A margin guard that shows which customers and features lose you money on AI
Joins your LLM cost with Stripe revenue per customer, flags accounts that cost more than they pay, and caps AI usage by plan before it eats your margin.
- Issue
- No. 01
- Published
- Nov 16, 2026
- Category
- Dev tools
- Who pays
- Seed to Series A SaaS teams (2 to 30 people) that sell AI features on flat or seat plans and call model APIs through OpenAI, Anthropic, OpenRouter or a gateway such as LiteLLM, Vercel or Cloudflare.
- Pricing
- Free up to a set amount of tracked AI spend, then a flat monthly tier by tracked spend. Avoid a percentage of spend: gateways like Vercel charge no token markup, so buyers compare against free.
01Problem
Model bills arrive as one number per provider. Gateways and observability tools can show cost per request or per tag, and Stripe shows revenue per customer, but nothing small teams use joins the two. So founders can't see which customers or features are unprofitable, plans get priced on guesses, and a few heavy accounts quietly push gross margin down. When they do cap usage, the cap is per API key, not per the plan the customer pays for.
02Why now
Tokens per task are rising faster than prices fall. Inference for the same capability gets about 10x cheaper a year, yet reasoning models now serve over half of tokens on OpenRouter and average prompts grew about 4x in 2025.
AI products run on thinner margins than classic SaaS (about 45% on average in 2025 per ICONIQ), so a few heavy users matter.
Pricing is moving to usage and credits. Cursor and GitHub Copilot both moved to metered limits in June 2025, and ICONIQ reports consumption pricing rising from 35% to 42% of AI companies. You can't set those limits without cost per customer.
The independent tools are being bought: Helicone (now in maintenance mode), Langfuse, Portkey, OpenMeter and Metronome were all acquired in the last 15 months. Teams want a neutral layer that works with whatever gateway they end up on.
03Signals
- Enterprises spent $37B on generative AI in 2025, up 3.2x from $11.5B, and $12.5B of it went to foundation-model APIs.Menlo Ventures, Dec 2025 ↗
- 98% of FinOps practitioners now manage AI spend, up from 31% two years earlier, and FinOps for AI is their top priority (n=1,192).State of FinOps 2026 ↗
- Fast-growing AI startups average about 25% gross margin, often negative; steadier ones average about 60%.Bessemer, State of AI 2025 ↗
- Average AI product gross margin was 45% in 2025, and 92% of execs say the true cost of AI is hard to predict, with token spend the top source of overruns.ICONIQ 2026 via SaaStr ↗
- 84% of companies say AI costs cut gross margin by more than 6 points, and 24% miss AI cost forecasts by more than 50% (n=372).CFO Dive, Sep 2025 ↗
- 53% of US enterprise leaders have no formal cost-per-outcome metric for LLMs (n=102).G2, Oct 2026 ↗
- Cursor moved its Pro plan from request limits to compute limits because new models spend more tokens per request, then refunded surprise charges.Simon Willison, Jul 2025 ↗
- Founders ask the exact question in public: how to track LLM cost per customer in a multi-tenant app.OpenAI developer forum ↗
- Stripe previewed token billing in March 2026, which meters token use per customer and adds a markup. It is waitlist only and handles pass-through, not margin on flat plans.TechCrunch, Mar 2026 ↗
04Alternatives
05Market analysis
Where the opening is
A generic proxy that tags LLM calls and alerts on overspend is crowded and consolidating: LiteLLM, Cloudflare, Vercel and OpenRouter already ship budgets per key or tag, and Stripe is previewing token billing. The gap is margin, not cost: join LLM cost with Stripe revenue per customer and plan, and cap usage by the plan a customer pays for. Only pre-launch or very small players do the Stripe join (Margine, MarginWard), and none pairs it with plan-aware caps.
Market size
- Enterprises spent $37B on generative AI in 2025, 3.2x the year before; $12.5B went to model APIs (Menlo Ventures).
- Gartner forecasts the AI platforms and models market growing from $39.3B in 2025 to $64.3B in 2026, with buyers focused on usage efficiency and cost control (Gartner).
- Consumption pricing rose from 35% to 42% of AI companies (ICONIQ via SaaStr), and every metered plan needs cost per customer to set its limits.
Competitors
| Product | What it does on cost | Price |
|---|---|---|
| LiteLLM | Spend and hard budgets per key, user, team or tag; Slack alerts | Open source; enterprise unpublished |
| Cloudflare AI Gateway | Spend limits by metadata such as user id (beta) | Not listed |
| Vercel AI Gateway | Cost by user and tag, budgets, no token markup | $0.075 per 1k tag writes |
| OpenRouter | Workspace and key budgets | Not listed |
| Langfuse | Tracing with cost | Free to $2,499/mo |
| Stripe token billing | Per-customer token metering with markup; waitlist | Not listed |
| CloudZero / Vantage | AI cost per customer and feature, for FinOps teams | Quote |
| Spendline | Proxy with customer header and budgets; revenue by CSV | Pilot |
Main risk
Stripe token billing reaching general availability, or the gateways adding margin views. Position it as a neutral layer on top of whatever gateway and Stripe setup a team already has.
After the first version
- Price simulator: the margin each plan would have at a new price or limit
- Write usage back to Stripe for overage billing
- Cost-aware routing to a cheaper model per feature
06First version
The smallest product worth charging for.
- Read-only Stripe connection: MRR and plan per customer
- Cost intake from where teams already are: OpenRouter, LiteLLM, Vercel or Cloudflare logs, or a one-line SDK wrapper with customer and feature tags
- Margin per customer, per plan and per feature, with a list of accounts that cost more than they pay
- Plan-aware limits: "Pro gets $X of AI a month", enforced by returning a cheaper model or a limit flag your app reads
- Slack or email alert when a customer or feature crosses its margin floor
Our team designs and ships AI SaaS products on a monthly subscription. Start with a 3-day free trial.
More ideas
All ideas →- No. 03Nov 16, 2026Vertical AIA student loan watchdog that keeps borrowers on the right plan and catches servicer mistakesReads your federal loan data and every servicer letter, picks the plan that fits under the 2026 rules, tracks each deadline, and flags the dropped payments and wrong bills that servicers now leave borrowers to find on their own.9 signals6 alternatives
- No. 02Nov 16, 2026Ops automationA payment guard that checks every vendor bank change before a small business paysReads the AP inbox and the vendor list, flags changed bank details and look-alike senders, verifies the payee through a channel the fraudster doesn't control, and keeps the evidence your bank now asks for.9 signals6 alternatives