The Economics of AI: Two-Week Online Course

A two-week online course on what AI applications cost and what they are worth. Smarter records every model call as a charge, with its tokens and its cost in USD, against the user, the application and the account that made it, priced from a table of prices per million tokens. Self-hosted GPU nodes are charged by the hour from the price of each compute profile, and Budget manifests turn spending limits into enforced controls, so students measure real costs and test real policies rather than estimate them. There are six 30-minute recorded lectures and a 90-minute live session each week. Week 1 measures costs, and week 2 builds business cases and controls. No programming is needed. See Course Formats and Grading for the format, the grading scale and the course policies, and The Economics of AI for an overview of the subject.

At a Glance

Students

Business, economics, finance and management students and professionals.

Deployment

A shared Smarter instance run by the instructor, with at least three models from two or more providers configured.

Workload

3 hours of instruction and 4 to 6 hours of assignments each week.

Cost

Expect less than $5 USD of model charges per student, capped by a per-student Budget.

Learning Objectives

By the end of the course, students can:

  1. Explain how model providers price requests, and what drives the cost of a conversation.

  2. Measure and compare the cost, latency and quality of AI tasks across models.

  3. Estimate the full cost per user per month of an AI application, including infrastructure.

  4. Build a business case with scenarios and sensitivity analysis.

  5. Design a budget policy and implement it as enforced controls.

Technical Prerequisites

  • Required: a computer with a current web browser, a reliable connection for video calls, and a spreadsheet program.

  • Required: spreadsheet basics: formulas, references and charts.

  • Required: reading a YAML file well enough to change a value (Learn YAML in Y minutes).

  • Helpful: introductory microeconomics and a basic business case or net present value calculation.

  • Not needed: programming.

Schedule

The files that the lectures refer to are listed in full under Smarter in the Source Code.

Week

Item

Content

1

Lecture 1 (30 min)

The unit of AI cost. Tokens, input and output prices, cached input, and why long conversations cost more than they seem to.

1

Lecture 2 (30 min)

Quality, latency and cost. Model tiers, reasoning models, tools and retrieval as cost multipliers, and comparing models fairly.

1

Lecture 3 (30 min)

Infrastructure. Fixed and variable costs, the baseline cost of a cluster (about $350 USD a month for Smarter), marginal cost per user, and self-hosting versus paying per token.

1

Live session 1 (90 min)

Live walk through Smarter’s cost records and the dashboard. Groups run a model comparison together. Start Assignment 1.

2

Lecture 4 (30 min)

The business case. Measurable benefits, adoption curves, scenarios, sensitivity, and what productivity research does and does not show.

2

Lecture 5 (30 min)

Controlling spending. Budgets, warnings and limits, cost accounting and chargeback.

2

Lecture 6 (30 min)

The bigger picture. Falling prices, the data center build-out, vendor concentration, labor market effects, and who captures the value.

2

Live session 2 (90 min)

Capstone presentations (about 5 minutes each, with 2 minutes of questions), then a debate: will AI be cheaper or more expensive for businesses in five years?

Assignment 1 is due at the end of week 1, Assignment 2 in the middle of week 2, and Assignment 3 and the capstone at the start of Live session 2.

Smarter in the Source Code

Every lecture, lab and assignment in this course measures costs with the records, prices and manifests below. Students see the records in the web console and the dashboard. The files are on the main branch of github.com/smarter-sh/smarter, for instructors and for students who want to see how a cost is computed. Python paths are under smarter/smarter/.

Measuring Cost

Topic

What Smarter does

Where to look

The charge

One record for each model call, plugin call or tool call, and for each hour of self-hosted compute: the account, the user, the session, the provider, the model, the prompt, completion and total tokens, and the cost in USD. A charge is created for every resource that took part, so the same request is seen by the user’s, the application’s and the account’s budgets. Cost Accounting.

account/models/charge.py

Prices

A price is USD per million tokens, markup included, for each combination of charge type, provider and model. A model without a price costs nothing, so the instructor loads the prices of the class’s models before Lab 1, as shown in Cost Accounting.

account/models/llm_prices.py (cost_of())

Totals for a period

Charges are aggregated by hour, and summed for any resource and time range, which is what a budget compares with its limits.

account/models/charge.py (AggregatedCharges, get_actuals())

The dashboard

Charts of charges and budgets in the web console, served by the dashboard’s API.

dashboard/views/views/api/charges.py, budgets.py

Comparing Models and Hosting

Topic

What Smarter does

Where to look

Changing the model

An LLMClient’s provider and defaultModel are two lines of its manifest, so the same tasks run on another provider’s model with no other change.

llmclient-simple.yaml

Providers

Each LLM provider that Smarter can call is a Provider resource, and new providers that speak the OpenAI protocol are added with a manifest. Smarter Provider.

apps/provider/

The price of a GPU node

Seven built-in compute profiles, from small CPU nodes to eight H100s, each with its AWS on-demand price per hour, from which self-hosted models are charged hourly.

llmhost/data/compute/, gpu-l4-1x.yaml

Open models

Twenty-four built-in LLMHost manifests for Llama, Gemma, Qwen, Mistral, Phi, DeepSeek and gpt-oss models, and embedding models, to price against per-token APIs. Smarter LLM Host.

llmhost/data/llmhost/

Platform infrastructure

The cluster, nodes, load balancer, database and other AWS services that a shared Smarter instance needs, all declared in Terraform, so each fixed cost can be traced to a resource.

smarter-infrastructure, aws/terraform/

Controlling Spending

Topic

What Smarter does

Where to look

Budgets

Limits in dollars or tokens, per hour, day, week or month and in total, for an account, a user, an application, a provider, a proxy or a self-hosted compute profile, that block or warn. Smarter Budget.

account/manifest/models/budget/spec.py

Seven budget policies

A student allowance ($10 a month, $100 in total), an account-wide monthly cap, a provider limited to $200 a day, a six-month project budget, a GPU budget of $500 a month, a warn-only budget on a proxy, and a token throttle for each user.

account/data/example-manifests/budgets/

Warnings and limits

A budget_warning signal at the warning threshold, once per period, and a budget_exceeded signal and a lock when a limit is reached.

account/signals.py, account/models/budget.py

Graded Assignments

Assignment 1: Measured Model Comparison

Students run twelve realistic tasks on three models from different providers, record cost, latency and a quality score for each from Smarter’s records, and reconstruct one cost from the price list.

