Replit's gross margins went from 36% to negative 14% in two months.

Same product. Same pricing. Same team. The only thing that changed: they launched a more autonomous AI agent that consumed more LLM resources than their pricing covered.

Traditional SaaS has 70-80% gross margins because one more subscriber costs almost nothing. AI products pay for compute on every prompt. Your best users are your most expensive users. That single fact breaks every pricing model designed for the SaaS era.

I mapped pricing across the top 50 AI startups by valuation with Moe Ali. Six patterns emerged.

The scariest finding: in most AI products, the P90 user costs 10-40x more than the P50 user. Both pay the same subscription. You're subsidizing your heaviest users with revenue from your lightest ones. And that subsidy grows as power users discover more ways to use the product.

Cursor learned this the hard way. They switched from flat 500 requests/month to a credit pool system. A developer burned the entire monthly allocation in a single day. $7,225 invoice. The CEO published a public apology on July 4th. The plan description quietly changed from "Unlimited" to "Extended" twelve days after launch.

Anthropic took a different approach. Their $17/$100/$200 tiers map to genuinely different user personas. A casual user, a power user, and a developer replacing an IDE. Those are different products with different willingness to pay. Then weekly rate limits targeting less than 5% of subscribers to push the heaviest users toward the API, where per-token pricing covers actual compute.

The pattern across all 50 companies: pure flat pricing is dying. Nearly half use two or three models simultaneously.

Here's the full breakdown:

1. Complete AI pricing guide: news.aakashg.com/p/how-…

2. Replit guide: news.aakashg.com/p/guid…

3. AI product strategy: news.aakashg.com/p/ai-p…

4. AI agents guide for PMs: news.aakashg.com/p/ai-a…

If you can't estimate your cost distribution across P10 to P90, you're not ready to set a price.