We just published our deepest research piece yet: State of AI Monetization, a strategy paper for product, technology, and finance leaders, with every price verified against primary sources. It runs twenty-five minutes.
This is the six-minute version.
Five findings, what they mean, and where to start. When a claim below makes you want the receipts, the full paper has the tables, the break-even math, and forty-plus sources.
Five findings
01The seat is being demoted, not killed.
Pure seat-based pricing fell from 21% to 15% of software companies in a year; hybrid pricing — a committed base plus a usage or outcome meter — jumped from 27% to 41%. But among established vendors adding AI, none abandoned seats entirely. The seat is becoming the unit of access; the meter is becoming the unit of value. Outcome pricing (Intercom's $0.99 per resolution, Zendesk's ~$1.50) is real but rare: 5% of companies today, 25% expecting it by 2028.
02Margins are the forcing function.
AI features carry real marginal cost, and classic SaaS pricing assumed there was none. AI-native gross margins run 25–60% against SaaS's 80%+; scaling AI companies spend roughly 23% of revenue on inference; Anthropic disclosed that some $200/month Claude Code users were costing it tens of thousands a month. Flat pricing plus power users is a loss machine — which is why everyone is installing meters.
03Intelligence is getting cheaper and dearer at once.
Constant-capability token prices fall 10–50x per year, and DeepSeek's flagship now costs $0.87 per million output tokens. Meanwhile frontier list prices are rising — GPT-5.5 launched at 4x GPT-5's price, and premium tiers reach $50 per million. The ~180x spread between those poles is the single biggest cost lever in AI products: route easy work to cheap models, reserve the flagship for queries that earn it, cache everything you re-send, batch everything that can wait.
04Rent-vs-own is a utilization question, not a hardware question.
A fully utilized owned GPU node produces output tokens for pennies per million — an order of magnitude below even budget APIs. But utilization is everything: halve it and costs double, and real self-hosting carries 3–5x hidden costs in labor and waste. The rule of thumb: owning beats hyperscaler on-demand above ~20% sustained utilization; against neoclouds and budget APIs the bar is far higher. Default to APIs, rent to prove utilization, buy for the baseline, burst for the peaks.
05Metering just became platform infrastructure.
In sixteen months, Stripe bought Metronome (the engine behind OpenAI's and Anthropic's billing), Salesforce signed for m3ter, and Zuora — which had bought Togai — went private at $1.7B. The market has decided: you don't build metering, and you don't rip out your billing stack either. You put a metering layer in front of the systems you already run, and you make usage data finance-grade — because under ASC 606, prepaid AI credits are a liability until consumed, and your usage pipeline is now a revenue-recognition system whether finance has noticed or not.
The unit of software value is no longer the person with a login. It is the work the machine did — and the margin question is whether you can meter it.
Where to start
Days 1–30: instrument one canonical usage event stream and produce your first per-customer AI margin report, however ugly. Days 31–60: pick a metering layer (buy, not build), run it shadow-mode, and agree your credit accounting policy with your auditors. Days 61–90: turn on routing, caching, and per-customer budgets, then launch the repriced offer to a cohort — grandfathered and dashboard-first.
The full paper walks through each step, the pricing tables behind the numbers above, the on-prem TCO math, and the operating model that makes Product, Tech, and Finance one pricing team.