The Missing Economics
AI inference costs are falling 10x per year. The economic frameworks we use to understand markets, productivity, and growth weren't built for this. Five critical pieces are missing.
In 1848, John Stuart Mill published Principles of Political Economy and more or less settled how the educated world thought about markets for the next century. Land, labor, capital. Supply, demand, price. The framework wasn't perfect, but it was good enough to explain a world where the scarce resource was human effort and the expensive thing was making stuff.
That framework still runs economics departments. It also can't explain what's happening right now.
The $0.15 Problem
In March 2024, OpenAI charged $30 per million input tokens for GPT-4. By July, GPT-4o mini did comparable work for $0.15. That's a 99.5% price drop in four months.
According to Epoch AI, the price of LLM inference at a given performance level has been falling at roughly 40x per year, and on some benchmarks, the decline is closer to 400x per year. For context, semiconductor costs fell at about 2x every two years during Moore's Law's golden era. This is faster by orders of magnitude.
Benjamin Shiller, a Brandeis economist and author of the recent AI Economics, applies standard economic tools to AI's impact on jobs, pricing, and market structure. His analysis of how AI reshapes existing markets is solid, but the book leaves a larger question untouched: what happens when the frameworks themselves stop fitting?
Where the Models Break
Classical economics assumes that factors of production (labor, capital, land) have marginal costs. Hire another worker, buy another machine, lease another acre. Prices adjust, equilibrium holds.
AI breaks this in a specific way. The marginal cost of cognitive work is approaching zero. Not the cost of building the model, which runs to billions. The cost of running it one more time, producing one more analysis, writing one more brief, reviewing one more contract. That cost is falling so fast that it will become a rounding error on most companies' balance sheets within a few years.
Economics has dealt with falling marginal costs before. Digital goods, software, and information all have near-zero marginal cost once created. Those were products. AI isn't a product. It's a factor of production that behaves like labor but scales like software. The standard production function (output as some combination of labor and capital) doesn't have a category for "an input that thinks, costs almost nothing to replicate, and gets better every quarter."
Nobody has proposed a widely accepted replacement.
The Subscription Trap
One place where the gap between economic reality and sticker price is already visible: consumer AI subscriptions.
OpenAI set ChatGPT Plus at $20 per month in early 2023, when heavy usage could cost them several dollars per session in inference. Two years later, equivalent-quality responses cost a fraction of a cent. The price hasn't moved.
This isn't unusual in software. SaaS margins improve as infrastructure gets cheaper. The speed of the underlying cost decline has no precedent in software. When Salesforce's server costs dropped, they dropped 20-30% per year. When AI inference costs drop, they drop 90% per year. The result is margin expansion that makes traditional SaaS economics look quaint.
The customers locked into annual enterprise contracts at 2024 per-token rates will look back the way someone who signed a 10-year lease on a Blockbuster franchise in 2006 looks back. The unit economics shifted under their feet while the contract held firm.
Jevons Paradox on Steroids
In 1865, William Stanley Jevons observed that making steam engines more efficient didn't reduce coal consumption. It increased it, because cheaper energy opened up uses that weren't economical before. Satya Nadella cited this directly after DeepSeek's efficiency breakthrough: "Jevons paradox strikes again. As AI gets more efficient and accessible, we will see its use skyrocket."
He's probably right. Jevons studied a world where a human had to decide each new use of coal. The AI version of this paradox operates differently. Agents can discover and exploit new use cases autonomously. Google has already launched agentic checkout, with AI agents executing purchases directly on merchant websites. Major tech companies are racing to build the interoperability and payment infrastructure for autonomous commerce.
When agents buy from agents, standard price theory gets strange. Prices exist, in part, to communicate information about scarcity to human decision-makers. An agent doesn't experience scarcity the way a person does. It processes price signals at machine speed and optimizes across thousands of variables simultaneously. The "market" still clears, but the mechanism that clears it has changed in ways that existing models don't capture.
The Toll Road That Can't Lose
One company has already figured out how to profit from AI without caring which direction the economics go: Stripe.
Stripe's metered billing API lets AI startups charge customers based on actual token usage. The company using Stripe's pass-through billing has a structural problem: as inference costs fall, per-user revenue falls with them. A startup charging a 50% markup on token costs watches its average revenue per user shrink every time the model provider cuts prices. The treadmill speeds up every quarter.
Stripe doesn't care. Their cut is a percentage of the transaction regardless of the amount. It's the Visa model: the percentage stays the same, the volume only goes up. The companies best positioned in AI economics aren't AI companies. They're infrastructure that processes the money.
What Gets Measured Gets Missed
The deeper problem is measurement. GDP, productivity statistics, labor market data. All of them were built on assumptions about what economic activity looks like. A lawyer researching a case for six hours generates measurable billable output. An AI doing the same research in four seconds generates... what, exactly? The billable hour collapses. The work was done. The output exists. The economic measurement framework sees a drop in legal services revenue and records it as contraction.
Acemoglu's recent work at MIT estimates that automation-heavy AI may actually depress measured labor demand and yield "only modest aggregate productivity gains." This might be an artifact of measuring the wrong thing. If AI produces equivalent output at near-zero cost, the value was created, but it just doesn't show up in metrics designed to track expensive human effort.
CNBC reported in January that AI-related components added roughly 0.9 percentage points to real U.S. GDP growth in 2025. That figure mostly captures spending on AI infrastructure: the GPUs, the data centers, the cloud contracts. It doesn't capture the value of work that AI performed but that nobody paid for in a way that national accounts can see. The measurement gap will widen as inference costs approach zero.
The Five Missing Chapters
What economics needs, and doesn't yet have, is a framework purpose-built for an economy where cognitive work is abundant and nearly free. Five pieces are missing:
A theory of data as a factor of production. Not data as an asset (that exists) but data as something with a rental rate, depreciation curve, and marginal product. When does data become a public good? When is it rivalrous? The answers matter for antitrust, taxation, and trade policy, and nobody has formalized them.
A model of agent-mediated markets. Price theory assumes human participants with bounded rationality and limited information. Markets populated by agents with superhuman processing speed and access to every public data point don't behave the same way. The efficient market hypothesis was controversial enough with human traders.
A revised production function. Output = f(labor, capital, data, intelligence), where intelligence is a factor that costs almost nothing at the margin but requires massive fixed investment to create. This is closer to increasing-returns models that economists like Brian Arthur explored in the 1990s, but nobody has built the full framework.
New measurement infrastructure. GDP was designed to count transactions. An economy where the most valuable work generates minimal transactions needs different metrics. Otherwise policymakers will be flying blind, reacting to statistics that describe an economy that no longer exists.
A distributional theory for zero-marginal-cost intelligence. If cognitive work costs nothing to produce, who captures the surplus? The model owners? The platform operators? The workers displaced by it? We know the answer won't be "everyone benefits equally" because that's never how a new factor of production distributes its gains. But we don't have a model that predicts the actual distribution.
The Textbook That Hasn't Been Written
The cost of cognitive work is collapsing. Autonomous agents are entering markets as participants, not just tools. The production function is missing a variable. GDP can't see the output.
The economics profession has a choice: build the framework before it's needed, or explain the wreckage after.
Where Do We Go From Here?
The companion piece, "Five Predictions for the Missing Economics," puts concrete stakes in the ground: outcome-based AI pricing by Q4 2027, the first autonomous agent liability dispute, the GDP measurement crisis, data costs emerging as a named line item in earnings reports, and agent-to-agent marketplaces crossing $100M in transactions. Each prediction has specific resolution criteria and a timeline. We track them publicly and publish post-mortems when we're wrong.