Many organizations lack full visibility into how much they’re spending on AI.

Not the annual budget. They know that. They’re spending $50 million on AI. Or $200 million. Pick a number.

What they don’t know is: per model, per use case, per business outcome. They don’t know if they’re spending $10,000 per model or $10 million. They don’t know if the training cost or the inference cost or the data preparation cost is the dominant expense. They don’t know if they could cut spending in half by consolidating infrastructure, or if they’re already at optimal consolidation.

They spend $50 million and they have no idea whether they got value.

This is not a small problem. This is the foundation of whether your AI program scales or whether it collapses under its own weight.

Why Most Organizations Get This Wrong

The structure of AI spending makes it invisible.

A model gets built by the data science team. The compute cost lives in the cloud bill. The data preparation cost is part of the analytics infrastructure budget. The monitoring cost is buried in the ops budget. The cost of maintaining the model in production is spread across multiple teams. The cost of retraining is charged to whoever requested the retraining.

So when you look at your P&L, you see: cloud bill, analytics infrastructure, ops budget, engineering headcount. You don’t see: AI.

You’ve spent hundreds of millions on cloud compute, and you have no way to trace it back to models. You’ve spent millions on data infrastructure, and you have no way to trace it back to business outcomes.

The usual response is to create an “AI budget.” Designate some money for AI initiatives. Track it by project. This gives you visibility into planned AI spending. It does nothing for actual spending.

The real problem is that AI spending is scattered across multiple cost centers, charged by different mechanisms, optimized independently. One team optimizes for compute cost. Another for data transfer. Another for headcount. Nobody optimizes for: “Is this model worth what we’re spending on it?”

What Unit Economics Actually Means

Unit economics is a simple concept: How much does it cost to deliver one unit of value?

In e-commerce, unit economics is: cost to acquire a customer, cost to fulfill an order, margin per order, lifetime value per customer. You can see, at a glance, whether your business is profitable. If acquisition costs $50, and margin per customer is $30, you’re losing money. If acquisition costs $5 and margin per customer is $30, you’re winning.

In AI, unit economics would be: cost to build one model, cost to maintain one model per year, cost per inference, cost per model update, cost to identify and fix one model problem.

Most organizations can’t answer these questions.

What You Should Be Measuring

Start with this:

Cost per model per year. Take all the costs associated with one model—training, inference, monitoring, maintenance, retraining, fixing problems—and divide by the number of years you expect to run it. If you’re paying $2 million per year to run a single model, that number should shock you into asking: Is this model worth $2 million?

Cost per inference. For high-volume inference models, what does each prediction cost? Infrastructure, data preparation, monitoring. Some models cost $0.001 per inference. Some cost $1. If you’re serving a billion inferences per year, the difference between $0.001 and $0.01 is $9 million.

Cost per business outcome. This is harder. It requires tying the model back to revenue, or risk reduction, or cost savings. But you need to try. If the model identifies fraud, your unit economics is: cost to run the model per year, divided by the fraud prevented. If the model increases conversion rate, your unit economics is: cost to run the model, divided by the incremental revenue. If the model is wrong, what’s the cost?

Cost per model update. Some models need to be retrained weekly. Some monthly. Some yearly. What’s the cost of each retraining cycle? If you’re retraining a model every week at a cost of $50,000 per retraining, that’s $2.6 million per year in retraining costs alone. Is the model’s accuracy improvement worth it?

Cost of model-induced problems. This is what most organizations never measure. When a model breaks, what’s the cost? How much customer impact? How much engineering time to fix? How long until you notice? Most organizations run models that, if broken, could cost millions in customer harm or revenue loss, but they’ve never calculated what that actually costs.

What Changes When You Know This

Once you know these numbers, everything changes.

You discover that you have seventeen models doing similar things, costing $500K each, when you could consolidate into two models costing $200K each. You don’t discover this unless you know the unit economics.

You discover that model A costs $2 million per year and model B costs $500K per year, but they deliver similar business value. You might retire model A. You don’t even ask the question unless you know the unit economics.

You discover that a model is being retrained weekly at massive cost, when retraining monthly would give almost the same accuracy. You could cut the cost by 75%. You can’t make this decision unless you know the cost per update.

You discover that you’re running a high-cost model that handles edge cases that occur once a month. You could build a cheaper model for the 99% case and handle the edge cases manually. You can’t make this tradeoff unless you know the unit economics of each approach.

Most importantly: You get a way to evaluate new models. Someone wants to build a model. You ask: What’s the unit economics? What’s this going to cost us? What business value will it deliver? Is the cost reasonable relative to that value? You make a decision based on first principles, not on enthusiasm for the technology.

The Coordination Problem

Here’s where it gets hard: Most of these costs are owned by different teams.

The data science team owns modeling cost. The infrastructure team owns compute cost. The ops team owns monitoring cost. The product team owns the decision about whether to retrain.

To calculate unit economics, you need to break down silos. You need the infrastructure team to report compute cost by model. You need the ops team to report monitoring cost by model. You need the data science team to report labor cost. You need the product team to report business impact.

This is annoying. Nobody’s systems are designed for this. You’ll hear: “We can’t easily attribute costs to individual models. That’s not how our cloud billing works.”

Here’s the truth: You’re not measuring it because it’s hard, not because it’s impossible.

And you’re making billions of dollars of decisions without the information you need.

What This Costs to Build

To get real unit economics, you need:

1. Clear cost attribution. Your cloud platform needs to tag costs by model (or you need to track this manually). This is doable. It requires discipline.

2. Centralized tracking. One system that aggregates costs from multiple sources. This is annoying to build. But it’s doable.

3. Regular reporting. Someone runs this every month. Identifies models that are expensive. Identifies cost trends. Brings this to leadership.

4. Governance tied to unit economics. New models get approved based on projected unit economics. Existing models get reviewed based on actual unit economics. You make decisions based on cost-value tradeoff, not on hope that the model will be valuable.

The cost of building this is real: engineering time, ongoing operational burden, discipline to maintain it.

The benefit is also real: You stop paying for models that aren’t worth it. You consolidate overlapping models. You optimize the expensive ones. You make better decisions about where to invest in AI.

Why This Matters

Your board asks: “What are we getting for our $200 million AI spend?”

If you have unit economics, you can answer: “This model generates $50 million in value per year and costs $5 million to run. This model costs $3 million and prevents $30 million in fraud risk. This model costs $500K but generates only $2 million in value, so we’re retiring it.”

If you don’t have unit economics, your answer is: “We’re investing in our AI future. We’re building capabilities. It’s hard to measure.”

The first answer is leadership. The second is theater.