalyst on this print.
Time horizon & position sizing (kept separate)
Horizon. Not a next-quarter setup — Q3 is guided to decelerate and the stock already popped. The thesis pays off (or fails) on a 3–5 year window: does the 2026–2027 capex super-cycle convert to billed AWS revenue at ROIC above the depreciation drag? Under ~2 years you're exposed mostly to sentiment on the capex/FCF headline and the Anthropic mark's swings; at 5 years you're underwriting AWS ROIC vs depreciation vs rates — the actual bet.
Sizing considerations (not a recommendation). AMZN's swing factor is now the same AI-capex cycle that drives Nvidia, Microsoft, and Google — so it adds capex-cycle beta, not diversification, to an AI-infra basket. Reported earnings now carry mark-to-market swings on the private Anthropic stake, a new source of non-operating volatility. And with no FCF cushion today, a normal multiple on an abnormal cash profile argues for entry discipline over conviction-max sizing.
What would invalidate this view
The discipline: name in advance what would prove each side wrong.
The bull breaks if…
- AWS growth rolls back toward the 20s% while depreciation lands and compresses margin (the 2026–2027 test).
- Backlog (RPO) stalls or utilization disappoints — the "demand-backed" defense evaporates.
- The Anthropic stake marks down materially — hits earnings, exposes the growth circularity.
- Capex keeps rising without FCF inflecting positive by ~2027 — "timing lag" becomes "structural sink."
The bear fails if…
- FCF inflects clearly positive in 2027 as front-loaded capex bills out — proving the timing-lag framing.
- AWS holds high-30s% growth and margin through the depreciation wave — proving the capex earns its return.
- Trainium wins a second major frontier-training tenant beyond Anthropic — breaking the single-customer critique.
Verification ledger
A separate skeptic pass tried to refute every load-bearing claim against primary sources. Where it bit:
Confirmed against Amazon’s own filing: AWS +37% / 39.4% margin · negative FCF & $220B capex raise · the $53.4B Anthropic mark and $27.5B durable operating income · power as the binding constraint · the $496B backlog duration bet.
Not found in the cited sources; public estimates of Anthropic’s 2026 AWS revenue span ~$1.9B–$10B. The circularity is real; its magnitude is unknown.
A Q1 2026 call quote spliced onto the Q2 39% margin — still management’s own framing, but temporally misattributed.
Throughput parity holds; the cost figure is AWS marketing that narrows vs H200 spot and disappears with the Neuron-SDK migration tax. The Uber savings stat was about Graviton4, not training.
Capex exceeds operating cash flow — that’s precisely why free cash flow went negative. A widely-repeated error.
Open questions
What is Anthropic’s actual 2026 AWS revenue contribution? Estimates span ~$1.9B–$10B — the circularity’s magnitude is genuinely unknown.
When exactly does the depreciation step-up hit AWS operating margin, and how big? The whole bear case hinges on timing not yet disclosed.
The precise TTM FCF figure — sources split −$7.6B vs −$11.6B on differing capex definitions. The most decision-relevant number has the least source agreement.
Sources
- 1Amazon Q2 2026 press release (Investor Relations) · SEC ex-99.1
- 2Investing.com — "AWS surges 37%, FCF turns negative" (2026-07-30)
- 3CNBC — Amazon / AWS Q2 2026 (2026-07-30)
- 4Yahoo Finance — Amazon Q2 2026 AWS (2026-07-30)
- 5TradingView / Wall St Engine — rev / OI / NI breakdown
- 6Variety — ad sales / profit / Anthropic (2026-07-30)
- 7stockanalysis.com — AMZN statistics (multiples, FCF)
- 8Benzinga — analyst price targets (2026-07-28)
- 9Morningstar — "$200B capex overshadows results"
- 10Fortune — Jassy on capex / demand / capacity (2026-07-30)
- 11DataCenterDynamics — customers try to buy all Graviton capacity; 3.9GW
- 12GeekWire — "AWS booming but FCF turns negative"
- 13TheNextWeb — Q1 2026 $16.8B Anthropic mark ≈ half of net income
- 14TechCrunch — Anthropic $5B / $100B AWS pledge (2026-04-20)
- 15CNBC — Amazon to invest up to $25B in Anthropic (2026-04-20)
- 16SemiAnalysis — AWS Trainium3 deep dive
- 17Spheron — Trainium3 vs Nvidia H200 / B200 (2026)
- 18Motley Fool — Mag-7 valuation lowest in a decade (2026-07-16)
- 19Investing.com analysis — "AWS growth must justify $200B AI spend"
- 20tech-insider.org — Google Cloud growth [low-confidence secondary]
- 21Google Cloud — AI infrastructure at Next ’26 (TPU 8th-gen)
- 22TradingKey — Q2 recap, price / multiples (2026-07-30)
- 23TheNextWeb — Uber + AWS Trainium (2026-04-07)
- 24Anthropic — Amazon / Trainium partnership
How this was made
Every Market Storm report is produced by STORM — a multi-agent research method adapted from Stanford's STORM. Several AI agents each take a different stake (a fundamentals analyst, a short-seller, an industry engineer, a valuation watcher), interview each other while grounded in live web search, and surface the load-bearing claims. A separate skeptic pass then tries to refute each of those claims against primary sources. What survives is written up here — with the caveats it earned. The finance is the payload; the method is the point.
Get the next Market Storm in your inbox
One email when a real market catalyst triggers a new report — the method, the numbers, and what the verification pass caught.