The End of Rationing

The era of rationing engineering time is over: coding agents like Claude Code have made building software extremely cheap while running AI-driven features imposes real marginal costs per user. Historically software was expensive to build and cheap to run, which produced doctrines like MVP, "do one thing," and product primitives that became universal - GitHub’s pull request workflow and Salesforce’s lead/opportunity/pipeline model are examples. With tokens and model inference driving per-user COGS, R&D fixed costs fall but operating margins compress; AI product gross margins sit well below classic SaaS levels. The result should be radically different products and experiments, but most offerings are still conservative rewrites of 2015-era patterns.

The gap reflects cultural, commercial, and organizational inertia: entrenched pre-AI playbooks, procurement that buys categories, venture capital favoring "X for Y" clones, and incumbents that effectively ship org charts. Enduring constraints - distribution moats, consumer trust, maintenance complexity, and token costs per user - slow change. Catch-up requires redesigning core product primitives for human-plus-agent workflows rather than bolting AI onto old interfaces, and embracing broader packaging and bundling to aggregate value and own customer relationships. Early signals include bundled platforms like Rippling and PostHog and strategic moves such as Stripe’s OpenRouter and Ramp’s LLM router; advantage accrues to teams that never learned to ration.