Amazon’s Durability

The Stratechery piece makes a structural argument that Amazon's apparent lag during the model training race was actually a form of optionality, and that the inference phase rewards exactly the operational discipline and infrastructure depth Amazon has been building for two decades. The buried point is that Amazon may not need a frontier model to win the enterprise AI market.

This connects directly to two threads we've been tracking. The May 1st piece 'AI Demand Is Outpacing the Scaffolding to Support It' identified infrastructure operationalization as the real bottleneck, which is precisely the gap Amazon's position is designed to fill. Separately, the Pentagon coverage from The Verge and TechCrunch around the same date confirmed AWS as one of the vendors awarded classified AI infrastructure contracts, suggesting Amazon's reliability credentials are already translating into high-stakes procurement wins. The $725 billion capex story from The Decoder adds useful context: at that spending scale, inference efficiency and cost-per-query matter enormously to buyers, which tilts the competitive advantage toward operators over model builders.

Watch whether AWS's share of enterprise inference workloads, as reported in quarterly earnings, grows relative to Azure and GCP over the next two reporting cycles. If it does, the structural thesis here has real numbers behind it; if Azure continues to pull ahead on AI revenue, Microsoft's OpenAI integration is outweighing Amazon's infrastructure depth.