MarginDash

Ryan Jeng and Jonny Wang were tired of watching their AI bills destroy their margins. Every startup building with AI faces the same brutal math—you ship a feature that uses OpenAI, charge the customer $50, and then find out Claude ran up $45 in API calls. Nobody knows what anything actually costs until the bill comes. So they built MarginDash, the cost tracking platform that ties every AI API call to actual revenue.

Here's what MarginDash does: you install their SDK—it weighs almost nothing—and it logs model names, token counts, and customer IDs on every call. Meanwhile, Stripe invoices sync automatically so you know exactly how much revenue came in from each customer. Their dashboards show you margins by customer, by feature, by model in real time. Not "estimated costs." Real costs mapped to real revenue. You can set budgets for anything—organization-wide, per-feature like "summarize," per-customer—and if someone hits 80%, 100%, whatever threshold you set, the SDK blocks the request before it even reaches the provider. They've got over 100 models covered across OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure, Groq, and more, with rates updating daily.

What makes this interesting is the architecture. There's no proxy hop your app has to route through. Your app still calls providers directly. MarginDash is a lightweight control plane that your SDK polls before each call. If the customer's budget is blown, the SDK returns early—no provider call, no charge. It's basically a budget guard that enforces decisions at runtime. You also get what-if simulators to compare model quality per dollar, so you can see "hey, switching from GPT-4o to Haiku saves us 90% on this feature with only 5% accuracy loss."

Ryan and Jonny are betting that as AI becomes infrastructure, the difference between a profitable AI company and a burning-money one is whether you actually know your unit economics. Every VC asking "what's your CAC?" is going to start asking "what's your AI margin per customer?" MarginDash answers that question.