Amazon faces a 'fundamental' shift in how AI startups spend on the cloud, internal documents show (www.businessinsider.com)

Internal Amazon documents obtained by Business Insider warn that AI startups are shifting first-dollar spend away from traditional AWS services toward models, GPU-backed training/fine-tuning, inference, and AI-as-a-service developer tools. The papers — marked “Amazon Confidential” from March and July — say founders increasingly buy API access to models (OpenAI, Anthropic) and “neocloud” GPU providers (e.g., CoreWeave) before adopting core cloud infrastructure, making early spending less sticky. AWS metrics cited include YC cohort usage falling (59% used >3 AWS services, down from 2022), heavy adoption of OpenAI/Anthropic (88%/72%), and very low take-up of AWS Bedrock (4.3%). Cursor is highlighted as spending <10% on traditional AWS infra versus newer AI categories. The shift matters because GPU-centric workloads and pay-as-you-go GPU slices alter vendor lock‑in dynamics and prioritize specialized hardware/software over CPU-based cloud staples, threatening AWS’s historical startup funnel and allowing neoclouds and rivals to capture early customer relationships. AWS growth (18% YoY) has trailed Google and Microsoft (>30%), and neocloud revenue has surged; internal docs flag pricing and small‑increment GPU access as weaknesses. AWS disputes the conclusions, citing recent wins, Anthropic partnership and price cuts (e.g., 45% off some EC2 Nvidia GPU instances), but the documents underline a broader “Cloud 2.0” realignment around GPUs, model APIs, and vertical AI stacks.