The Economics of AI

Ten big ideas

The technology arrives. Work has to change.
AI is a general-purpose technology: it can be used across many sectors. As with electricity and the internet, its gains depend on new skills, workflows, and institutions. With AI, capability can keep improving while people are still adapting.

02–05

Expensive to build. Cheaper to use.
Training the most capable models costs a lot. Using them has become cheaper per unit. Lower prices can encourage enough extra use that total spending rises.
Training cost ↑Price per use ↓
Explore the falling cost of useful work ↗
Every layer needs to earn its keep.
Suppliers sell hardware and power; operators sell compute; model labs and applications sell services. Each layer’s revenue is the next layer’s cost. Payment to a supplier today does not prove that the final user gets a worthwhile return.
How long the hardware lasts matters.
Depreciation spreads an asset’s cost over its useful life. A shorter useful life raises the annual expense. Returns depend on whether the hardware earns enough before it loses its economic value.
Value created is not value captured.
AI can save users time and improve their work without all that benefit becoming revenue for its providers. AI can enrich the world while disappointing some of the investors who financed it.

06–07

Tasks change before jobs and firms do.
A job bundles many tasks. AI can automate some, help with others, and create new work. Productivity depends on how firms reorganize the whole job—including review, responsibility, and how people learn.
There is no single AI jobs effect.
Some early-career workers show stress; some firms using AI intensively expand. Evidence about smaller teams and solo businesses adds other possibilities. These findings concern different groups and measures; they do not establish one effect for the whole economy.

08–09

India is exposed where it earns.
India’s IT and business-services strength rests partly on tasks AI can increasingly perform. Exposure is concentrated in particular occupations, firms, and cities. A national average can hide the places where adjustment is hardest.
India has options. We need institutions.
English, an IT base, digital public infrastructure, and a large domestic market give India places to experiment. Those assets yield gains when firms, schools, and public institutions help people build and enter better forms of work.

The future is a coordination problem.
Measure changes early, make useful experiments easier, and protect people through adjustment. When reading an AI headline, ask what changed, who gains, who pays, and what evidence would establish the claim.
MeasureExperimentCheckAdjust

How value is created, who captures it, and why depreciation matters.

Tasks, jobs, firms, and India’s routes into better work.

How useful ideas spread through practice, competition, and institutions.

Interested in reading more? Start with one of these, or explore the full collection.

General Purpose TechnologiesWhy the gains depend on changes around the technology.
The AI Capex LedgerHow to trace required returns through the AI industry.
The task is not the jobWhy automating tasks can change the work that remains.