Amazon Just Admitted AI Capex Now Outruns Its Own Cash Engine — A $220B Verdict on Hyperscaler Survival

Andy Jassy put the entire $20 billion capex increase on memory, not on NVIDIA silicon or new data centers. That framing matters: HBM4 stacks run roughly $500 each, HBM3E about $300, and HBM3 around $200, with SK Hynix controlling 50–55% of the market. Amazon's $169 billion in TTM property and equipment purchases — up $66.1 billion year over year — implies hyperscaler memory spend is now the swing factor in the entire AI supply chain, not the accelerators themselves.

With TTM FCF of -$7.6 billion, Amazon is no longer the pristine cash machine of the past decade. The 18.2-to-negative swing is the largest absolute FCF reversal of any hyperscaler to date, but the directional signal was telegraphed: Oracle posted a -$23.7 billion TTM FCF on $55.7 billion of FY2026 capex, and Alphabet just printed -$5.85 billion in Q2 alone on a raised $195–205 billion 2026 capex guide. Microsoft (+$67B TTM) and Meta (+$41B TTM) still print positive FCF, but Meta is the laggard in the comparison set and Microsoft is one missed Azure quarter away from joining the negative club.

Amazon reported AWS backlog of $496 billion, up triple-digits year over year, on a $169 billion annualized run rate. Jassy reiterated that 2026 AI demand already outstrips capacity, that 2027 will look the same, and that 2028 demand is 'striking.' The 3.5–4x backlog-to-revenue ratio at AWS is the most concrete evidence that current capex is pre-funding contracted revenue two to three years out, and Jassy's own rule of thumb — data centers require capital roughly two years before servers generate revenue — means the 2026 $220 billion lands in the P&L as AWS revenue around 2028.

Power capacity doubles by end-2027, which is the only physical constraint left. The risk for shareholders is not that the AI capex cycle is over — it is that the gap between capex and cash generation is being closed with mark-to-market gains on private AI investments, not operating cash. The $53.4B Q2 non-cash pre-tax gain came from revaluing Amazon's stake in Anthropic; that stake is now worth more than $70 billion on a roughly $13 billion invested cost basis, with a potential 2026 IPO on the horizon.

Every dollar of the $20 billion capex increase flows to a small set of beneficiaries. SK Hynix reported Q2 2026 revenue of 79.3 trillion won (+257% YoY) and started HBM4 mass shipments in the same quarter, with operating profit of 60.5 trillion won. Micron is sold out of HBM for 2026, guided fiscal Q3 revenue to roughly $33.5 billion at an 81% gross margin, and is targeting 20–25% HBM market share by 2028. Samsung Electronics is the swing supplier: qualifying HBM4 to NVIDIA and now absorbing the HBM3E surge SK Hynix cannot fully meet.

Q2 net income of $62.6 billion looks historic, but $53.4 billion of it is a pre-tax, non-operating, mark-to-market gain on Anthropic, reflecting the revaluation of a private AI company. Operating income was $27.5 billion, up 43% — a strong print, but $53.4B in gains means the quarter's bottom line would have been roughly $9 billion without Anthropic revaluation. This is structurally similar to Alphabet's pattern, where mark-to-market gains on private AI investments are doing increasing amounts of the work that operating margins used to do.