I sat down with Werner Vogels, the CTO of Amazon

Summary

i sat down with werner vogels 30 Aug, 2025 a few days ago, i had the surreal experience of joining a private fireside chat with werner vogels, the cto of amazon, during startup summit 2025 in florianópolis. it's one thing to follow his talks online – it's another to be in the same room, listening to him unpack two decades of lessons from building some of the most critical infrastructure on the internet. werner is sharp, real, and doesn't sugarcoat. what follows is a distillation of the ideas that hit me hardest. solve problems, not hype werner didn't flinch: "there are no magic technologies." he recalled how, five years ago, everyone wanted to "do something with blockchain." now, it's generative ai. his point was simple – you should be driven by a problem, not by fomo. most use cases asking for a blockchain can be handled by a simple immutable ledger. most uses of gen ai? not needed. don't chase tech for its own sake. plan for failure, not perfection at amazon's scale, every 9 you add to availability costs millions. so the real art lies in deciding what can fail. classify your systems: tier 1 (can never go down – checkout, search), tier 2 (can degrade a bit), and tier 3 (can disappear and no one will care). not everything needs 99.999% uptime. trade-offs are the game. reversible vs. irreversible decisions this one's gold. according to werner, good decision-making starts with this split: reversible decisions (two-way doors): move fast. 70% of the info is enough. irreversible decisions (one-way doors): slow down. think deeply. you don't need perfect data to start – unless you're making a bet you can't walk back from (like a public api or a core product shift). security > operations > cost this was his framework for tech leadership: 5. ai gives us time back – that's the real magic security: never an add-on. bake it in from day zero. operations: software lives longer in production than in dev. build for maintainability. cost: design not just for scale-up, but for scale-...