Turn in: the spreadsheet, a chart of cost against quality, and a one-paragraph recommendation.

Criterion

Pts

Exemplary

Developing

Data

6

Complete and taken from Smarter’s records.

Missing runs or estimates.

Price list check

4

One cost rebuilt from the price list, with any difference explained.

Not attempted, or unexplained mismatch.

Quality scoring

5

Stated rubric applied consistently.

Impressions.

Recommendation

5

Fits a stated use, with the trade-off explained.

Cheapest or best, without reasons.

Assignment 2: Cost per User per Month

Students estimate the monthly cost per user of an AI assistant for organizations of 50, 500 and 5,000 users, combining measured model charges, a usage assumption, and infrastructure, and explain how the cost per user changes with scale.

Turn in: the spreadsheet and a one-page explanation with a chart.

Criterion

Pts

Exemplary

Developing

Model charges

5

From measurements, with a stated usage assumption.

Guessed.

Infrastructure

5

Fixed and variable costs separated.

Infrastructure ignored.

Scale analysis

6

Shows and explains economies of scale.

Three numbers without explanation.

Assumptions

4

Stated and reasonable.

Hidden.

Assignment 3: Budget Policy

Students design a budget policy for a team of twenty with two AI applications, implement it in Smarter, use up a small budget to show the limit working, and propose the warnings that should come first.

Turn in: a one-page policy, the budget manifests and screenshots of the limit.

Criterion

Pts

Exemplary

Developing

Policy

6

Limits based on expected use, with reasons.

Arbitrary.

Implementation

6

Manifests enforce the policy.

Do not match the policy.

Evidence

4

Limit reached and shown.

Not shown.

User experience

4

Warnings and messages designed for the user.

Users cut off without warning.

Capstone Projects

Students choose one, and present it in Live session 2.

Capstone A: An Investment Proposal for an AI Application

Learning objective. Build an evidence-based investment proposal, from measured unit costs to a three-year financial model, and present it to decision makers.

Students choose a use case in a real or realistic company, build a prototype assistant in Smarter, measure its cost and quality on twenty-five realistic tasks, and use the measurements in a financial model.

What students turn in:

  • The prototype manifest and the measured task results.

  • A three-year financial model with scenarios and sensitivity analysis.

  • A three-page investment memo, and a five-minute presentation.

How it is graded (100 points):

Criterion

Pts

Exemplary

Developing

Measured unit costs

20

From the prototype on realistic tasks, quality checked.

From price lists.

Financial model

25

Complete, with scenarios and sensitivity.

Missing categories or scenarios.

Assumptions and risks

20

Stated, sourced, tested; risks include price changes and vendor lock-in.

Hidden assumptions; risks ignored.

Memo

20

Clear recommendation a CFO could act on.

No recommendation.

Presentation

15

Persuasive and on time.

Overruns or reads slides.

Capstone B: Build, Buy or Self-Host

Learning objective. Compare the total cost of ownership of different ways to provide an AI capability, and identify the volume at which the best choice changes.

A fictional organization needs an internal assistant for 2,000 employees. Students compare three options: a commercial chat product with per-seat pricing, an assistant built on Smarter with a commercial model provider, and an assistant on Smarter with a self-hosted open model on rented GPU nodes. They measure what they can in Smarter, and research the rest.

What students turn in:

  • A total cost of ownership model for each option over three years, including people.

  • A break-even chart showing which option is cheapest at different usage levels.

  • A three-page recommendation, and a five-minute presentation.

How it is graded (100 points):

Criterion

Pts

Exemplary

Developing

Cost completeness

25

Licenses, model charges, infrastructure, people and switching costs.

Only the obvious costs.

Evidence

20

Measured where possible, sourced otherwise.

Unsourced numbers.

Break-even analysis

20

Correct, with the crossover points explained.

Missing or wrong.

Non-financial factors

15

Privacy, control, quality and risk weighed.

Cost only.

Recommendation and presentation

20

Clear, conditional and on time.

No clear choice.

Further Study

  • Price and performance data: Artificial Analysis, which compares models on quality, speed and price, and the providers’ own pricing pages.

  • Cloud cost management: the FinOps Foundation, its framework, and its working group material on the cost of AI.

  • Reports: the annual AI Index from Stanford HAI, with chapters on the economy and on the cost of training and running models.

  • Podcasts: Exponential View with Azeem Azhar, on technology and the economy, Odd Lots from Bloomberg, which regularly covers AI and the data center build-out, and the a16z Podcast.

  • Writing: Ethan Mollick’s One Useful Thing on AI and productivity at work, and Benedict Evans’s newsletter and presentations on technology markets.

  • YouTube: talks by Ajay Agrawal, Joshua Gans and Avi Goldfarb on the economics of prediction, and Daron Acemoglu’s lectures on automation and inequality.