Part of why I love my work is I get to learn from founders all the time. I just got introduced to a new metric to measure how good an AI application is, and I think it’s intuitive, so I’m sharing with permission.
We work with an application layer company called Layer9 - they created an AI platform and services for telecom, an industry battling decades of tech debt, network designs that live only in retired engineers’ heads, and undocumented hardware integrations. Layer9 measures the quality of their application with a simple metric called RoTo (”row-tow”). Going deep with them on this metric over the past couple days makes me believe RoTo is the best way to understand the health of a software business today.
RoTo — Return on Token: “for every dollar of inference spend, how many dollars of revenue do you create?”
ROTO = $value or $revenue captured / $inference spend
How to measure RoTo
High and rising RoTo signals the product is accumulating proprietary context, deterministic workflows, daily usage, and building value for the customer over time—not simply reselling tokens at a markup. RoTo measures the economic value an application-layer company captures relative to its underlying model/inference spend, rather than treating token cost or gross margin alone as the outcome.
For application-layer AI companies, RoTo is the best north-star metric because it measures whether each unit of model spend is converted into repeatable, customer-valued work that generalized AI cannot easily replace.
Your customer is choosing to spend on your platform rather than buy frontier model tokens and the shape of that spend defines the quality of your business. RoTo helps assess that quality across four dimensions—differentiation, specialization, durability, and the ability to turn customer value into lasting margins. Here’s how to think about it more tactically:
RoTo distinguishes a compounding application from a thin wrapper.
If you have a really low RoTo, you’re a thin wrapper because you’re not adding that much value on top of the tokens. RoTo is high and rising when the company has created proprietary workflow, context, integration, and outcome value above generalized model capability.
High RoTo is evidence of valuable specialization rather than generic intelligence.
If you have low RoTo or it’s declining with new model capabilities or price reductions on existing models this indicates that what you are doing can be done by generalized intelligence. As the models continue the march of greater capabilities and lower cost, your ability to capture value will decline/disappear.
Conversely, high RoTo likely indicates the data is not readily available in the pre-training set and you are adding measurable value on top of the tokens. Your value can rise as model capability rises, or value creation/capture remains static as token prices drop - driving RoTo up.
High and rising RoTo points to a durable product or service.
A high RoTo vertical AI business indicates the solution is specific and sufficient. Customers are not returning to frontier models to resolve a problem and in the best business case, the frontier model can’t solve it on its own. The core application has embedded enough domain context and reliability to become the place where work gets completed. This is the proposed moat: not the base model, but the learning loop embedded in the specialized application that reliably completes domain work at lowest cost.
It separates value creation from cost optimization.
The stated framework has two independent levers:
How valuable is your thing versus just raw tokens? (the harness)
How good are you at delivering sufficient intelligence for a cheaper actual cost? (model choice and optimization)
Pure cost savings impacts RoTo but is an advantage that erodes as the underlying infrastructure (the models) marches down the cost curve. But in high RoTo applications, you can achieve price stability even as the underlying model cost drops. Value derived by your customers is not correlated to model costs and your margin just grows.
Thin wrappers are a function of weak models
Most people believe that scale wins, specialization loses, and anything you hand-built on top of the model eventually gets absorbed by the model. I am not convinced that as the models keep getting better, the dominant way to get that intelligence into an actual enterprise is through an FDE task force. When I hear FDE, I think immature software. The value of a technology compounds when software makes it simple to adopt quickly and apply broadly. Today’s need for embedded engineers is a sign we haven’t built that software yet.
Two years ago model use was simple. A model could handle a few thousand tokens of context and one turn of reasoning so there was only one thing you could build on top of it. “Wrapper” was an accurate description of the possible design space and that design space was tiny.
But now, interaction with the models is way more complex and you can do a ton of domain-specific work inside the application. The surface area for being more than a wrapper is expanding with every model release, the opposite of what the bitter lesson teaches.
A strong application-layer company will show that RoTo rises as it gains customer scale, workflow depth, and specialized context—not just as models get cheaper.
The goal at the application layer is to create a specialized daily driver with high and measurably increasing RoTo. It is a great operating metric to optimize and a very elegant way to measure the durability of the business you’re building.