Skip to content
Multigrid

AI economics, markets & business models

The structure underneath the industry's numbers — where value is captured, why prices fall, and what a metered cost of goods does to a software business.

Most writing about the economics of this industry is a number with a date attached, and the date is doing more work than anyone admits. Spend totals, round sizes, market shares and price charts are all snapshots of a fast-moving system, and a page built on one is wrong within months in a way that is invisible to the reader.

These pages take the other half of the subject. Underneath the numbers is a structure that does not move: a variable cost of goods behaves in a particular way under a flat price; an efficiency gain raises total spend under a specific elasticity condition and lowers it otherwise; a new capability tier does a predictable thing to the price of the tier below it; a moat is only a moat if it accumulates faster than it decays. All of that can be derived, and once you have it you can read this quarter’s numbers yourself rather than being told what they mean.

So there are no figures here for revenue, valuation, funding, spending or market share — not for a company and not for the industry. Where a page needs one, it gives the calculation and names the kind of primary source the input has to come from. Every worked example is labelled as an assumption and chosen to be round. Substitute your own.

The Economics of Inference: Why Prices Keep Falling

The three independent terms behind a per-token price, why they compound, and why the price of a fixed capability collapses while the price of the best available model does not.

5 min read

Is Inference Profitable? The Model, Not the Rumour

The cost model of a token-serving business, why utilisation rather than price decides the sign of its margin, and which disclosures would actually answer the question.

5 min read

The Capex Cycle Behind AI: Depreciation as the Hidden Variable

How capital spending on accelerators becomes a cost per hour, why the assumed useful life moves that cost more than anything physical, and why a fixed-cost industry produces price wars.

5 min read

Business Models for AI Products: Seats, Usage and Outcomes

What each way of charging for an AI product rewards, the failure it produces, and the working-capital problem that appears once cost of goods is incurred before revenue.

5 min read

Why Usage-Based Pricing Is Hard to Get Right

What metering actually transfers to the buyer, why a cheaper expected bill can still lose to a flat one, and the roadmap constraint a metered unit quietly imposes on you.

5 min read

Gross Margin in AI-Native SaaS: Where the Ceiling Comes From

Why scale improves a classic software margin and cannot improve the inference part of an AI one, what that ceiling does to lifetime value, and the one move that restores scale economies.

5 min read

The Commoditisation of Model Capability

A testable definition of commoditisation, where switching costs actually live as opposed to where people assume they live, and what a new capability tier does to the price of the one below it.

5 min read

Moats in AI: The Test Each Claimed Advantage Has to Pass

Three questions that separate a durable advantage from a temporary lead, applied to weights, data, distribution, switching costs, scale and regulation.

5 min read

The Open-Weight Strategy: Why Give a Model Away?

The complement-commoditising argument stated properly, the four other reasons weights get released, what the release costs, and the inequality that decides it.

5 min read

Token Prices Over Time: How to Build a Series Worth Reading

Why a chart of price per million tokens is a price index with three unfixed problems, and how to construct a quality-adjusted, per-task series that answers the question people think that chart answers.

5 min read

Jevons Paradox and AI Demand: The Condition Nobody States

The elasticity condition under which cheaper inference raises total spending rather than lowering it, why AI demand is lumpy rather than smooth, and what the paradox does not predict.

5 min read

Who Captures the Value in the AI Stack?

Four bargaining-power questions applied to each layer of the stack, why the answers migrate over time, and how to identify which constraint is currently binding.

4 min read

The Application Layer: Thin Wrapper or Real Product?

What the accusation actually claims, three tests that settle it for a given product, and the supplier-exposure arithmetic that decides how much the answer costs you.

4 min read

AI Startup Funding: Reading the Numbers Without Being Misled

Why a company with a variable cost of goods needs more capital than classic software for the same growth, why recurring revenue is not comparable across margin structures, and how survivorship distorts every funding statistic.

5 min read

Enterprise AI Spend: Reading the Surveys, and the Bottom-Up Model

Seven questions that decide whether a spend survey means anything, a bottom-up cost model of an enterprise AI programme, and the single term that determines whether it pays back.

5 min read

Cost Curves: Training and Inference, Frontier and Fixed

Why there are four cost curves rather than two, which of them rise by definition, and the multiplier that lets cost per task climb while cost per token falls.

5 min read

Pricing an AI Feature Inside an Existing Product

The margin dilution a variable-cost feature causes when it is bundled into a fixed price, the price rise that would hold margin constant, and the conditions under which each of bundling, upselling and metering is right.

5 min read

Free Tiers as Customer Acquisition When Free Costs Money

The effective acquisition cost of a free tier with real marginal cost, the correlation that makes it worse than it looks, and the ceiling formula that tells you how generous you can be.

5 min read

The Bubble Question, Taken Seriously

How to turn the question into one that can be answered — the implied-expectation calculation, the load-bearing assumption on each side, and why the technology being real and the investment being mispriced are compatible.

5 min read

Second-Order Effects: What Cheap Cognition Changes

The viability threshold that decides which use cases switch on as cost falls, the order in which they do, and the two costs that do not fall and therefore end up binding.

7 min read

Other topics

AI economics, markets & business models · Multigrid