No Priors Ep. 78 | With AWS CEO Matt Garman

AWS succeeded early by offering developers standard compute, storage, and databases as on-demand services instead of forcing new programming models, lowering friction compared to competitors that tried to reshape how apps were built.

Garman notes that roughly 80–85% of workloads remain on-premise, with mainframes, tightly coupled enterprise systems, telco, and factory-floor workloads representing massive remaining opportunity and requiring more tooling and modernization support.

AWS designed Bedrock and its AI stack around three assumptions: customers won’t compromise on security, there will be many models (large and small), and enterprise data is the key IP that must never leak back into shared models.

AWS offers first-party Titan models alongside partners like Anthropic and Meta’s LLaMA, leans heavily into open source and open weights, and aims to let customers mix and match models without being locked into a single vendor or license.

Garman expects AI capacity to remain tight for years given fab, memory, and power lead times; AWS is investing tens to hundreds of billions in land, power, renewable energy, and its own Trainium/Inferentia chips while partnering deeply with NVIDIA.

He argues most companies won’t build frontier models; they’ll win by applying existing models to sales, manufacturing, customer support, and other workflows where AI drives measurable efficiency and new capabilities.

Drawing on his own failed startup, Garman advises founders to have a clear path to monetization, maintain flexibility in spending on training/infrastructure, and remember that running out of money—not lack of ideas—kills companies.