AI Business Models.

Consumer AI subscriptions look like SaaS freemium, and they convert like it too: ChatGPT turns roughly 5–6% of its weekly users into subscribers³. What's different is everything around that conversion. Free users cost real money to serve, subscribers churn faster, and the heaviest users can cost more than they pay. The model works when two things are true. Paid tiers sell more usage, not a smarter model, so free users get a taste of the best experience and pay when they hit a limit. And something brings people back between bursts of use, so subscribers don't cancel in quiet months.

Every free user runs up a compute bill

In SaaS, an extra free user costs close to nothing. In AI, every prompt from every free user runs on GPUs¹. OpenAI has worked to cut the cost of serving even logged-out ChatGPT visitors, reportedly halving its inference cost for them². The free tier becomes a marketing expense you pay every month, and routing simple prompts to cheaper models is one of the main ways companies keep it affordable¹. Fraud makes it worse. Elena Verna, who leads growth at Lovable, points to "bots, stolen accounts, and people abusing free tiers"¹², and in AI each of them runs up a real compute bill.

Conversion holds up at 5–6%

OpenAI reported more than 50 million consumer subscribers against more than 900 million weekly ChatGPT users in February 2026³, roughly 5–6%. That sits at the top of the range for self-serve freemium SaaS, where Kyle Poyar, who writes the Growth Unhinged newsletter, and Lenny Rachitsky, who writes Lenny's Newsletter, put good at 3–5% and great at 6–8%⁴. App data points the same way. RevenueCat, which runs subscription billing for more than 115,000 apps, most of them mobile, found that AI subscription apps convert trials to paid at a higher median rate than non-AI apps (8.5% vs 5.6%)⁵. They also earn 41% more per paying user over the first year^5,6.

Keeping subscribers is the hard part

The same RevenueCat data shows AI subscription apps' monthly plans retain 36% worse over 12 months than non-AI apps⁶, with a higher median refund rate (4.2% vs 3.5%)⁵. Data from 4,029 software companies collected by ChartMogul, which makes subscription analytics software, shows SaaS products get nearly three times as many plan upgrades as AI-native products⁷. Part of the reason is that usage comes in bursts. Vikas Kansal, who leads product for Google's AI subscription plans, describes a subscriber who writes ten reports in October and none in November, then cancels¹. The upside of that pattern is that people come back: AI-native products win back nearly 10% of churned customers⁷.

Google gates usage, not intelligence

When Google first put a paywall on Gemini, it used the classic SaaS play: a single $20-a-month Gemini Advanced tier for access to its smartest model. According to Kansal, two things went wrong. The free Gemini app was already capable enough that users asked why they should pay $20 a month, and the power users who did pay consumed so much compute that the unit economics were, in his words, "terrifying"¹. Google rebuilt the offer as Plus, Pro and Ultra tiers with rising usage limits and larger context windows, and found that gating usage intensity was a stronger lever than gating model intelligence¹. The catch is that people have to understand the limit. Verna points out that ChatGPT describes its tiers' usage as "Limited," "More," "Expanded" and "Maximum," and Anthropic sells its Max plan as 5x or 20x Pro usage. In her words, "Some companies promise 5x more usage without explaining 5x of WHAT"¹². Spending is concentrated at the top as well. Olivia Moore, a partner at the venture firm Andreessen Horowitz (a16z), found in US receipt data that the top 10% of US consumers who pay for AI account for about half of observed consumer AI spend, and the top 1% average $903 a month against a $25 median⁸.