This is an opinion essay on market structure, not a technical result. The claim: AI infrastructure won't follow classic cloud's playbook. Four arguments. Classic cloud served web workloads where deployment was the whole job, while ML front-loads massive compute before anything ships — yet most app builders consume models purely via APIs and never touch GPUs. Concentration inverts: classic cloud had three landlords and everyone as customer; AI infra has fragmented providers (hyperscalers plus GPU neoclouds) but a concentrated buyer base — a handful of training labs, model builders, and open-weight inference providers. PaaS wins over IaaS this time because enterprises have no legacy GPU-ops competence, and closed models are only available API-first — structurally Twilio, not EC2. And buyer requirements are immature: the supply crunch makes "we have GPUs" the entire sales pitch, evidenced by CoreWeave drawing over half its revenue from Microsoft at IPO. Actionable takeaway for practitioners: unless you're training, fine-tuning, or self-hosting, API consumption is the sane default, and GPU ops is a niche competence. The evidence is mostly assertion beyond the CoreWeave figure. The sharp tension: point four concedes current dynamics come from a cyclical supply crunch — if requirements sophistication "always comes," the divergence may be less structural than the framing claims.