Blog posts
Events
Cloudgov University
Quietly, a new utility bill has attached itself to the economy. Content platforms are metering AI crawlers per fetch. Agents pay per task completed. Every AI feature in every product carries a per interaction cost that moves with model prices, context lengths and retry behavior. Tokens have become a metered input to business, like power or bandwidth, except the price list changes every six weeks and almost nobody has installed the meter.
The short answer
AI unit economics is the discipline of knowing, per outcome, what intelligence costs you: per customer, per feature, per ticket, per agent workflow. It requires four capabilities:
- Metering every model interaction against an identity
- Attributing token spend to products, customers and outcomes
- Enforcing budgets in real time, especially per customer for AI products
- Routing each task to the cheapest model that meets its quality bar
Companies with these four price with confidence; companies without them discover their margins in the invoice.
Tokens behave like no cost you have managed before
- The price list rewrites itself every six weeks. New models, new tiers, price cuts, capability jumps. A unit cost calculated in January describes a different market by March, so unit economics must be computed continuously, not annually.
- Consumption is elastic in the worst way. Longer contexts, richer features, chattier agents: every product improvement tends to consume more tokens per outcome unless routing and summarization discipline push back. Costs drift upward by default.
- The variance is behavioral, not seasonal. One power user, one retry loop or one prompt change can move a customer’s cost profile by an order of magnitude overnight. Averages mislead; distributions govern.
- Spend is decided by people who never see the bill. Developers choose models, product managers design features, agents chain their own calls. Without identity level attribution, the connection between decision and cost never closes.
The margin trap in AI products
For anyone shipping AI features, the token line is cost of goods sold, and it hides a trap that has already caught real products: a feature priced on the proof of concept’s assumptions meets production behavior, longer contexts, heavier usage, a premium model someone switched to for quality, and the cost per customer quietly climbs past the price per customer. The product succeeds its way into negative margin.
The trap has a known anatomy:
- Estimated once, never recomputed. The unit cost lives in a launch slide while the real number moves every week.
- Averaged across customers. The mean looks fine while the top decile of users consumes multiples of their price. Per customer distributions are the honest view.
- Unbounded by design. Nothing in the stack enforces a ceiling, so the worst case is discovered rather than decided.
The escape is per customer token budgets: a decided ceiling, enforced in real time, that converts margin risk into a product design parameter. Above the threshold, the product degrades gracefully, upsells, or routes to cheaper models, your choice, made in advance, instead of the invoice’s choice made for you.
Cost per outcome: the number that matters
Token counts and per million prices are inputs. The governable metric is cost per outcome: what does a resolved ticket, a processed document, a completed agent workflow, a served customer actually cost in intelligence?
- It is comparable across model changes: when the landscape shifts, cost per outcome tells you instantly whether you won or lost
- It is the pricing floor for AI products, and the efficiency scoreboard for internal workloads
- It exposes the design levers: summarization between agent hops, caching, routing tiers, prompt discipline, each visible as a measured improvement
- Joined with infrastructure costs, it becomes true unit economics: the AI feature’s tokens plus the cluster it runs on plus the storage behind it, per outcome
That last join matters more than it looks: AI features run on Kubernetes and clouds, so cost per outcome only computes when token, container and infrastructure spend meet in one attributed plane. A token point tool and a separate cloud tool leave you doing the join in a spreadsheet forever.
Routing: the margin lever hiding in plain sight
The price gap between frontier models and capable smaller models exceeds an order of magnitude, and a large share of production tasks cannot tell the difference in output quality. Smart routing, sending each task to the cheapest model that meets its quality bar, under policy, with the quality bar measured, is therefore the single largest recurring margin lever in AI operations:
- It requires no code changes when done at the control plane
- It converts every market price cut into automatic savings
- It reserves premium capacity for the work that genuinely needs it, which usually improves the premium work too
- Its savings compound with volume, exactly when they matter most
A FinOps ROI score per interaction keeps routing honest: quality and cost measured together, so the cheap model earns its traffic rather than being imposed.
What this looks like by role
For the CFO
Add one line to product reviews: cost per outcome, per product, trending. Any AI feature that cannot produce it is running on estimated margins, and estimated margins in a market that reprices every six weeks are not margins.
For the product leader
Design the ceiling in, per customer budgets, graceful degradation, routing tiers, rather than discovering it in the invoice. The best AI products treat token behavior as a first class design surface, and their pricing pages show it.
For the FinOps leader
The token layer is the fastest growing spend class you govern, and the one where your levers, budgets, routing, attribution, act in real time rather than at renewal. Bring product and engineering one number, cost per outcome, and you own the conversation that decides margins.
For the engineering leader
Prompt discipline, context summarization and retry hygiene are now margin engineering. Give teams their own attributed numbers and the optimization happens without mandates; engineers cannot resist a measured inefficiency.
How Cloudgov.ai runs the meter
- TokenShield is the token economy’s control plane: every interaction metered against an identity, per customer and per team budgets that intervene before overspend, smart routing across the major providers, and a FinOps ROI score on every interaction.
- MulticloudShield and ContainerShield complete the unit economics join: the clouds and clusters under the AI features, attributed in the same plane, so cost per outcome is one query rather than one spreadsheet.
- AgentShield extends the meter to agents and their workflows, the fastest growing consumers in the economy.
Frequently asked questions
What are AI unit economics?
The cost of intelligence per business outcome: per customer served, ticket resolved, document processed or agent workflow completed. Computing it requires identity level metering, attribution to outcomes, and joining token spend with the infrastructure spend underneath.
Why do AI product margins erode?
Unit costs estimated once at launch meet production behavior: longer contexts, heavier users, premium model switches and retries. Without per customer budgets and continuous recomputation, cost per customer climbs past price per customer unnoticed.
What is a per customer token budget?
A decided spend ceiling per customer identity, enforced in real time. Above the threshold the product degrades gracefully, upsells or reroutes, converting margin risk into a design parameter instead of an invoice surprise.
How much can model routing save?
The price gap between frontier and capable smaller models exceeds an order of magnitude, and a large share of tasks show no quality difference. Routing at the control plane captures that gap continuously, with no code changes, and converts market price cuts into automatic savings.
What is cost per outcome?
Total intelligence cost divided by completed business outcomes, per workload. It is the metric that survives model landscape changes, prices AI products, and exposes design levers like summarization, caching and routing tiers.
Why should token and cloud costs live in one platform?
Because AI features run on clusters and clouds: true unit economics require joining token, container and infrastructure spend at matching granularity. Separate tools leave the join, and therefore the margin question, unanswered.
Install the meter before the bill installs it for you
Cloudgov.ai meters, attributes, budgets and routes the token economy, joined with the clouds underneath, so your cost per outcome is a number you decide rather than discover. Talk to our team, or start with our resources.