Amazon Added $20B to Capex Because Memory Got Expensive, Not Because It Bought More

The number looks familiar. Amazon lifted its 2026 cash capital expenditure plan from about $200 billion to $220 billion. Twenty billion more.

The reason is the unfamiliar part. It is not that Amazon bought more servers. The company pointed to memory prices coming in higher than expected (Data Center Knowledge, Calcalist).

Same Capex, Different Meaning

For two years, capex headlines read as good news for founders. The more data centers hyperscalers build, the more supply there is, and more supply pushes inference prices down. That is roughly what happened. The capex war piece from July landed in the same place: you cannot win on capital, but you can compete on top of cheaper tokens.

This increase touches the premise underneath that logic. If spending rises and buying power rises with it, that is added supply. If the same hardware simply costs more, supply is flat and only the bill moved. Amazon’s stated reason sits closer to the second.

The capacity outlook points the same way. Andy Jassy said the company will not have enough capacity to meet all of its 2026 demand, and expects constraints to persist through 2027. Demand is already booked into 2028. That came in the same release where AWS posted its fastest growth in eighteen quarters.

What the $1.1 Trillion Is Made Of

Amazon, Google, Meta and Microsoft have together put more than $1.1 trillion into AI infrastructure since 2023. Another $700-plus billion is expected to land in 2026 alone. Tallies differ, ranging from about $725 billion to $745 billion depending on what gets counted, but either figure is a long way up from roughly $410 billion in 2025 (Tom’s Hardware).

Read as a growth rate, it looks like demand exploding. But if part of the increase came from component pricing, as it did at Amazon, the total starts describing the invoice rather than the capability. Those are different stories.

Nothing to Share on Funding

Asked how Amazon is paying for the extra $20 billion, Jassy said he had nothing to share (Benzinga).

One sentence from one company is not much to build on. Still, whether spending at this scale stays inside operating cash flow or moves onto the balance sheet eventually reaches customers. Once financing costs attach to infrastructure, cloud providers recover them somewhere, and the usual somewhere is long-term commitments and unit pricing.

What Founders Should Recalculate

Three things show up in practice.

First, do not hard-code falling inference prices into the budget. They may keep falling. This quarter’s signal pointed the other way once.

Second, if capacity stays tight through 2027, now is a reasonable time to check how tied you are to a specific instance type. A team that can swap models and a team optimized around one accelerator have very different leverage in this stretch.

Third, memory being the bottleneck is a different kind of information depending on where you sit. Korean memory makers are on the supply side of this cycle. For teams working in hardware or anywhere along the semiconductor value chain, this increase reads as demand news rather than cost news.

The same number inverts depending on which side of it you stand on. Which is why the source of the increase is worth checking before the total.