Amazon's $220B capex: memory prices set your cloud floor

Amazon reported Q2 on Wednesday and raised its 2026 capital expenditure plan by roughly $20 billion in a single quarter. The number that matters to anyone renting infrastructure isn't the $220 billion — it's the reason given for the increase. Higher memory costs. Not more data centers, not more demand. The same components got more expensive, and the largest buyer on earth ate it.

What actually happened

Per Amazon's earnings release, Q2 net sales hit $200.6 billion, up 20% from $167.7 billion a year ago. Operating income was $27.5 billion, up 43%. AWS did $42.2 billion, up 37% year over year — its fastest growth since 2021, and well past the ~31% analysts expected. Advertising added $19.8 billion, up 26%. Q3 guidance is $197–202 billion in net sales.

On the call, Amazon lifted 2026 cash capex to about $220 billion from roughly $200 billion. CNBC reports the hike was driven by higher memory costs. Andy Jassy's other line is the one worth pinning to the wall: even at $220 billion, Amazon expects it still won't have enough capacity to meet demand in 2026, and expects the same to be true in 2027.

Why AI infrastructure spend matters for your business

We've written about Alphabet's capex and Meta's. This one adds a specific mechanism. When a hyperscaler's spend rises because it's building more, that's a growth story. When it rises because DRAM and HBM got more expensive, that's an input-cost story — and input costs travel downhill. They land in instance pricing, in per-token pricing, in the SaaS tools that resell both, and eventually in the quote your vendor sends you at renewal.

Add the capacity comment and you get the second half. A supplier who is sold out through next year has no reason to discount. If you were planning your 2027 budget around "compute gets cheaper every year," the largest cloud provider just told you, on the record, that it can't build fast enough to meet demand two years out.

None of that means panic. It means three concrete things. Stop assuming your cloud line item shrinks. Model it flat at best, up 10–15% at worst, and make the automation you're building justify itself against that number rather than a fantasy discount. Know which of your workloads actually need a hyperscaler — a nightly ETL job and a document classifier do not need the same tier of infrastructure, and one of them can run somewhere much cheaper. Keep the model layer swappable, because token pricing moves with the same physics. An abstraction that lets you change providers in an afternoon is cheap insurance against a bill you don't control.

The cost floor for AI infrastructure is being set right now by memory fabs, not by cloud vendors. Budget accordingly.

Sources: Amazon Investor Relations, CNBC.