Amazon just posted its best AWS quarter and its first negative-cash quarter — in the same release

Amazon's cloud business had its best quarter in four years — and in the same release, the company burned more cash than it earned for the first time ever. Both are true, and they are the same story: AWS is growing fast because Amazon is spending roughly $220 billion a year building the data centres that run it. The bet is that the spending turns into billed revenue before the accounting catches up with it. The scorecard below is that argument in numbers — the operating business on one side, the cash statement on the other.

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.

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

1. 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.
2. 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%.
3. 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.
4. Guided deceleration (already visible). Q3 revenue guide +9–12% (half of Q2's +20%), operating-income midpoint flat-to-down sequentially, below Street.
5. 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.
6. 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.