a16z · Oct 24, 2026 · 1h 1m · 53m saved
With Andy Jassy, CEO, Amazon
Photo via amazon.com
Amazon is spending roughly $220 billion on capital expenditure in 2026, and Andy Jassy says that still isn't enough to cover what customers want to buy. Markets have openly wondered whether AI infrastructure spending can even be financed. His own numbers argue the opposite problem: AWS has sold out 2026, mostly sold out 2027, and already committed a chunk of 2028.
"Even spending $220 billion of capex this year, we're leaving a meaningful amount of demand on the table."
Andy has run Amazon since 2021 and built AWS before that, the business whose generative-AI unit he says has already reached a $25 billion run rate in three years.
The full interview is covered here so you can skip it. 61 minutes of audio, 8 minutes of reading.
Here are the 7 takeaways that matter.
Key Takeaways
Amazon expects to spend about $220 billion in capex in 2026 and is still leaving demand on the table, with 2026 capacity sold out, 2027 largely sold out, and part of 2028 already committed
Data-center buildings get committed roughly two years before they pay off but last about 30 years, while servers break even in about three years and last five to six, which Andy says creates cash-flow "headwinds" during a fast buildout
AI demand looks like a "barbell," with AI labs consuming "gobs and gobs" of compute on one end and enterprises mostly using AI for cost avoidance on the other
Inference costs have fallen but are still "an order of magnitude too expensive" for pervasive enterprise use, he says
Golf courses nationally use 200 times more water than all of Amazon's data centers combined, Andy says, and backup generators actually run only about 1% of the year
He blames rising electricity rates on decades of underinvestment in the grid, not data centers themselves, and says the U.S. "cannot afford not to build data centers" or cede AI infrastructure leadership to China
The 2015 Annapurna Labs acquisition, bought to offload work from EC2 servers rather than for AI, became the base for Graviton and Trainium chips and helped AWS's AI business hit $25 billion in run-rate in three years
Amazon's investments in Anthropic and OpenAI reflect belief in both companies long-term, plus a jointly built "stateful runtime environment" for agents inside Bedrock
Andy predicts companies will have to rewrite almost all of their code to be "agentically capable," with fewer humans doing today's work in smaller, cross-functional teams
He says the more urgent AI risk isn't speculative doomsday scenarios but cybersecurity, since AI finds far more vulnerabilities than humans did and patches will need to deploy in minutes
Amazon chose to fully rebuild Alexa's "brains" for generative AI, a call Andy calls obvious, while retail and Prime Video are still being reimagined from scratch
Andy says AI doesn't replace critical reasoning, even as he tells students not to fight using it, because human judgment is still needed to turn ideas into a plan
1. $220B still isn't enough
Andy Jassy told a16z's Erik Torenberg and Ben Horowitz that Amazon expects to spend about $220 billion on capital expenditure in 2026, and that even at that level the company is turning away business it could otherwise book.
AWS's 2026 and 2027 capacity is already sold out, and part of 2028 is committed too
"26 is obviously sold out. 27 is largely sold out. A good chunk of 28 is sold out." — Andy Jassy
Andy explained the mismatch between AWS's cash cycle and its spending: data-center buildings are committed roughly two years before they generate revenue but have useful lives of about 30 years, while servers and networking gear are ordered only months ahead, break even in about three years, and last five to six years. That timing gap, he said, is why free cash flow shows "headwinds" during a fast buildout even though the investment is expected to pay back.
