Working with several AI-powered subscription apps recently, I started noticing this pattern. The very behavior you are trying to encourage — more usage, more exploration, more interaction — can now compress your margins if monetization and infrastructure are not designed in tandem.
AI is not just another product feature. It is infrastructure. So if you’re not modeling AI usage against ARPU, churn, and LTV before you ship it, you may be increasing engagement while quietly destroying your economics.
User engagement isn’t free anymore
Subscription businesses are structurally efficient. Or they used to be. Once you’ve built the core product experience, the marginal cost of serving an additional subscriber inside the app is typically close to zero, and the economics compound as you scale. (Thomas Petit talks about this in his blog on why hybrid monetization should now be the default for subscription apps.)
AI, on the other hand, disrupts that elegance. By introducing AI-powered features, you introduce variable cost at the feature level. Every time a user triggers an AI interaction, tokens are consumed, inference endpoints are called, and a third-party provider bills you for compute (the hardware resources used to make AI models work).
In short, your cost structure becomes inextricably linked to usage.
This creates a subtle but important tension — the same engagement you’ve worked so hard to increase now drives incremental cost. Higher engagement increases AI calls → more AI calls increase infrastructure spend. And unless revenue expands proportionally, your gross margin begins to shrink.
5 Ways to reduce AI spend in your app
So what does this mean? AI features are inherently a margin shrinker? No, not quite. AI means subscription businesses need to think more like cloud infrastructure businesses. Meaning usage is no longer just a growth metric, but also a cost driver.
1. Apps should buy AI infrastructure, rather than build it
I recently spoke with a portfolio Ops Manager working across several AI products and they described a familiar problem — the music generation API powering one of their apps became unstable, and suddenly even paying users couldn’t access the core feature. Complaints rose, reviews worsened, and monetization performance became harder to interpret.
This is what makes AI different from traditional app features. The question isn’t just whether users want the feature, it’s whether your infrastructure can deliver it reliably enough to support retention and revenue (without breaking your economics).
This is why I’d generally encourage thinking carefully about the infrastructure you choose. Training large datasets or creating your own models can make sense if you’re running a large-scale AI platform, but most subscription apps are far better served by using third-party APIs (e.g. Open AI’s ChatGPT, Google’s Gemini, Anthropic’s Claude) — particularly if you’re still experimenting with monetization and feature or product-market fit.
Running your own models introduces a number of challenges:
- GPU overhead
- DevOps complexity
- Model maintenance risk
- Fixed monthly burn (regardless of usage)
It’s a dangerous position to be in. So for many growth-stage subscription apps, using API-based foundation models makes more sense:
- Paying pe